diff --git a/run.json b/run.json
index e031c502e..3f46c83bb 100644
--- a/run.json
+++ b/run.json
@@ -493,7 +493,7 @@
     "kind": "running"
   },
   "status_updated_at": "2026-05-23T16:30:00.579662Z",
-  "last_event_at": "2026-05-23T17:14:03.422520Z",
+  "last_event_at": "2026-05-23T17:36:14.461171Z",
   "pending_control": null,
   "checkpoints": [
     {
@@ -741,9 +741,9 @@
       }
     },
     {
-      "seq": 0,
+      "seq": 1024,
       "checkpoint": {
-        "timestamp": "2026-05-23T17:14:03.544702Z",
+        "timestamp": "2026-05-23T17:14:07.818229Z",
         "current_node": "implement",
         "completed_nodes": [
           "start",
@@ -754,30 +754,157 @@
         ],
         "node_retries": {},
         "context_values": {
+          "response.implement": "Implemented the mid-stage agent interview tools and metadata plumbing.\n\nSummary of changes:\n- Extended canonical `InterviewOption` with optional `description` and `preview`.\n- Updated OpenAPI and regenerated the TypeScript API client:\n  - `ApiQuestion.options` now uses `InterviewOption`.\n  - Removed stale `ApiQuestionOption` export/file.\n- Added model-native agent question tools:\n  - OpenAI: `request_user_input`\n  - Anthropic: `AskUserQuestion`\n  - Gemini: no question tool\n- Added per-turn `AgentToolRuntime` so cached sessions receive current workflow/stage interview context dynamically.\n- Enforced question-tool round rules:\n  - question tool runs alone\n  - mixed tool peers get model-visible errors\n  - multiple separate question-tool calls execute only the first and tell the model to batch\n- Added shared run-level interview blocker/guard used by both human gates and agent questions.\n- Added workflow runtime for agent-originated batched questions:\n  - emits/registers all questions first\n  - blocks once per batch\n  - waits concurrently\n  - emits completion/interruption/timeout events per question\n  - cleans up pending questions on cancellation/drop\n- Preserved answer endpoints/paths through existing pending interview projection and submission flow.\n- Updated web/Slack display:\n  - web shows option descriptions\n  - Slack renders option descriptions where practical\n  - preview is captured/exposed but not specially rendered\n- Added/updated tests across agent, workflow, API, store, Slack, and web.\n\nValidation run:\n- `cargo check -q -p fabro-agent -p fabro-workflow -p fabro-api -p fabro-server -p fabro-slack -p fabro-store -p fabro-cli --tests` ✅\n- `cargo +nightly-2026-04-14 fmt --check --all` ✅\n- `cargo +nightly-2026-04-14 clippy -q -p fabro-agent -p fabro-workflow -p fabro-api -p fabro-server -p fabro-slack -p fabro-store -p fabro-cli --all-targets -- -D warnings` ✅\n- `cargo nextest run -p fabro-agent -p fabro-workflow -p fabro-api ...focused question/API tests...` ✅\n- `cargo nextest run -p fabro-interview` ✅\n- `cargo nextest run -p fabro-slack -p fabro-store --lib --tests` ✅\n- `cd apps/fabro-web && bun test ./app/components/interview-dock.test.tsx ./app/components/stage-renderers/helpers.test.ts` ✅\n- `cd apps/fabro-web && bun run typecheck` ✅\n- `cd lib/packages/fabro-api-client && bun run typecheck` ✅\n\nNoted failures:\n- `cargo nextest run -p fabro-server ...` fails on `server::tests::get_graph_returns_svg`; rerun of the single test also fails. The error is the graph renderer subprocess returning nextest output (`running 0 tests`) instead of SVG, unrelated to the interview changes.\n- Full `cd apps/fabro-web && bun test` still has existing unrelated failures in run-detail/lifecycle tests; focused changed web tests pass and typecheck passes.",
+          "command.output": "blob://sha256/12ae32cb1ec02d01eda3581b127c1fee3b0dc53572ed6baf239721a03d82e126",
+          "failure_signature": "",
           "graph.goal": "# Mid-Stage Agent Interview Tools\n\n## Summary\n\nAdd model-native question tools that let agents pause mid-stage and ask the human for input through Fabro's existing interview system.\n\nOpenAI-profile agents get `request_user_input`; Anthropic-profile agents get `AskUserQuestion`. When either tool is called, Fabro creates pending interview questions, surfaces them through the existing web/API/Slack paths, waits for answers, then returns provider-shaped tool results so the model can continue the same stage.\n\n## Key Changes\n\n- Extend the existing interview contract without type sprawl:\n  - Add optional `description` and `preview` fields to the canonical `fabro_types::InterviewOption`; reuse that type through `fabro-api` replacements instead of introducing `AgentQuestionOption`, API-only aliases, or adapter-only duplicate types.\n  - Update OpenAPI, generated Rust/TypeScript clients, event conversion, projection, Slack/web mappers, and the existing `with_replacement(\"InterviewOption\", \"fabro_types::InterviewOption\", ...)` parity tests.\n  - Treat both fields as untrusted model-authored display data. Store and expose them after enforcing bounded lengths; truncate or reject oversized values consistently before persistence.\n  - Initial UI behavior: display `description` under option labels where practical. Capture and expose `preview`, but do not render preview content specially in web or Slack v1.\n\n- Add a shared run-level interview runtime:\n  - Move the private human-node blocked-state refcount into a reusable run-level guard used by both `HumanHandler` and agent question tools, so `RunUnblocked` is emitted only when all human and agent interviews for the run are resolved.\n  - Runtime accepts the interviewer, workflow emitter, stage scope, stage id, tool call id, and normalized questions.\n  - Support batch asks as a first-class operation: emit/register all questions first, mark the run blocked once, await all answers concurrently, then emit completion/timeout/interrupted events per question and unblock when the batch resolves.\n  - Batch support applies only to multiple `questions[]` inside one question-tool call. Do not aggregate multiple separate question-tool calls from the same model round.\n  - Generate safe internal question IDs with a ULID/UUID plus stage visit/tool-call context; store original model question IDs/text in question metadata for provider result mapping.\n\n- Add provider-specific agent tools:\n  - `request_user_input` for `AgentProfileKind::OpenAi`.\n    - Accept Codex-compatible schema: `questions[]` with `id`, `header`, `question`, and `options[] { label, description }`.\n    - Normalize each question to `QuestionType::MultipleChoice` with `allow_freeform: true`.\n    - Return JSON text matching Codex shape, keyed by the original model question ID: `{\"answers\":{\"id\":{\"answers\":[\"...\"]}}}`.\n  - `AskUserQuestion` for `AgentProfileKind::Anthropic`.\n    - Accept Claude-compatible schema: `questions[]` with `question`, `header`, `options[] { label, description, preview? }`, and `multiSelect`.\n    - Normalize single-select to `MultipleChoice`, multi-select to `MultiSelect`, always with `allow_freeform: true`.\n    - Return Claude-style tool result text keyed by the original question text: `User has answered your questions: \"...question...\"=\"answer\". You can now continue...`.\n  - Answer formatting for both tools returns user-facing option labels to the model. Preserve internal option keys for validation and event storage. For multi-select, preserve the submission order supplied by the answer path.\n\n- Thread workflow interview context into agent tool execution:\n  - Add an explicit per-turn agent tool runtime context passed into `process_input` or an adjacent `process_input_with_runtime` API. It carries the interviewer, workflow emitter, stage scope/id, shared block guard, and provider answer formatter.\n  - Do not capture stage-specific interview handles in the profile registry or cached session construction; cached full-fidelity sessions must receive the current turn's stage context dynamically.\n  - Child/subagent sessions must not expose these question tools. If somehow called outside the root session, return a model-visible error.\n  - Question tools must execute alone in a model tool round. If a round contains one question tool plus any other tool call, execute the question tool and return model-visible error results for the non-question peers, preserving tool-call/tool-result ordering. If a round contains multiple separate question-tool calls, execute only the first and return model-visible error results for the later question-tool calls instructing the model to combine questions into one `questions[]` batch.\n  - Agent-originated questions have no per-question timeout in v1 because the provider schemas do not include timeout. They rely on existing stage timeout, wall-clock timeout, cancellation, and interruption behavior.\n\n- Preserve existing answer paths:\n  - Do not add a new answer endpoint.\n  - Continue using `GET /runs/{id}/questions` and `POST /runs/{id}/questions/{qid}/answer`.\n  - Keep `ControlInterviewer`, web `InterviewDock`, Slack blocks, and run projection as the delivery mechanism.\n\n## Test Plan\n\n- Unit tests for schema parsing and normalization:\n  - Codex request with descriptions maps to Fabro multiple-choice questions and returns answers by model question ID.\n  - Claude request with `multiSelect: true` maps to `MultiSelect` and returns comma-separated answer text.\n  - Batched Codex and Claude requests surface all questions as pending before awaiting answers, then return one result with every answer mapped to the original model ID/text.\n  - Optional `preview` and `description` survive event, projection, API conversion, OpenAPI replacement tests, and TypeScript client generation.\n  - Oversized `description`/`preview` values are bounded before persistence and never rendered as trusted HTML.\n\n- Workflow and agent tests:\n  - OpenAI-profile session advertises `request_user_input`; Anthropic-profile session advertises `AskUserQuestion`; Gemini advertises neither.\n  - Subagent profiles do not advertise the question tools.\n  - Root agent can ask a question and resume after the answer.\n  - Subagent or missing interview context returns a clear tool error.\n  - Cached full-fidelity session emits interview events against the current stage, not the original cached stage.\n  - A mixed tool round containing a human-question tool plus another tool preserves all required tool results and rejects the peer calls with model-visible errors.\n  - A round with multiple separate question-tool calls executes only the first and rejects later question-tool calls with model-visible errors.\n\n- Server, projection, and UI tests:\n  - `InterviewStarted` with option metadata appears in pending questions.\n  - Submitting valid selected, multi-selected, and freeform answers unblocks the waiting tool.\n  - Duplicate answer submission remains rejected through existing accepted-question logic.\n  - Parallel human gate plus agent question keeps the run blocked until both are answered.\n  - Pause, cancel, and interrupt while an agent question is waiting resolve pending questions consistently and do not leave the run blocked.\n  - Stage timeout or wall-clock timeout while an agent question is waiting interrupts the batch; no per-question timeout event is expected unless a future schema adds timeout.\n  - Slack answer submissions work for agent-originated questions using the same pending interview transport.\n\n- Run checks:\n  - `cargo nextest run -p fabro-interview -p fabro-workflow -p fabro-server -p fabro-agent`\n  - `cd apps/fabro-web && bun test && bun run typecheck`\n  - Regenerate and verify OpenAPI-derived Rust and TypeScript clients after schema changes.\n\n## Assumptions\n\n- This feature is only for in-process/API-backed agent sessions, not ACP external agents in v1.\n- `preview` is stored and exposed but not rendered specially in the first implementation.\n- Human-question tools are available only during root agent execution inside a workflow run with an active interviewer.\n- Existing interview events remain the source of truth for pending questions; no separate agent-question event family is added.\n- The implementation should prefer extending existing interview structs and replacement mappings over adding parallel API DTOs or conversion-only aliases.\n",
+          "graph.rankdir": "LR",
+          "internal.node_visit_count": 1,
+          "internal.retry_count.implement": 0,
+          "failure_class": "",
+          "internal.retry_count.preflight_lint": 0,
+          "internal.retry_count.start": 0,
+          "internal.run_id": "01KSATQNAXG41FHKV0QH5N1QGC",
+          "internal.thread_id": "preflight_lint",
           "internal.work_dir": "/home/daytona/workspace/fabro",
+          "outcome": "succeeded",
+          "current_node": "implement",
+          "thread.preflight_compile.current_node": "preflight_lint",
+          "thread.toolchain.current_node": "preflight_compile",
+          "thread.preflight_lint.current_node": "implement",
+          "internal.retry_count.toolchain": 0,
+          "last_response": "Implemented the mid-stage agent interview tools and metadata plumbing.\n\nSummary of changes:\n- Extended canonical `InterviewOption` with optional `description` and `preview`.\n- Updated OpenAPI and rege",
+          "internal.fidelity": "compact",
+          "thread.start.current_node": "toolchain",
+          "internal.retry_count.preflight_compile": 0,
+          "last_stage": "implement",
+          "graph.model_stylesheet": "\n            * { model: claude-opus-4-7; }\n        "
+        },
+        "node_outcomes": {
+          "preflight_lint": {
+            "status": "succeeded",
+            "context_updates": {
+              "command.output": "blob://sha256/12ae32cb1ec02d01eda3581b127c1fee3b0dc53572ed6baf239721a03d82e126"
+            },
+            "notes": "Script completed: cargo +nightly-2026-04-14 clippy -q --workspace --all-targets -- -D warnings 2>&1",
+            "usage": null
+          },
+          "implement": {
+            "status": "succeeded",
+            "context_updates": {
+              "last_response": "Implemented the mid-stage agent interview tools and metadata plumbing.\n\nSummary of changes:\n- Extended canonical `InterviewOption` with optional `description` and `preview`.\n- Updated OpenAPI and rege",
+              "last_stage": "implement",
+              "response.implement": "Implemented the mid-stage agent interview tools and metadata plumbing.\n\nSummary of changes:\n- Extended canonical `InterviewOption` with optional `description` and `preview`.\n- Updated OpenAPI and regenerated the TypeScript API client:\n  - `ApiQuestion.options` now uses `InterviewOption`.\n  - Removed stale `ApiQuestionOption` export/file.\n- Added model-native agent question tools:\n  - OpenAI: `request_user_input`\n  - Anthropic: `AskUserQuestion`\n  - Gemini: no question tool\n- Added per-turn `AgentToolRuntime` so cached sessions receive current workflow/stage interview context dynamically.\n- Enforced question-tool round rules:\n  - question tool runs alone\n  - mixed tool peers get model-visible errors\n  - multiple separate question-tool calls execute only the first and tell the model to batch\n- Added shared run-level interview blocker/guard used by both human gates and agent questions.\n- Added workflow runtime for agent-originated batched questions:\n  - emits/registers all questions first\n  - blocks once per batch\n  - waits concurrently\n  - emits completion/interruption/timeout events per question\n  - cleans up pending questions on cancellation/drop\n- Preserved answer endpoints/paths through existing pending interview projection and submission flow.\n- Updated web/Slack display:\n  - web shows option descriptions\n  - Slack renders option descriptions where practical\n  - preview is captured/exposed but not specially rendered\n- Added/updated tests across agent, workflow, API, store, Slack, and web.\n\nValidation run:\n- `cargo check -q -p fabro-agent -p fabro-workflow -p fabro-api -p fabro-server -p fabro-slack -p fabro-store -p fabro-cli --tests` ✅\n- `cargo +nightly-2026-04-14 fmt --check --all` ✅\n- `cargo +nightly-2026-04-14 clippy -q -p fabro-agent -p fabro-workflow -p fabro-api -p fabro-server -p fabro-slack -p fabro-store -p fabro-cli --all-targets -- -D warnings` ✅\n- `cargo nextest run -p fabro-agent -p fabro-workflow -p fabro-api ...focused question/API tests...` ✅\n- `cargo nextest run -p fabro-interview` ✅\n- `cargo nextest run -p fabro-slack -p fabro-store --lib --tests` ✅\n- `cd apps/fabro-web && bun test ./app/components/interview-dock.test.tsx ./app/components/stage-renderers/helpers.test.ts` ✅\n- `cd apps/fabro-web && bun run typecheck` ✅\n- `cd lib/packages/fabro-api-client && bun run typecheck` ✅\n\nNoted failures:\n- `cargo nextest run -p fabro-server ...` fails on `server::tests::get_graph_returns_svg`; rerun of the single test also fails. The error is the graph renderer subprocess returning nextest output (`running 0 tests`) instead of SVG, unrelated to the interview changes.\n- Full `cd apps/fabro-web && bun test` still has existing unrelated failures in run-detail/lifecycle tests; focused changed web tests pass and typecheck passes."
+            },
+            "notes": "Stage completed: implement",
+            "usage": {
+              "input": {
+                "usage": {
+                  "model": {
+                    "provider": "openai",
+                    "model_id": "gpt-5.5"
+                  },
+                  "tokens": {
+                    "input_tokens": 483262,
+                    "output_tokens": 43089,
+                    "reasoning_tokens": 20029,
+                    "cache_read_tokens": 63283712,
+                    "cache_write_tokens": 0
+                  }
+                },
+                "facts": {
+                  "algorithm": "openai"
+                }
+              },
+              "total_usd_micros": 35951706
+            },
+            "files_touched": [
+              "/home/daytona/workspace/fabro/lib/crates/fabro-agent/src/question_tools.rs",
+              "/home/daytona/workspace/fabro/lib/crates/fabro-workflow/src/interview_runtime.rs"
+            ]
+          },
+          "preflight_compile": {
+            "status": "succeeded",
+            "context_updates": {
+              "command.output": "blob://sha256/12ae32cb1ec02d01eda3581b127c1fee3b0dc53572ed6baf239721a03d82e126"
+            },
+            "notes": "Script completed: cargo check -q --workspace 2>&1",
+            "usage": null
+          },
+          "start": {
+            "status": "succeeded",
+            "usage": null
+          },
+          "toolchain": {
+            "status": "succeeded",
+            "context_updates": {
+              "command.output": "blob://sha256/fc14b2ba2d770e5cd3169df7a29525c962adfc4cfa3097b9098c63ebd61a748c"
+            },
+            "notes": "Script completed: command -v cargo >/dev/null || { curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y && sudo ln -sf $HOME/.cargo/bin/* /usr/local/bin/; }; cargo --version 2>&1",
+            "usage": null
+          }
+        },
+        "next_node_id": "simplify_opus",
+        "git_commit_sha": "61317d13c1ac536f9d663790c11cb95d8c7de458",
+        "node_visits": {
+          "implement": 1,
+          "preflight_compile": 1,
+          "start": 1,
+          "toolchain": 1,
+          "preflight_lint": 1
+        }
+      },
+      "diff": {
+        "patch": "diff --git a/apps/fabro-web/app/components/interview-dock.test.tsx b/apps/fabro-web/app/components/interview-dock.test.tsx\nindex 34a58269f..6fb48bfb9 100644\n--- a/apps/fabro-web/app/components/interview-dock.test.tsx\n+++ b/apps/fabro-web/app/components/interview-dock.test.tsx\n@@ -133,6 +133,27 @@ describe(\"InterviewDock\", () => {\n     expect(buttons.Revise).toBeDefined();\n   });\n \n+  test(\"multiple choice renders option descriptions as display text\", () => {\n+    const question = makeQuestion({\n+      question_type: QuestionType.MULTIPLE_CHOICE,\n+      options: [\n+        {\n+          key: \"A\",\n+          label: \"[A] Approve\",\n+          description: \"Deploy the current patch\",\n+          preview: \"<b>not rendered specially</b>\",\n+        },\n+      ],\n+    });\n+    const tree = render(\n+      <InterviewDock runId=\"run-1\" questions={[question]} />,\n+    );\n+    const text = textContent(tree.toJSON());\n+    expect(text).toContain(\"Approve\");\n+    expect(text).toContain(\"Deploy the current patch\");\n+    expect(text).not.toContain(\"<b>not rendered specially</b>\");\n+  });\n+\n   test(\"freeform question renders a textarea and disables send when empty\", () => {\n     const question = makeQuestion({\n       question_type: QuestionType.FREEFORM,\ndiff --git a/apps/fabro-web/app/components/interview-dock.tsx b/apps/fabro-web/app/components/interview-dock.tsx\nindex 35eb2fcf5..967044907 100644\n--- a/apps/fabro-web/app/components/interview-dock.tsx\n+++ b/apps/fabro-web/app/components/interview-dock.tsx\n@@ -15,7 +15,7 @@ import {\n import { QuestionType } from \"@qltysh/fabro-api-client\";\n import type {\n   ApiQuestion,\n-  ApiQuestionOption,\n+  InterviewOption,\n } from \"@qltysh/fabro-api-client\";\n \n import {\n@@ -270,7 +270,7 @@ function ChoiceBody({\n   submitting,\n   onSubmit,\n }: {\n-  options: ApiQuestionOption[];\n+  options: InterviewOption[];\n   allowFreeform: boolean;\n   submitting: boolean;\n   onSubmit: (answer: SubmitInterviewAnswer) => Promise<void>;\n@@ -287,7 +287,7 @@ function ChoiceBody({\n               onClick={() => void onSubmit({ kind: \"selected\", option_key: option.key })}\n               className={CHOICE_BUTTON}\n             >\n-              {displayLabel(option.label)}\n+              <OptionLabel option={option} />\n             </button>\n           ))}\n         </div>\n@@ -314,7 +314,7 @@ function MultiSelectBody({\n   submitting,\n   onSubmit,\n }: {\n-  options: ApiQuestionOption[];\n+  options: InterviewOption[];\n   submitting: boolean;\n   onSubmit: (answer: SubmitInterviewAnswer) => Promise<void>;\n }) {\n@@ -346,7 +346,7 @@ function MultiSelectBody({\n               className={isSelected ? CHOICE_BUTTON_SELECTED : CHOICE_BUTTON}\n             >\n               {isSelected && <CheckIcon className=\"size-3.5\" aria-hidden=\"true\" />}\n-              {displayLabel(option.label)}\n+              <OptionLabel option={option} />\n             </button>\n           );\n         })}\n@@ -455,6 +455,19 @@ function FreeformBody({\n   );\n }\n \n+function OptionLabel({ option }: { option: InterviewOption }) {\n+  return (\n+    <span className=\"text-left\">\n+      <span className=\"block\">{displayLabel(option.label)}</span>\n+      {option.description && (\n+        <span className=\"mt-0.5 block text-xs/5 font-normal text-fg-muted\">\n+          {option.description}\n+        </span>\n+      )}\n+    </span>\n+  );\n+}\n+\n function Spinner() {\n   return <ArrowPathIcon className=\"size-4 animate-spin\" aria-hidden=\"true\" />;\n }\ndiff --git a/apps/fabro-web/app/components/stage-renderers/helpers.test.ts b/apps/fabro-web/app/components/stage-renderers/helpers.test.ts\nindex 6c03541c0..8426ae141 100644\n--- a/apps/fabro-web/app/components/stage-renderers/helpers.test.ts\n+++ b/apps/fabro-web/app/components/stage-renderers/helpers.test.ts\n@@ -79,6 +79,36 @@ describe(\"parseHumanInterviewPairs\", () => {\n     expect(pairs[0].resolution).toBeNull();\n   });\n \n+  test(\"preserves option description and preview metadata from started events\", () => {\n+    const events: EventEnvelope[] = [\n+      envelope(1, {\n+        event: \"interview.started\",\n+        properties: {\n+          question_id: \"q-1\",\n+          question: \"Pick a path\",\n+          question_type: \"multiple_choice\",\n+          options: [\n+            {\n+              key: \"ship\",\n+              label: \"Ship\",\n+              description: \"Deploy the current patch\",\n+              preview: \"diff preview\",\n+            },\n+          ],\n+        },\n+      }),\n+    ];\n+\n+    const pairs = parseHumanInterviewPairs(events);\n+\n+    expect(pairs[0].question.options[0]).toEqual({\n+      key: \"ship\",\n+      label: \"Ship\",\n+      description: \"Deploy the current patch\",\n+      preview: \"diff preview\",\n+    });\n+  });\n+\n   test(\"captures timeout and interrupted resolutions\", () => {\n     const events: EventEnvelope[] = [\n       envelope(1, {\ndiff --git a/apps/fabro-web/app/components/stage-renderers/helpers.ts b/apps/fabro-web/app/components/stage-renderers/helpers.ts\nindex 561779a55..676e372ee 100644\n--- a/apps/fabro-web/app/components/stage-renderers/helpers.ts\n+++ b/apps/fabro-web/app/components/stage-renderers/helpers.ts\n@@ -5,6 +5,8 @@ import { getArray, getNumber, getObject, getString, type UnknownRecord } from \".\n export interface InterviewOption {\n   key: string;\n   label: string;\n+  description?: string | null;\n+  preview?: string | null;\n }\n \n export interface HumanQuestion {\n@@ -54,7 +56,14 @@ function parseInterviewOptions(value: unknown): InterviewOption[] {\n     const record = item as UnknownRecord;\n     const key = getString(record, \"key\");\n     const label = getString(record, \"label\");\n-    if (key && label) out.push({ key, label });\n+    if (key && label) {\n+      const option: InterviewOption = { key, label };\n+      const description = getString(record, \"description\");\n+      const preview = getString(record, \"preview\");\n+      if (description !== null) option.description = description;\n+      if (preview !== null) option.preview = preview;\n+      out.push(option);\n+    }\n   }\n   return out;\n }\ndiff --git a/apps/fabro-web/app/components/stage-renderers/human-qa.tsx b/apps/fabro-web/app/components/stage-renderers/human-qa.tsx\nindex 50dc48cdd..37464ce12 100644\n--- a/apps/fabro-web/app/components/stage-renderers/human-qa.tsx\n+++ b/apps/fabro-web/app/components/stage-renderers/human-qa.tsx\n@@ -184,7 +184,14 @@ function QuestionBlock({\n                 <span className=\"inline-flex size-5 shrink-0 items-center justify-center rounded bg-overlay-strong font-mono text-[11px] text-fg-2\">\n                   {option.key}\n                 </span>\n-                <span className=\"text-sm text-fg-3\">{option.label}</span>\n+                <span className=\"min-w-0 text-sm text-fg-3\">\n+                  <span>{option.label}</span>\n+                  {option.description && (\n+                    <span className=\"mt-0.5 block text-xs/5 text-fg-muted\">\n+                      {option.description}\n+                    </span>\n+                  )}\n+                </span>\n               </li>\n             ))}\n             {question.allowFreeform && (\ndiff --git a/docs/public/api-reference/fabro-api.yaml b/docs/public/api-reference/fabro-api.yaml\nindex 44bfe394c..933a7228e 100644\n--- a/docs/public/api-reference/fabro-api.yaml\n+++ b/docs/public/api-reference/fabro-api.yaml\n@@ -6941,22 +6941,6 @@ components:\n         id:\n           type: string\n \n-    ApiQuestionOption:\n-      description: A selectable option for a multiple-choice or multi-select question.\n-      type: object\n-      required:\n-        - key\n-        - label\n-      properties:\n-        key:\n-          type: string\n-          description: Machine-readable option key used when submitting an answer.\n-          example: option_a\n-        label:\n-          type: string\n-          description: Human-readable label displayed to the user.\n-          example: Accept changes\n-\n     ApiQuestion:\n       description: A pending human-in-the-loop question generated by a workflow stage.\n       type: object\n@@ -6986,7 +6970,7 @@ components:\n           type: array\n           description: Available options for selection-based questions. Empty for freeform questions.\n           items:\n-            $ref: \"#/components/schemas/ApiQuestionOption\"\n+            $ref: \"#/components/schemas/InterviewOption\"\n         allow_freeform:\n           type: boolean\n           description: Whether the user may provide freeform text in addition to selecting options.\n@@ -7758,8 +7742,16 @@ components:\n       properties:\n         key:\n           type: string\n+          description: Machine-readable option key used when submitting an answer.\n         label:\n           type: string\n+          description: Human-readable label displayed to the user.\n+        description:\n+          type: [\"string\", \"null\"]\n+          description: Optional untrusted model-authored option description for display.\n+        preview:\n+          type: [\"string\", \"null\"]\n+          description: Optional untrusted model-authored option preview captured for clients.\n \n     InterviewQuestionRecord:\n       description: Storage shape of an interview question recorded in the event log.\n@@ -11651,4 +11643,4 @@ components:\n         login:\n           type: string\n           description: User's login identifier (e.g. GitHub username).\n-          example: octocat\n\\ No newline at end of file\n+          example: octocat\ndiff --git a/lib/crates/fabro-agent/src/lib.rs b/lib/crates/fabro-agent/src/lib.rs\nindex 04a1c6fdf..010004694 100644\n--- a/lib/crates/fabro-agent/src/lib.rs\n+++ b/lib/crates/fabro-agent/src/lib.rs\n@@ -15,6 +15,7 @@ pub mod loop_detection;\n pub mod mcp_integration;\n pub mod memory;\n pub mod profiles;\n+pub mod question_tools;\n pub mod read_before_write_sandbox;\n pub mod sandbox;\n pub mod session;\n@@ -46,6 +47,11 @@ pub use local_sandbox::LocalSandbox;\n pub use loop_detection::detect_loop;\n pub use memory::{MemoryDocument, discover_memory};\n pub use profiles::{AnthropicProfile, EnvContext, GeminiProfile, OpenAiProfile};\n+pub use question_tools::{\n+    ANTHROPIC_ASK_USER_QUESTION_TOOL, AgentQuestion, AgentQuestionAnswer,\n+    AgentQuestionAnswerStatus, AgentQuestionRuntime, AgentToolRuntime,\n+    OPENAI_REQUEST_USER_INPUT_TOOL, register_question_tools,\n+};\n pub use read_before_write_sandbox::ReadBeforeWriteSandbox;\n pub use sandbox::{\n     CommandOutputCallback, DirEntry, ExecResult, ExecStreamingResult, GrepOptions, Sandbox,\ndiff --git a/lib/crates/fabro-agent/src/question_tools.rs b/lib/crates/fabro-agent/src/question_tools.rs\nnew file mode 100644\nindex 000000000..fd9c7deda\n--- /dev/null\n+++ b/lib/crates/fabro-agent/src/question_tools.rs\n@@ -0,0 +1,586 @@\n+//! Model-native tools that let a root workflow agent ask the human for input.\n+\n+use std::collections::BTreeMap;\n+use std::future::Future;\n+use std::sync::Arc;\n+\n+use async_trait::async_trait;\n+use fabro_llm::types::ToolDefinition;\n+use fabro_model::AgentProfileKind;\n+use fabro_types::{InterviewOption, QuestionType};\n+use serde::Deserialize;\n+use serde_json::json;\n+use tokio_util::sync::CancellationToken;\n+\n+use crate::tool_registry::{RegisteredTool, ToolContext, ToolRegistry};\n+\n+tokio::task_local! {\n+    static CURRENT_AGENT_TOOL_RUNTIME: AgentToolRuntime;\n+}\n+\n+pub const OPENAI_REQUEST_USER_INPUT_TOOL: &str = \"request_user_input\";\n+pub const ANTHROPIC_ASK_USER_QUESTION_TOOL: &str = \"AskUserQuestion\";\n+\n+pub const OPTION_DESCRIPTION_MAX_CHARS: usize = 2_000;\n+pub const OPTION_PREVIEW_MAX_CHARS: usize = 4_000;\n+\n+#[derive(Clone, Default)]\n+pub struct AgentToolRuntime {\n+    question_runtime: Option<Arc<dyn AgentQuestionRuntime>>,\n+}\n+\n+impl AgentToolRuntime {\n+    #[must_use]\n+    pub fn new() -> Self {\n+        Self::default()\n+    }\n+\n+    #[must_use]\n+    pub fn with_question_runtime(runtime: Arc<dyn AgentQuestionRuntime>) -> Self {\n+        Self {\n+            question_runtime: Some(runtime),\n+        }\n+    }\n+\n+    #[must_use]\n+    pub fn question_runtime(&self) -> Option<Arc<dyn AgentQuestionRuntime>> {\n+        self.question_runtime.clone()\n+    }\n+}\n+\n+pub async fn scope_agent_tool_runtime<F>(runtime: AgentToolRuntime, future: F) -> F::Output\n+where\n+    F: Future,\n+{\n+    CURRENT_AGENT_TOOL_RUNTIME.scope(runtime, future).await\n+}\n+\n+fn current_agent_tool_runtime() -> AgentToolRuntime {\n+    CURRENT_AGENT_TOOL_RUNTIME\n+        .try_with(Clone::clone)\n+        .unwrap_or_default()\n+}\n+\n+#[derive(Debug, Clone, PartialEq, Eq)]\n+pub struct AgentQuestion {\n+    pub original_id:       Option<String>,\n+    pub original_question: String,\n+    pub header:            Option<String>,\n+    pub text:              String,\n+    pub question_type:     QuestionType,\n+    pub options:           Vec<InterviewOption>,\n+    pub allow_freeform:    bool,\n+}\n+\n+#[derive(Debug, Clone, Copy, PartialEq, Eq)]\n+pub enum AgentQuestionAnswerStatus {\n+    Answered,\n+    Cancelled,\n+    Interrupted,\n+    Skipped,\n+    Timeout,\n+}\n+\n+#[derive(Debug, Clone, PartialEq, Eq)]\n+pub struct AgentQuestionAnswer {\n+    pub original_id:       Option<String>,\n+    pub original_question: String,\n+    pub answers:           Vec<String>,\n+    pub status:            AgentQuestionAnswerStatus,\n+}\n+\n+#[async_trait]\n+pub trait AgentQuestionRuntime: Send + Sync {\n+    async fn ask_questions(\n+        &self,\n+        tool_call_id: &str,\n+        questions: Vec<AgentQuestion>,\n+        cancel_token: CancellationToken,\n+    ) -> Result<Vec<AgentQuestionAnswer>, String>;\n+}\n+\n+#[derive(Debug, Deserialize)]\n+struct OpenAiQuestionToolArgs {\n+    questions: Vec<OpenAiQuestion>,\n+}\n+\n+#[derive(Debug, Deserialize)]\n+struct OpenAiQuestion {\n+    id:       String,\n+    header:   String,\n+    question: String,\n+    #[serde(default)]\n+    options:  Vec<OpenAiOption>,\n+}\n+\n+#[derive(Debug, Deserialize)]\n+struct OpenAiOption {\n+    label:       String,\n+    #[serde(default)]\n+    description: Option<String>,\n+}\n+\n+#[derive(Debug, Deserialize)]\n+struct AnthropicQuestionToolArgs {\n+    questions: Vec<AnthropicQuestion>,\n+}\n+\n+#[derive(Debug, Deserialize)]\n+#[serde(rename_all = \"camelCase\")]\n+struct AnthropicQuestion {\n+    question:     String,\n+    #[serde(default)]\n+    header:       Option<String>,\n+    #[serde(default)]\n+    options:      Vec<AnthropicOption>,\n+    #[serde(default)]\n+    multi_select: bool,\n+}\n+\n+#[derive(Debug, Deserialize)]\n+struct AnthropicOption {\n+    label:       String,\n+    #[serde(default)]\n+    description: Option<String>,\n+    #[serde(default)]\n+    preview:     Option<String>,\n+}\n+\n+#[must_use]\n+pub fn is_question_tool(name: &str) -> bool {\n+    matches!(\n+        name,\n+        OPENAI_REQUEST_USER_INPUT_TOOL | ANTHROPIC_ASK_USER_QUESTION_TOOL\n+    )\n+}\n+\n+pub fn register_question_tools(profile_kind: AgentProfileKind, registry: &mut ToolRegistry) {\n+    match profile_kind {\n+        AgentProfileKind::OpenAi => registry.register(make_openai_question_tool()),\n+        AgentProfileKind::Anthropic => registry.register(make_anthropic_question_tool()),\n+        AgentProfileKind::Gemini => {}\n+    }\n+}\n+\n+fn make_openai_question_tool() -> RegisteredTool {\n+    RegisteredTool {\n+        definition: ToolDefinition {\n+            name:        OPENAI_REQUEST_USER_INPUT_TOOL.to_string(),\n+            description: \"Ask the human one or more questions and wait for their answers before continuing this stage.\".to_string(),\n+            parameters:  json!({\n+                \"type\": \"object\",\n+                \"required\": [\"questions\"],\n+                \"properties\": {\n+                    \"questions\": {\n+                        \"type\": \"array\",\n+                        \"minItems\": 1,\n+                        \"items\": {\n+                            \"type\": \"object\",\n+                            \"required\": [\"id\", \"header\", \"question\", \"options\"],\n+                            \"properties\": {\n+                                \"id\": { \"type\": \"string\" },\n+                                \"header\": { \"type\": \"string\" },\n+                                \"question\": { \"type\": \"string\" },\n+                                \"options\": {\n+                                    \"type\": \"array\",\n+                                    \"items\": {\n+                                        \"type\": \"object\",\n+                                        \"required\": [\"label\"],\n+                                        \"properties\": {\n+                                            \"label\": { \"type\": \"string\" },\n+                                            \"description\": { \"type\": \"string\" }\n+                                        }\n+                                    }\n+                                }\n+                            }\n+                        }\n+                    }\n+                }\n+            }),\n+        },\n+        executor:   Arc::new(|args, ctx| {\n+            Box::pin(async move {\n+                let parsed: OpenAiQuestionToolArgs = parse_tool_args(args)?;\n+                let questions = normalize_openai_questions(parsed)?;\n+                let answers = execute_question_tool(ctx, questions).await?;\n+                format_openai_answers(&answers)\n+            })\n+        }),\n+    }\n+}\n+\n+fn make_anthropic_question_tool() -> RegisteredTool {\n+    RegisteredTool {\n+        definition: ToolDefinition {\n+            name:        ANTHROPIC_ASK_USER_QUESTION_TOOL.to_string(),\n+            description: \"Ask the human one or more questions and wait for their answers before continuing this stage.\".to_string(),\n+            parameters:  json!({\n+                \"type\": \"object\",\n+                \"required\": [\"questions\"],\n+                \"properties\": {\n+                    \"questions\": {\n+                        \"type\": \"array\",\n+                        \"minItems\": 1,\n+                        \"items\": {\n+                            \"type\": \"object\",\n+                            \"required\": [\"question\", \"options\", \"multiSelect\"],\n+                            \"properties\": {\n+                                \"question\": { \"type\": \"string\" },\n+                                \"header\": { \"type\": \"string\" },\n+                                \"options\": {\n+                                    \"type\": \"array\",\n+                                    \"items\": {\n+                                        \"type\": \"object\",\n+                                        \"required\": [\"label\"],\n+                                        \"properties\": {\n+                                            \"label\": { \"type\": \"string\" },\n+                                            \"description\": { \"type\": \"string\" },\n+                                            \"preview\": { \"type\": \"string\" }\n+                                        }\n+                                    }\n+                                },\n+                                \"multiSelect\": { \"type\": \"boolean\" }\n+                            }\n+                        }\n+                    }\n+                }\n+            }),\n+        },\n+        executor:   Arc::new(|args, ctx| {\n+            Box::pin(async move {\n+                let parsed: AnthropicQuestionToolArgs = parse_tool_args(args)?;\n+                let questions = normalize_anthropic_questions(parsed)?;\n+                let answers = execute_question_tool(ctx, questions).await?;\n+                format_anthropic_answers(&answers)\n+            })\n+        }),\n+    }\n+}\n+\n+fn parse_tool_args<T: for<'de> Deserialize<'de>>(args: serde_json::Value) -> Result<T, String> {\n+    serde_json::from_value(args).map_err(|err| format!(\"invalid question tool arguments: {err}\"))\n+}\n+\n+async fn execute_question_tool(\n+    ctx: ToolContext,\n+    questions: Vec<AgentQuestion>,\n+) -> Result<Vec<AgentQuestionAnswer>, String> {\n+    let session_id = ctx.session_id.as_deref().ok_or_else(|| {\n+        \"human-question tools are available only during a root workflow agent session\".to_string()\n+    })?;\n+    let root_session_id = ctx.root_session_id.as_deref().ok_or_else(|| {\n+        \"human-question tools are available only during a root workflow agent session\".to_string()\n+    })?;\n+    if session_id != root_session_id {\n+        return Err(\n+            \"human-question tools are only available to the root agent; subagents must report back to their parent\".to_string(),\n+        );\n+    }\n+    let tool_call_id = ctx\n+        .tool_call_id\n+        .as_deref()\n+        .ok_or_else(|| \"human-question tool call is missing a provider tool_call_id\".to_string())?;\n+    let runtime = current_agent_tool_runtime().question_runtime().ok_or_else(|| {\n+        \"human-question tools are available only inside a workflow run with an active interviewer\".to_string()\n+    })?;\n+    runtime\n+        .ask_questions(tool_call_id, questions, ctx.cancel.clone())\n+        .await\n+}\n+\n+fn normalize_openai_questions(args: OpenAiQuestionToolArgs) -> Result<Vec<AgentQuestion>, String> {\n+    if args.questions.is_empty() {\n+        return Err(\"questions must contain at least one question\".to_string());\n+    }\n+    args.questions\n+        .into_iter()\n+        .map(|question| {\n+            let original_question = question.question.trim().to_string();\n+            Ok(AgentQuestion {\n+                original_id: Some(non_empty(&question.id, \"question id\")?),\n+                text: display_text(Some(question.header.as_str()), &question.question),\n+                header: Some(question.header),\n+                original_question,\n+                question_type: QuestionType::MultipleChoice,\n+                options: options_from_openai(question.options),\n+                allow_freeform: true,\n+            })\n+        })\n+        .collect()\n+}\n+\n+fn normalize_anthropic_questions(\n+    args: AnthropicQuestionToolArgs,\n+) -> Result<Vec<AgentQuestion>, String> {\n+    if args.questions.is_empty() {\n+        return Err(\"questions must contain at least one question\".to_string());\n+    }\n+    args.questions\n+        .into_iter()\n+        .map(|question| {\n+            let original_question = non_empty(&question.question, \"question\")?;\n+            Ok(AgentQuestion {\n+                original_id: None,\n+                text: display_text(question.header.as_deref(), &question.question),\n+                header: question.header,\n+                original_question,\n+                question_type: if question.multi_select {\n+                    QuestionType::MultiSelect\n+                } else {\n+                    QuestionType::MultipleChoice\n+                },\n+                options: options_from_anthropic(question.options),\n+                allow_freeform: true,\n+            })\n+        })\n+        .collect()\n+}\n+\n+fn options_from_openai(options: Vec<OpenAiOption>) -> Vec<InterviewOption> {\n+    options\n+        .into_iter()\n+        .enumerate()\n+        .map(|(idx, option)| InterviewOption {\n+            key:         option_key(idx),\n+            label:       option.label,\n+            description: option\n+                .description\n+                .map(|value| bounded_display_field(&value, OPTION_DESCRIPTION_MAX_CHARS)),\n+            preview:     None,\n+        })\n+        .collect()\n+}\n+\n+fn options_from_anthropic(options: Vec<AnthropicOption>) -> Vec<InterviewOption> {\n+    options\n+        .into_iter()\n+        .enumerate()\n+        .map(|(idx, option)| InterviewOption {\n+            key:         option_key(idx),\n+            label:       option.label,\n+            description: option\n+                .description\n+                .map(|value| bounded_display_field(&value, OPTION_DESCRIPTION_MAX_CHARS)),\n+            preview:     option\n+                .preview\n+                .map(|value| bounded_display_field(&value, OPTION_PREVIEW_MAX_CHARS)),\n+        })\n+        .collect()\n+}\n+\n+fn option_key(idx: usize) -> String {\n+    format!(\"option_{}\", idx + 1)\n+}\n+\n+fn non_empty(value: &str, field: &str) -> Result<String, String> {\n+    let trimmed = value.trim();\n+    if trimmed.is_empty() {\n+        Err(format!(\"{field} must not be empty\"))\n+    } else {\n+        Ok(trimmed.to_string())\n+    }\n+}\n+\n+fn display_text(header: Option<&str>, question: &str) -> String {\n+    let header = header.map(str::trim).filter(|value| !value.is_empty());\n+    let question = question.trim();\n+    match (header, question.is_empty()) {\n+        (Some(header), false) => format!(\"{header}\\n\\n{question}\"),\n+        (Some(header), true) => header.to_string(),\n+        (None, false) => question.to_string(),\n+        (None, true) => String::new(),\n+    }\n+}\n+\n+#[must_use]\n+pub fn bounded_display_field(value: &str, max_chars: usize) -> String {\n+    if value.chars().count() <= max_chars {\n+        return value.to_string();\n+    }\n+    value.chars().take(max_chars).collect()\n+}\n+\n+fn ensure_all_answered(answers: &[AgentQuestionAnswer]) -> Result<(), String> {\n+    if let Some(answer) = answers\n+        .iter()\n+        .find(|answer| answer.status != AgentQuestionAnswerStatus::Answered)\n+    {\n+        return Err(format!(\n+            \"human-question request ended before the user answered `{}`: {}\",\n+            answer.original_question,\n+            answer_status_label(answer.status)\n+        ));\n+    }\n+    Ok(())\n+}\n+\n+fn answer_status_label(status: AgentQuestionAnswerStatus) -> &'static str {\n+    match status {\n+        AgentQuestionAnswerStatus::Answered => \"answered\",\n+        AgentQuestionAnswerStatus::Cancelled => \"cancelled\",\n+        AgentQuestionAnswerStatus::Interrupted => \"interrupted\",\n+        AgentQuestionAnswerStatus::Skipped => \"skipped\",\n+        AgentQuestionAnswerStatus::Timeout => \"timed out\",\n+    }\n+}\n+\n+fn format_openai_answers(answers: &[AgentQuestionAnswer]) -> Result<String, String> {\n+    ensure_all_answered(answers)?;\n+    let mut answer_map = BTreeMap::new();\n+    for answer in answers {\n+        let Some(original_id) = answer.original_id.as_ref() else {\n+            return Err(\n+                \"OpenAI question answer is missing the original model question id\".to_string(),\n+            );\n+        };\n+        answer_map.insert(original_id.clone(), json!({ \"answers\": answer.answers }));\n+    }\n+    serde_json::to_string(&json!({ \"answers\": answer_map }))\n+        .map_err(|err| format!(\"failed to serialize answers: {err}\"))\n+}\n+\n+fn format_anthropic_answers(answers: &[AgentQuestionAnswer]) -> Result<String, String> {\n+    ensure_all_answered(answers)?;\n+    let pairs = answers\n+        .iter()\n+        .map(|answer| {\n+            let question = serde_json::to_string(&answer.original_question)\n+                .unwrap_or_else(|_| \"\\\"question\\\"\".to_string());\n+            let answer_text = serde_json::to_string(&answer.answers.join(\", \"))\n+                .unwrap_or_else(|_| \"\\\"\\\"\".to_string());\n+            format!(\"{question}={answer_text}\")\n+        })\n+        .collect::<Vec<_>>()\n+        .join(\", \");\n+    Ok(format!(\n+        \"User has answered your questions: {pairs}. You can now continue with the task.\"\n+    ))\n+}\n+\n+#[cfg(test)]\n+mod tests {\n+    use super::*;\n+\n+    fn answered(\n+        original_id: Option<&str>,\n+        question: &str,\n+        answers: &[&str],\n+    ) -> AgentQuestionAnswer {\n+        AgentQuestionAnswer {\n+            original_id:       original_id.map(str::to_string),\n+            original_question: question.to_string(),\n+            answers:           answers.iter().map(|value| (*value).to_string()).collect(),\n+            status:            AgentQuestionAnswerStatus::Answered,\n+        }\n+    }\n+\n+    #[test]\n+    fn openai_request_with_descriptions_normalizes_to_multiple_choice() {\n+        let args: OpenAiQuestionToolArgs = serde_json::from_value(json!({\n+            \"questions\": [{\n+                \"id\": \"q1\",\n+                \"header\": \"Decision\",\n+                \"question\": \"Which path?\",\n+                \"options\": [{ \"label\": \"Ship\", \"description\": \"Deploy now\" }]\n+            }]\n+        }))\n+        .unwrap();\n+\n+        let questions = normalize_openai_questions(args).unwrap();\n+\n+        assert_eq!(questions.len(), 1);\n+        assert_eq!(questions[0].original_id.as_deref(), Some(\"q1\"));\n+        assert_eq!(questions[0].question_type, QuestionType::MultipleChoice);\n+        assert!(questions[0].allow_freeform);\n+        assert_eq!(questions[0].text, \"Decision\\n\\nWhich path?\");\n+        assert_eq!(questions[0].options[0].key, \"option_1\");\n+        assert_eq!(questions[0].options[0].label, \"Ship\");\n+        assert_eq!(\n+            questions[0].options[0].description.as_deref(),\n+            Some(\"Deploy now\")\n+        );\n+    }\n+\n+    #[test]\n+    fn anthropic_multiselect_preserves_preview_and_formats_comma_joined_answers() {\n+        let args: AnthropicQuestionToolArgs = serde_json::from_value(json!({\n+            \"questions\": [{\n+                \"header\": \"Pick features\",\n+                \"question\": \"Which features?\",\n+                \"multiSelect\": true,\n+                \"options\": [{\n+                    \"label\": \"Auth\",\n+                    \"description\": \"Login support\",\n+                    \"preview\": \"auth diff\"\n+                }]\n+            }]\n+        }))\n+        .unwrap();\n+\n+        let questions = normalize_anthropic_questions(args).unwrap();\n+\n+        assert_eq!(questions[0].question_type, QuestionType::MultiSelect);\n+        assert_eq!(\n+            questions[0].options[0].preview.as_deref(),\n+            Some(\"auth diff\")\n+        );\n+        let text =\n+            format_anthropic_answers(&[answered(None, \"Which features?\", &[\"Auth\", \"Billing\"])])\n+                .unwrap();\n+        assert!(text.contains(\"\\\"Which features?\\\"=\\\"Auth, Billing\\\"\"));\n+    }\n+\n+    #[test]\n+    fn openai_answers_are_keyed_by_original_model_question_id() {\n+        let text = format_openai_answers(&[\n+            answered(Some(\"first\"), \"First?\", &[\"Yes\"]),\n+            answered(Some(\"second\"), \"Second?\", &[\"No\"]),\n+        ])\n+        .unwrap();\n+\n+        assert_eq!(\n+            serde_json::from_str::<serde_json::Value>(&text).unwrap(),\n+            json!({\n+                \"answers\": {\n+                    \"first\": { \"answers\": [\"Yes\"] },\n+                    \"second\": { \"answers\": [\"No\"] }\n+                }\n+            })\n+        );\n+    }\n+\n+    #[test]\n+    fn option_description_and_preview_are_bounded() {\n+        let long = \"x\".repeat(OPTION_PREVIEW_MAX_CHARS + 10);\n+\n+        assert_eq!(\n+            bounded_display_field(&long, OPTION_DESCRIPTION_MAX_CHARS)\n+                .chars()\n+                .count(),\n+            OPTION_DESCRIPTION_MAX_CHARS\n+        );\n+        assert_eq!(\n+            bounded_display_field(&long, OPTION_PREVIEW_MAX_CHARS)\n+                .chars()\n+                .count(),\n+            OPTION_PREVIEW_MAX_CHARS\n+        );\n+    }\n+\n+    #[test]\n+    fn question_tool_registration_is_profile_specific() {\n+        let mut openai = ToolRegistry::new();\n+        register_question_tools(AgentProfileKind::OpenAi, &mut openai);\n+        assert!(openai.get(OPENAI_REQUEST_USER_INPUT_TOOL).is_some());\n+        assert!(openai.get(ANTHROPIC_ASK_USER_QUESTION_TOOL).is_none());\n+\n+        let mut anthropic = ToolRegistry::new();\n+        register_question_tools(AgentProfileKind::Anthropic, &mut anthropic);\n+        assert!(anthropic.get(ANTHROPIC_ASK_USER_QUESTION_TOOL).is_some());\n+        assert!(anthropic.get(OPENAI_REQUEST_USER_INPUT_TOOL).is_none());\n+\n+        let mut gemini = ToolRegistry::new();\n+        register_question_tools(AgentProfileKind::Gemini, &mut gemini);\n+        assert!(gemini.names().is_empty());\n+    }\n+}\ndiff --git a/lib/crates/fabro-agent/src/session.rs b/lib/crates/fabro-agent/src/session.rs\nindex 2d0d0b4e8..0d4fd749f 100644\n--- a/lib/crates/fabro-agent/src/session.rs\n+++ b/lib/crates/fabro-agent/src/session.rs\n@@ -32,6 +32,7 @@ use crate::history::History;\n use crate::loop_detection::detect_loop;\n use crate::memory::{BUDGET_BYTES, MemoryDocument, discover_memory};\n use crate::profiles::EnvContext;\n+use crate::question_tools::AgentToolRuntime;\n use crate::sandbox::Sandbox;\n use crate::skills::{\n     ExpandedInput, Skill, default_skill_dirs, discover_skills, expand_skill, make_use_skill_tool,\n@@ -1119,6 +1120,15 @@ impl Session {\n     }\n \n     pub async fn process_input(&mut self, input: &str) -> Result<(), Error> {\n+        self.process_input_with_runtime(input, AgentToolRuntime::default())\n+            .await\n+    }\n+\n+    pub async fn process_input_with_runtime(\n+        &mut self,\n+        input: &str,\n+        agent_tool_runtime: AgentToolRuntime,\n+    ) -> Result<(), Error> {\n         if self.state == SessionState::Closed {\n             return Err(Error::SessionClosed);\n         }\n@@ -1142,7 +1152,7 @@ impl Session {\n         });\n \n         // Process the initial input, then drain any followups\n-        let mut result = self.run_single_input(input).await;\n+        let mut result = self.run_single_input(input, &agent_tool_runtime).await;\n \n         if result.is_ok() {\n             loop {\n@@ -1152,7 +1162,7 @@ impl Session {\n                     .expect(\"followup queue lock poisoned\")\n                     .pop_front();\n                 let Some(followup) = followup else { break };\n-                result = self.run_single_input(&followup).await;\n+                result = self.run_single_input(&followup, &agent_tool_runtime).await;\n                 if result.is_err() {\n                     break;\n                 }\n@@ -1172,7 +1182,11 @@ impl Session {\n         result\n     }\n \n-    async fn run_single_input(&mut self, input: &str) -> Result<(), Error> {\n+    async fn run_single_input(\n+        &mut self,\n+        input: &str,\n+        agent_tool_runtime: &AgentToolRuntime,\n+    ) -> Result<(), Error> {\n         const STREAM_CONSUME_RETRIES: usize = 3;\n \n         if self.state == SessionState::Closed {\n@@ -1623,6 +1637,7 @@ impl Session {\n                 &self.id,\n                 &self.root_session_id,\n                 self.tool_env_provider.as_ref(),\n+                agent_tool_runtime,\n             )\n             .await;\n             composite_watcher.abort();\ndiff --git a/lib/crates/fabro-agent/src/tool_execution.rs b/lib/crates/fabro-agent/src/tool_execution.rs\nindex 939e6a4e9..3015b6147 100644\n--- a/lib/crates/fabro-agent/src/tool_execution.rs\n+++ b/lib/crates/fabro-agent/src/tool_execution.rs\n@@ -7,6 +7,7 @@ use tracing::debug;\n \n use crate::config::{SessionOptions, ToolHookCallback, ToolHookDecision};\n use crate::event::{Emitter, SessionBoundEmitter};\n+use crate::question_tools::{self, AgentToolRuntime, is_question_tool};\n use crate::sandbox::Sandbox;\n use crate::session::ToolEnvProvider;\n use crate::tool_registry::{AgentEventEmitter, RegisteredTool, ToolContext, ToolRegistry};\n@@ -31,7 +32,25 @@ pub async fn execute_tool_calls(\n     session_id: &str,\n     root_session_id: &str,\n     tool_env_provider: Option<&Arc<dyn ToolEnvProvider>>,\n+    agent_tool_runtime: &AgentToolRuntime,\n ) -> Vec<ToolResult> {\n+    if tool_calls.iter().any(|tc| is_question_tool(&tc.name)) {\n+        return execute_question_tool_round(\n+            tool_calls,\n+            registry,\n+            env,\n+            tool_hooks,\n+            cancel_token,\n+            config,\n+            emitter,\n+            session_id,\n+            root_session_id,\n+            tool_env_provider,\n+            agent_tool_runtime,\n+        )\n+        .await;\n+    }\n+\n     if parallel && tool_calls.len() > 1 {\n         execute_tool_calls_parallel(\n             tool_calls,\n@@ -44,6 +63,7 @@ pub async fn execute_tool_calls(\n             session_id,\n             root_session_id,\n             tool_env_provider,\n+            agent_tool_runtime,\n         )\n         .await\n     } else {\n@@ -58,6 +78,7 @@ pub async fn execute_tool_calls(\n             session_id,\n             root_session_id,\n             tool_env_provider,\n+            agent_tool_runtime,\n         )\n         .await\n     }\n@@ -78,6 +99,7 @@ async fn execute_tool_calls_sequential(\n     session_id: &str,\n     root_session_id: &str,\n     tool_env_provider: Option<&Arc<dyn ToolEnvProvider>>,\n+    agent_tool_runtime: &AgentToolRuntime,\n ) -> Vec<ToolResult> {\n     let mut results = Vec::new();\n     for tc in tool_calls {\n@@ -86,7 +108,7 @@ async fn execute_tool_calls_sequential(\n             continue;\n         }\n \n-        let result = execute_and_emit_one_tool(\n+        let result = execute_and_emit_one_tool_with_runtime(\n             tc,\n             registry,\n             env.clone(),\n@@ -97,6 +119,7 @@ async fn execute_tool_calls_sequential(\n             session_id,\n             root_session_id,\n             tool_env_provider,\n+            agent_tool_runtime,\n         )\n         .await;\n         results.push(result);\n@@ -119,8 +142,10 @@ async fn execute_tool_calls_parallel(\n     session_id: &str,\n     root_session_id: &str,\n     tool_env_provider: Option<&Arc<dyn ToolEnvProvider>>,\n+    agent_tool_runtime: &AgentToolRuntime,\n ) -> Vec<ToolResult> {\n     let tool_env_provider = tool_env_provider.cloned();\n+    let agent_tool_runtime = agent_tool_runtime.clone();\n     let futures: Vec<_> = tool_calls\n         .iter()\n         .map(|tc| {\n@@ -133,6 +158,7 @@ async fn execute_tool_calls_parallel(\n             let root_session_id = root_session_id.to_owned();\n             let tool_hooks = tool_hooks.cloned();\n             let tool_env_provider = tool_env_provider.clone();\n+            let agent_tool_runtime = agent_tool_runtime.clone();\n             let access_denial = config.tool_access_denial_reason(&tc.name);\n             // Look up the tool before spawning since ToolRegistry is not Send.\n             let registered_tool = if access_denial.is_none() {\n@@ -153,6 +179,7 @@ async fn execute_tool_calls_parallel(\n                     &session_id,\n                     &root_session_id,\n                     tool_env_provider.as_ref(),\n+                    &agent_tool_runtime,\n                 )\n                 .await\n             }\n@@ -162,6 +189,99 @@ async fn execute_tool_calls_parallel(\n     future::join_all(futures).await\n }\n \n+#[allow(\n+    clippy::too_many_arguments,\n+    reason = \"Question-tool round handling needs the same execution context as normal tool dispatch.\"\n+)]\n+async fn execute_question_tool_round(\n+    tool_calls: &[ToolCall],\n+    registry: &ToolRegistry,\n+    env: Arc<dyn Sandbox>,\n+    tool_hooks: Option<&Arc<dyn ToolHookCallback>>,\n+    cancel_token: &CancellationToken,\n+    config: &SessionOptions,\n+    emitter: &Emitter,\n+    session_id: &str,\n+    root_session_id: &str,\n+    tool_env_provider: Option<&Arc<dyn ToolEnvProvider>>,\n+    agent_tool_runtime: &AgentToolRuntime,\n+) -> Vec<ToolResult> {\n+    let first_question_index = tool_calls\n+        .iter()\n+        .position(|tc| is_question_tool(&tc.name))\n+        .expect(\"question-tool round should contain a question tool\");\n+    let mut results = Vec::with_capacity(tool_calls.len());\n+\n+    for (index, tc) in tool_calls.iter().enumerate() {\n+        if cancel_token.is_cancelled() {\n+            results.push(ToolResult::error(tc.id.clone(), \"Cancelled\"));\n+            continue;\n+        }\n+\n+        if index == first_question_index {\n+            results.push(\n+                execute_and_emit_one_tool_with_runtime(\n+                    tc,\n+                    registry,\n+                    env.clone(),\n+                    tool_hooks,\n+                    cancel_token.child_token(),\n+                    config,\n+                    emitter,\n+                    session_id,\n+                    root_session_id,\n+                    tool_env_provider,\n+                    agent_tool_runtime,\n+                )\n+                .await,\n+            );\n+        } else if is_question_tool(&tc.name) {\n+            results.push(error_tool_result_with_events(\n+                tc,\n+                emitter,\n+                session_id,\n+                config,\n+                \"Only one human-question tool call may be used in a tool round. Combine all questions into a single questions[] batch and call the question tool once.\",\n+            ));\n+        } else {\n+            results.push(error_tool_result_with_events(\n+                tc,\n+                emitter,\n+                session_id,\n+                config,\n+                \"This tool call was not executed because human-question tools must run alone in a tool round. Retry non-question tools in a later round after the user answers.\",\n+            ));\n+        }\n+    }\n+\n+    results\n+}\n+\n+fn error_tool_result_with_events(\n+    tc: &ToolCall,\n+    emitter: &Emitter,\n+    session_id: &str,\n+    config: &SessionOptions,\n+    message: &str,\n+) -> ToolResult {\n+    emitter.emit(session_id.to_owned(), AgentEvent::ToolCallStarted {\n+        tool_name:    tc.name.clone(),\n+        tool_call_id: tc.id.clone(),\n+        arguments:    tc.arguments.clone(),\n+    });\n+    let result = ToolResult::error(&tc.id, message);\n+    emitter.emit(session_id.to_owned(), AgentEvent::ToolCallOutputDelta {\n+        delta: result.content.to_string(),\n+    });\n+    emitter.emit(session_id.to_owned(), AgentEvent::ToolCallCompleted {\n+        tool_name:    tc.name.clone(),\n+        tool_call_id: tc.id.clone(),\n+        output:       result.content.clone(),\n+        is_error:     true,\n+    });\n+    truncate_tool_result(&result, &tc.name, config)\n+}\n+\n /// Execute a single tool call with event emission and output truncation.\n #[allow(\n     clippy::too_many_arguments,\n@@ -178,6 +298,39 @@ pub async fn execute_and_emit_one_tool(\n     session_id: &str,\n     root_session_id: &str,\n     tool_env_provider: Option<&Arc<dyn ToolEnvProvider>>,\n+) -> ToolResult {\n+    execute_and_emit_one_tool_with_runtime(\n+        tc,\n+        registry,\n+        env,\n+        tool_hooks,\n+        cancel_token,\n+        config,\n+        emitter,\n+        session_id,\n+        root_session_id,\n+        tool_env_provider,\n+        &AgentToolRuntime::default(),\n+    )\n+    .await\n+}\n+\n+#[allow(\n+    clippy::too_many_arguments,\n+    reason = \"Single-tool execution needs the tool, runtime handles, and emission context.\"\n+)]\n+async fn execute_and_emit_one_tool_with_runtime(\n+    tc: &ToolCall,\n+    registry: &ToolRegistry,\n+    env: Arc<dyn Sandbox>,\n+    tool_hooks: Option<&Arc<dyn ToolHookCallback>>,\n+    cancel_token: CancellationToken,\n+    config: &SessionOptions,\n+    emitter: &Emitter,\n+    session_id: &str,\n+    root_session_id: &str,\n+    tool_env_provider: Option<&Arc<dyn ToolEnvProvider>>,\n+    agent_tool_runtime: &AgentToolRuntime,\n ) -> ToolResult {\n     let access_denial = config.tool_access_denial_reason(&tc.name);\n     let registered_tool = if access_denial.is_none() {\n@@ -197,6 +350,7 @@ pub async fn execute_and_emit_one_tool(\n         session_id,\n         root_session_id,\n         tool_env_provider,\n+        agent_tool_runtime,\n     )\n     .await\n }\n@@ -219,6 +373,7 @@ async fn execute_and_emit_one_tool_with_lookup(\n     session_id: &str,\n     root_session_id: &str,\n     tool_env_provider: Option<&Arc<dyn ToolEnvProvider>>,\n+    agent_tool_runtime: &AgentToolRuntime,\n ) -> ToolResult {\n     emitter.emit(session_id.to_owned(), AgentEvent::ToolCallStarted {\n         tool_name:    tc.name.clone(),\n@@ -276,6 +431,7 @@ async fn execute_and_emit_one_tool_with_lookup(\n         session_id,\n         root_session_id,\n         tool_env_provider,\n+        agent_tool_runtime,\n     )\n     .await;\n \n@@ -329,6 +485,7 @@ async fn execute_one_tool(\n     session_id: &str,\n     root_session_id: &str,\n     tool_env_provider: Option<&Arc<dyn ToolEnvProvider>>,\n+    agent_tool_runtime: &AgentToolRuntime,\n ) -> ToolResult {\n     match registered_tool {\n         Some(tool) => {\n@@ -355,7 +512,10 @@ async fn execute_one_tool(\n                 tool_call_id: Some(tc.id.clone()),\n                 agent_event_emitter,\n             };\n-            match (tool.executor)(tc.arguments.clone(), ctx).await {\n+            let execution = (tool.executor)(tc.arguments.clone(), ctx);\n+            match question_tools::scope_agent_tool_runtime(agent_tool_runtime.clone(), execution)\n+                .await\n+            {\n                 Ok(output) => ToolResult::success(&tc.id, serde_json::json!(output)),\n                 Err(err) => ToolResult::error(&tc.id, err),\n             }\n@@ -418,9 +578,11 @@ pub fn validate_tool_args(\n #[cfg(test)]\n mod tests {\n     use std::collections::HashMap;\n-    use std::sync::Mutex;\n+    use std::sync::{Arc, Mutex};\n \n+    use async_trait::async_trait;\n     use fabro_llm::types::{ToolCall, ToolDefinition};\n+    use fabro_model::AgentProfileKind;\n \n     use super::*;\n     use crate::config::{\n@@ -428,6 +590,10 @@ mod tests {\n     };\n     use crate::event::Emitter;\n     use crate::local_sandbox::LocalSandbox;\n+    use crate::question_tools::{\n+        AgentQuestion, AgentQuestionAnswer, AgentQuestionAnswerStatus, AgentQuestionRuntime,\n+        AgentToolRuntime, register_question_tools,\n+    };\n     use crate::read_before_write_sandbox::ReadBeforeWriteSandbox;\n     use crate::test_support::MutableMockSandbox;\n     use crate::tool_registry::{RegisteredTool, ToolContext, ToolRegistry};\n@@ -505,6 +671,125 @@ mod tests {\n         }\n     }\n \n+    struct StubQuestionRuntime;\n+\n+    #[async_trait]\n+    impl AgentQuestionRuntime for StubQuestionRuntime {\n+        async fn ask_questions(\n+            &self,\n+            _tool_call_id: &str,\n+            questions: Vec<AgentQuestion>,\n+            _cancel_token: CancellationToken,\n+        ) -> Result<Vec<AgentQuestionAnswer>, String> {\n+            Ok(questions\n+                .into_iter()\n+                .map(|question| AgentQuestionAnswer {\n+                    original_id:       question.original_id,\n+                    original_question: question.original_question,\n+                    answers:           vec![\"Ship\".to_string()],\n+                    status:            AgentQuestionAnswerStatus::Answered,\n+                })\n+                .collect())\n+        }\n+    }\n+\n+    #[tokio::test]\n+    async fn question_tool_round_rejects_non_question_peers_and_preserves_order() {\n+        let mut registry = ToolRegistry::new();\n+        register_question_tools(AgentProfileKind::OpenAi, &mut registry);\n+        registry.register(make_echo_tool());\n+        let tool_calls = vec![\n+            make_tool_call(\n+                \"request_user_input\",\n+                \"call_question\",\n+                serde_json::json!({\n+                    \"questions\": [{\n+                        \"id\": \"q1\",\n+                        \"header\": \"Decision\",\n+                        \"question\": \"Ship it?\",\n+                        \"options\": [{ \"label\": \"Ship\" }]\n+                    }]\n+                }),\n+            ),\n+            make_tool_call(\"echo\", \"call_echo\", serde_json::json!({\"text\": \"hello\"})),\n+        ];\n+        let runtime = AgentToolRuntime::with_question_runtime(Arc::new(StubQuestionRuntime));\n+\n+        let results = execute_tool_calls(\n+            &tool_calls,\n+            true,\n+            &registry,\n+            Arc::new(LocalSandbox::new(std::env::current_dir().unwrap())),\n+            None,\n+            &CancellationToken::new(),\n+            &SessionOptions::default(),\n+            &Emitter::new(),\n+            \"root\",\n+            \"root\",\n+            None,\n+            &runtime,\n+        )\n+        .await;\n+\n+        assert_eq!(results.len(), 2);\n+        assert_eq!(results[0].tool_call_id, \"call_question\");\n+        assert!(!results[0].is_error);\n+        assert_eq!(results[1].tool_call_id, \"call_echo\");\n+        assert!(results[1].is_error);\n+        assert!(\n+            results[1]\n+                .content\n+                .as_str()\n+                .unwrap()\n+                .contains(\"human-question tools must run alone\")\n+        );\n+    }\n+\n+    #[tokio::test]\n+    async fn multiple_question_tool_calls_execute_only_first() {\n+        let mut registry = ToolRegistry::new();\n+        register_question_tools(AgentProfileKind::OpenAi, &mut registry);\n+        let question_args = serde_json::json!({\n+            \"questions\": [{\n+                \"id\": \"q1\",\n+                \"header\": \"Decision\",\n+                \"question\": \"Ship it?\",\n+                \"options\": [{ \"label\": \"Ship\" }]\n+            }]\n+        });\n+        let tool_calls = vec![\n+            make_tool_call(\"request_user_input\", \"call_first\", question_args.clone()),\n+            make_tool_call(\"request_user_input\", \"call_second\", question_args),\n+        ];\n+        let runtime = AgentToolRuntime::with_question_runtime(Arc::new(StubQuestionRuntime));\n+\n+        let results = execute_tool_calls(\n+            &tool_calls,\n+            true,\n+            &registry,\n+            Arc::new(LocalSandbox::new(std::env::current_dir().unwrap())),\n+            None,\n+            &CancellationToken::new(),\n+            &SessionOptions::default(),\n+            &Emitter::new(),\n+            \"root\",\n+            \"root\",\n+            None,\n+            &runtime,\n+        )\n+        .await;\n+\n+        assert!(!results[0].is_error);\n+        assert!(results[1].is_error);\n+        assert!(\n+            results[1]\n+                .content\n+                .as_str()\n+                .unwrap()\n+                .contains(\"Combine all questions into a single questions[] batch\")\n+        );\n+    }\n+\n     struct MockHookCallback {\n         pre_decision:       ToolHookDecision,\n         post_calls:         Arc<Mutex<Vec<(String, String, String)>>>,\ndiff --git a/lib/crates/fabro-api/tests/interview_option_round_trip.rs b/lib/crates/fabro-api/tests/interview_option_round_trip.rs\nindex 40fb51b23..06f4cff82 100644\n--- a/lib/crates/fabro-api/tests/interview_option_round_trip.rs\n+++ b/lib/crates/fabro-api/tests/interview_option_round_trip.rs\n@@ -13,7 +13,9 @@ fn interview_option_reuses_canonical_type() {\n fn interview_option_round_trips_representative_json() {\n     let value = json!({\n         \"key\": \"approve\",\n-        \"label\": \"Approve\"\n+        \"label\": \"Approve\",\n+        \"description\": \"Approve the proposed changes.\",\n+        \"preview\": \"diff --stat output\"\n     });\n \n     let option: InterviewOption = serde_json::from_value(value.clone()).unwrap();\ndiff --git a/lib/crates/fabro-api/tests/interview_question_record_round_trip.rs b/lib/crates/fabro-api/tests/interview_question_record_round_trip.rs\nindex 8e8f743b6..14338c4b4 100644\n--- a/lib/crates/fabro-api/tests/interview_question_record_round_trip.rs\n+++ b/lib/crates/fabro-api/tests/interview_question_record_round_trip.rs\n@@ -17,7 +17,12 @@ fn interview_question_record_round_trips_representative_json() {\n         \"stage\": \"gate\",\n         \"question_type\": \"multiple_choice\",\n         \"options\": [\n-            { \"key\": \"approve\", \"label\": \"Approve\" },\n+            {\n+                \"key\": \"approve\",\n+                \"label\": \"Approve\",\n+                \"description\": \"Deploy now\",\n+                \"preview\": \"deploy --prod\"\n+            },\n             { \"key\": \"reject\", \"label\": \"Reject\" }\n         ],\n         \"allow_freeform\": true,\ndiff --git a/lib/crates/fabro-api/tests/pending_interview_record_round_trip.rs b/lib/crates/fabro-api/tests/pending_interview_record_round_trip.rs\nindex 9ce7c4c9c..dae81fdc3 100644\n--- a/lib/crates/fabro-api/tests/pending_interview_record_round_trip.rs\n+++ b/lib/crates/fabro-api/tests/pending_interview_record_round_trip.rs\n@@ -18,7 +18,12 @@ fn pending_interview_record_round_trips_populated_question() {\n             \"stage\": \"gate\",\n             \"question_type\": \"multiple_choice\",\n             \"options\": [\n-                { \"key\": \"approve\", \"label\": \"Approve\" },\n+                {\n+                    \"key\": \"approve\",\n+                    \"label\": \"Approve\",\n+                    \"description\": \"Deploy now\",\n+                    \"preview\": \"deploy --prod\"\n+                },\n                 { \"key\": \"reject\", \"label\": \"Reject\" }\n             ],\n             \"allow_freeform\": true,\ndiff --git a/lib/crates/fabro-api/tests/run_projection_round_trip.rs b/lib/crates/fabro-api/tests/run_projection_round_trip.rs\nindex 00aeb7ae2..64a00df91 100644\n--- a/lib/crates/fabro-api/tests/run_projection_round_trip.rs\n+++ b/lib/crates/fabro-api/tests/run_projection_round_trip.rs\n@@ -54,7 +54,12 @@ fn run_projection_round_trips_populated_projection() {\n                     \"stage\": \"gate\",\n                     \"question_type\": \"multiple_choice\",\n                     \"options\": [\n-                        { \"key\": \"approve\", \"label\": \"Approve\" },\n+                        {\n+                            \"key\": \"approve\",\n+                            \"label\": \"Approve\",\n+                            \"description\": \"Deploy now\",\n+                            \"preview\": \"deploy --prod\"\n+                        },\n                         { \"key\": \"reject\", \"label\": \"Reject\" }\n                     ],\n                     \"allow_freeform\": true,\ndiff --git a/lib/crates/fabro-cli/src/commands/run/attach.rs b/lib/crates/fabro-cli/src/commands/run/attach.rs\nindex d66249aad..c8ad2fff2 100644\n--- a/lib/crates/fabro-cli/src/commands/run/attach.rs\n+++ b/lib/crates/fabro-cli/src/commands/run/attach.rs\n@@ -426,8 +426,10 @@ fn api_question_to_question(question: &types::ApiQuestion) -> Question {\n         .options\n         .iter()\n         .map(|option| InterviewOption {\n-            key:   option.key.clone(),\n-            label: option.label.clone(),\n+            key:         option.key.clone(),\n+            label:       option.label.clone(),\n+            description: option.description.clone(),\n+            preview:     option.preview.clone(),\n         })\n         .collect();\n     converted.allow_freeform = question.allow_freeform;\n@@ -966,8 +968,10 @@ mod tests {\n     fn invalid_multiple_choice_without_freeform_is_user_correctable() {\n         let mut question = Question::new(\"Pick one.\", QuestionType::MultipleChoice);\n         question.options = vec![InterviewOption {\n-            key:   \"A\".to_string(),\n-            label: \"Approve\".to_string(),\n+            key:         \"A\".to_string(),\n+            label:       \"Approve\".to_string(),\n+            description: None,\n+            preview:     None,\n         }];\n \n         let response = parse_choice_response(&question, PromptRead::Line(\"bogus\".to_string()));\n@@ -979,8 +983,10 @@ mod tests {\n     fn unmatched_multiple_choice_with_freeform_remains_text() {\n         let mut question = Question::new(\"Pick one.\", QuestionType::MultipleChoice);\n         question.options = vec![InterviewOption {\n-            key:   \"A\".to_string(),\n-            label: \"Approve\".to_string(),\n+            key:         \"A\".to_string(),\n+            label:       \"Approve\".to_string(),\n+            description: None,\n+            preview:     None,\n         }];\n         question.allow_freeform = true;\n \n@@ -1000,12 +1006,16 @@ mod tests {\n         let mut question = Question::new(\"Pick many.\", QuestionType::MultiSelect);\n         question.options = vec![\n             InterviewOption {\n-                key:   \"A\".to_string(),\n-                label: \"Approve\".to_string(),\n+                key:         \"A\".to_string(),\n+                label:       \"Approve\".to_string(),\n+                description: None,\n+                preview:     None,\n             },\n             InterviewOption {\n-                key:   \"N\".to_string(),\n-                label: \"Notify\".to_string(),\n+                key:         \"N\".to_string(),\n+                label:       \"Notify\".to_string(),\n+                description: None,\n+                preview:     None,\n             },\n         ];\n \ndiff --git a/lib/crates/fabro-interview/src/auto_approve.rs b/lib/crates/fabro-interview/src/auto_approve.rs\nindex faeba9a7b..66a323548 100644\n--- a/lib/crates/fabro-interview/src/auto_approve.rs\n+++ b/lib/crates/fabro-interview/src/auto_approve.rs\n@@ -72,12 +72,16 @@ mod tests {\n         let mut q = Question::new(\"Choose:\", QuestionType::MultipleChoice);\n         q.options = vec![\n             InterviewOption {\n-                key:   \"A\".to_string(),\n-                label: \"Alpha\".to_string(),\n+                key:         \"A\".to_string(),\n+                label:       \"Alpha\".to_string(),\n+                description: None,\n+                preview:     None,\n             },\n             InterviewOption {\n-                key:   \"B\".to_string(),\n-                label: \"Beta\".to_string(),\n+                key:         \"B\".to_string(),\n+                label:       \"Beta\".to_string(),\n+                description: None,\n+                preview:     None,\n             },\n         ];\n         let answer = interviewer.ask(q).await.answer;\n@@ -85,8 +89,10 @@ mod tests {\n         assert_eq!(\n             answer.selected_option,\n             Some(InterviewOption {\n-                key:   \"A\".to_string(),\n-                label: \"Alpha\".to_string(),\n+                key:         \"A\".to_string(),\n+                label:       \"Alpha\".to_string(),\n+                description: None,\n+                preview:     None,\n             })\n         );\n     }\ndiff --git a/lib/crates/fabro-interview/src/console.rs b/lib/crates/fabro-interview/src/console.rs\nindex 437025041..0005b1aca 100644\n--- a/lib/crates/fabro-interview/src/console.rs\n+++ b/lib/crates/fabro-interview/src/console.rs\n@@ -296,12 +296,16 @@ mod tests {\n     fn find_matching_option_by_key() {\n         let options = vec![\n             InterviewOption {\n-                key:   \"A\".to_string(),\n-                label: \"Approve\".to_string(),\n+                key:         \"A\".to_string(),\n+                label:       \"Approve\".to_string(),\n+                description: None,\n+                preview:     None,\n             },\n             InterviewOption {\n-                key:   \"R\".to_string(),\n-                label: \"Reject\".to_string(),\n+                key:         \"R\".to_string(),\n+                label:       \"Reject\".to_string(),\n+                description: None,\n+                preview:     None,\n             },\n         ];\n         let result = find_matching_option(\"A\", &options);\n@@ -313,8 +317,10 @@ mod tests {\n     #[test]\n     fn find_matching_option_by_key_case_insensitive() {\n         let options = vec![InterviewOption {\n-            key:   \"Y\".to_string(),\n-            label: \"Yes\".to_string(),\n+            key:         \"Y\".to_string(),\n+            label:       \"Yes\".to_string(),\n+            description: None,\n+            preview:     None,\n         }];\n         let result = find_matching_option(\"y\", &options);\n         assert!(result.is_some());\n@@ -324,12 +330,16 @@ mod tests {\n     fn find_matching_option_by_index() {\n         let options = vec![\n             InterviewOption {\n-                key:   \"A\".to_string(),\n-                label: \"Alpha\".to_string(),\n+                key:         \"A\".to_string(),\n+                label:       \"Alpha\".to_string(),\n+                description: None,\n+                preview:     None,\n             },\n             InterviewOption {\n-                key:   \"B\".to_string(),\n-                label: \"Beta\".to_string(),\n+                key:         \"B\".to_string(),\n+                label:       \"Beta\".to_string(),\n+                description: None,\n+                preview:     None,\n             },\n         ];\n         let result = find_matching_option(\"2\", &options);\n@@ -341,8 +351,10 @@ mod tests {\n     #[test]\n     fn find_matching_option_no_match() {\n         let options = vec![InterviewOption {\n-            key:   \"A\".to_string(),\n-            label: \"Alpha\".to_string(),\n+            key:         \"A\".to_string(),\n+            label:       \"Alpha\".to_string(),\n+            description: None,\n+            preview:     None,\n         }];\n         let result = find_matching_option(\"zzz\", &options);\n         assert!(result.is_none());\n@@ -351,8 +363,10 @@ mod tests {\n     #[test]\n     fn find_matching_option_index_out_of_range() {\n         let options = vec![InterviewOption {\n-            key:   \"A\".to_string(),\n-            label: \"Alpha\".to_string(),\n+            key:         \"A\".to_string(),\n+            label:       \"Alpha\".to_string(),\n+            description: None,\n+            preview:     None,\n         }];\n         let result = find_matching_option(\"5\", &options);\n         assert!(result.is_none());\n@@ -362,8 +376,10 @@ mod tests {\n     fn non_tty_multiple_choice_eof_returns_interrupted() {\n         let mut question = Question::new(\"Approve?\", QuestionType::MultipleChoice);\n         question.options = vec![InterviewOption {\n-            key:   \"A\".to_string(),\n-            label: \"Approve\".to_string(),\n+            key:         \"A\".to_string(),\n+            label:       \"Approve\".to_string(),\n+            description: None,\n+            preview:     None,\n         }];\n \n         let answer = parse_non_tty_choice_response(&question, PromptRead::Eof);\ndiff --git a/lib/crates/fabro-interview/src/lib.rs b/lib/crates/fabro-interview/src/lib.rs\nindex c242cf14a..4c050c4f7 100644\n--- a/lib/crates/fabro-interview/src/lib.rs\n+++ b/lib/crates/fabro-interview/src/lib.rs\n@@ -301,8 +301,10 @@ mod tests {\n     #[test]\n     fn answer_selected() {\n         let opt = InterviewOption {\n-            key:   \"A\".to_string(),\n-            label: \"Approve\".to_string(),\n+            key:         \"A\".to_string(),\n+            label:       \"Approve\".to_string(),\n+            description: None,\n+            preview:     None,\n         };\n         let a = Answer::selected(\"A\", opt.clone());\n         assert_eq!(a.value, AnswerValue::Selected(\"A\".to_string()));\n@@ -319,12 +321,16 @@ mod tests {\n     #[test]\n     fn question_option_eq() {\n         let a = InterviewOption {\n-            key:   \"Y\".to_string(),\n-            label: \"Yes\".to_string(),\n+            key:         \"Y\".to_string(),\n+            label:       \"Yes\".to_string(),\n+            description: None,\n+            preview:     None,\n         };\n         let b = InterviewOption {\n-            key:   \"Y\".to_string(),\n-            label: \"Yes\".to_string(),\n+            key:         \"Y\".to_string(),\n+            label:       \"Yes\".to_string(),\n+            description: None,\n+            preview:     None,\n         };\n         assert_eq!(a, b);\n     }\ndiff --git a/lib/crates/fabro-server/src/demo/mod.rs b/lib/crates/fabro-server/src/demo/mod.rs\nindex a71d9c748..cf533295b 100644\n--- a/lib/crates/fabro-server/src/demo/mod.rs\n+++ b/lib/crates/fabro-server/src/demo/mod.rs\n@@ -1677,13 +1677,17 @@ mod runs {\n                 stage:           \"review\".into(),\n                 question_type:   QuestionType::YesNo,\n                 options:         vec![\n-                    ApiQuestionOption {\n-                        key:   \"yes\".into(),\n-                        label: \"Yes\".into(),\n+                    InterviewOption {\n+                        key:         \"yes\".into(),\n+                        label:       \"Yes\".into(),\n+                        description: None,\n+                        preview:     None,\n                     },\n-                    ApiQuestionOption {\n-                        key:   \"no\".into(),\n-                        label: \"No\".into(),\n+                    InterviewOption {\n+                        key:         \"no\".into(),\n+                        label:       \"No\".into(),\n+                        description: None,\n+                        preview:     None,\n                     },\n                 ],\n                 allow_freeform:  false,\n@@ -1696,13 +1700,17 @@ mod runs {\n                 stage:           \"migration\".into(),\n                 question_type:   QuestionType::MultipleChoice,\n                 options:         vec![\n-                    ApiQuestionOption {\n-                        key:   \"incremental\".into(),\n-                        label: \"Incremental migration\".into(),\n+                    InterviewOption {\n+                        key:         \"incremental\".into(),\n+                        label:       \"Incremental migration\".into(),\n+                        description: None,\n+                        preview:     None,\n                     },\n-                    ApiQuestionOption {\n-                        key:   \"big_bang\".into(),\n-                        label: \"Big-bang rewrite\".into(),\n+                    InterviewOption {\n+                        key:         \"big_bang\".into(),\n+                        label:       \"Big-bang rewrite\".into(),\n+                        description: None,\n+                        preview:     None,\n                     },\n                 ],\n                 allow_freeform:  true,\ndiff --git a/lib/crates/fabro-server/src/server.rs b/lib/crates/fabro-server/src/server.rs\nindex 9bbf93cff..701697a2c 100644\n--- a/lib/crates/fabro-server/src/server.rs\n+++ b/lib/crates/fabro-server/src/server.rs\n@@ -22,10 +22,10 @@ use base64::Engine as _;\n use base64::engine::general_purpose::STANDARD as BASE64_STANDARD;\n use bytes::Bytes;\n pub use fabro_api::types::{\n-    AggregateBilling, AggregateBillingTotals, ApiQuestion, ApiQuestionOption, AppendEventResponse,\n-    ArtifactEntry, ArtifactListResponse, BillingByModel, BillingStageRef,\n-    CloseRunPullRequestResponse, CompletionContentPart, CompletionMessage, CompletionMessageRole,\n-    CompletionResponse, CompletionToolChoiceMode, CompletionUsage, CreateCompletionRequest,\n+    AggregateBilling, AggregateBillingTotals, ApiQuestion, AppendEventResponse, ArtifactEntry,\n+    ArtifactListResponse, BillingByModel, BillingStageRef, CloseRunPullRequestResponse,\n+    CompletionContentPart, CompletionMessage, CompletionMessageRole, CompletionResponse,\n+    CompletionToolChoiceMode, CompletionUsage, CreateCompletionRequest,\n     CreateRunPullRequestRequest, CreateSecretRequest, DeleteRunResponse, DeleteRunSandbox,\n     DeleteSecretRequest, DiskUsageResponse, DiskUsageRunRow, DiskUsageSummaryRow, ForkRequest,\n     ForkResponse, LinkRunPullRequestRequest, MergeRunPullRequestRequest,\n@@ -2894,14 +2894,7 @@ fn api_question_from_interview_record(question: &InterviewQuestionRecord) -> Api\n         text:            question.text.clone(),\n         stage:           question.stage.clone(),\n         question_type:   question.question_type,\n-        options:         question\n-            .options\n-            .iter()\n-            .map(|option| ApiQuestionOption {\n-                key:   option.key.clone(),\n-                label: option.label.clone(),\n-            })\n-            .collect(),\n+        options:         question.options.clone(),\n         allow_freeform:  question.allow_freeform,\n         timeout_seconds: question.timeout_seconds,\n         context_display: question.context_display.clone(),\ndiff --git a/lib/crates/fabro-server/src/server/tests.rs b/lib/crates/fabro-server/src/server/tests.rs\nindex a9ad98fd5..5829dfccc 100644\n--- a/lib/crates/fabro-server/src/server/tests.rs\n+++ b/lib/crates/fabro-server/src/server/tests.rs\n@@ -4861,8 +4861,10 @@ async fn submit_pending_interview_answer_rejects_invalid_answer_shape() {\n             stage:           \"gate\".to_string(),\n             question_type:   QuestionType::MultipleChoice,\n             options:         vec![fabro_types::run_event::InterviewOption {\n-                key:   \"approve\".to_string(),\n-                label: \"Approve\".to_string(),\n+                key:         \"approve\".to_string(),\n+                label:       \"Approve\".to_string(),\n+                description: None,\n+                preview:     None,\n             }],\n             allow_freeform:  false,\n             timeout_seconds: None,\n@@ -4948,8 +4950,10 @@ fn answer_from_typed_selected_request_validates_and_attaches_option() {\n         stage:           \"gate\".to_string(),\n         question_type:   QuestionType::MultipleChoice,\n         options:         vec![fabro_types::run_event::InterviewOption {\n-            key:   \"approve\".to_string(),\n-            label: \"Approve\".to_string(),\n+            key:         \"approve\".to_string(),\n+            label:       \"Approve\".to_string(),\n+            description: None,\n+            preview:     None,\n         }],\n         allow_freeform:  false,\n         timeout_seconds: None,\n@@ -4979,12 +4983,16 @@ fn answer_from_typed_multi_selected_request_validates_option_keys() {\n         question_type:   QuestionType::MultiSelect,\n         options:         vec![\n             fabro_types::run_event::InterviewOption {\n-                key:   \"approve\".to_string(),\n-                label: \"Approve\".to_string(),\n+                key:         \"approve\".to_string(),\n+                label:       \"Approve\".to_string(),\n+                description: None,\n+                preview:     None,\n             },\n             fabro_types::run_event::InterviewOption {\n-                key:   \"notify\".to_string(),\n-                label: \"Notify\".to_string(),\n+                key:         \"notify\".to_string(),\n+                label:       \"Notify\".to_string(),\n+                description: None,\n+                preview:     None,\n             },\n         ],\n         allow_freeform:  false,\ndiff --git a/lib/crates/fabro-slack/src/blocks.rs b/lib/crates/fabro-slack/src/blocks.rs\nindex 83a94543b..7632ab97e 100644\n--- a/lib/crates/fabro-slack/src/blocks.rs\n+++ b/lib/crates/fabro-slack/src/blocks.rs\n@@ -153,6 +153,32 @@ fn lead_blocks(question: &Question, run_web_url: Option<&str>) -> Vec<Value> {\n     blocks\n }\n \n+fn option_descriptions_section(question: &Question) -> Option<Value> {\n+    let rows = question\n+        .options\n+        .iter()\n+        .filter_map(|option| {\n+            let description = option.description.as_deref()?.trim();\n+            if description.is_empty() {\n+                return None;\n+            }\n+            Some(format!(\n+                \"• *{}* — {}\",\n+                escape_slack_controls(&option.label),\n+                escape_slack_controls(description)\n+            ))\n+        })\n+        .collect::<Vec<_>>();\n+    if rows.is_empty() {\n+        return None;\n+    }\n+    Some(text_block(&truncate_to_limit(\n+        &rows.join(\"\\n\"),\n+        SLACK_SECTION_TEXT_LIMIT,\n+        HEADER_TRUNCATION_SUFFIX,\n+    )))\n+}\n+\n pub fn answered_blocks(question_text: &str, answer_text: &str) -> Vec<Value> {\n     vec![text_block(&format!(\n         \"~{}~\\n*Answer:* {}\",\n@@ -168,6 +194,9 @@ pub fn question_to_blocks(\n     run_web_url: Option<&str>,\n ) -> Vec<Value> {\n     let mut blocks = lead_blocks(question, run_web_url);\n+    if let Some(descriptions) = option_descriptions_section(question) {\n+        blocks.push(descriptions);\n+    }\n \n     match question.question_type {\n         QuestionType::YesNo | QuestionType::Confirmation => {\n@@ -212,10 +241,26 @@ pub fn question_to_blocks(\n                 .options\n                 .iter()\n                 .map(|opt| {\n-                    json!({\n+                    let mut option = json!({\n                         \"text\": { \"type\": \"plain_text\", \"text\": opt.label },\n                         \"value\": opt.key\n-                    })\n+                    });\n+                    if let Some(description) = opt\n+                        .description\n+                        .as_deref()\n+                        .map(str::trim)\n+                        .filter(|value| !value.is_empty())\n+                    {\n+                        option[\"description\"] = json!({\n+                            \"type\": \"plain_text\",\n+                            \"text\": truncate_to_limit(\n+                                description,\n+                                75,\n+                                HEADER_TRUNCATION_SUFFIX\n+                            )\n+                        });\n+                    }\n+                    option\n                 })\n                 .collect();\n             blocks.push(json!({\n@@ -293,16 +338,22 @@ mod tests {\n         let mut q = Question::new(\"Pick a language:\", QuestionType::MultipleChoice);\n         q.options = vec![\n             InterviewOption {\n-                key:   \"rs\".to_string(),\n-                label: \"Rust\".to_string(),\n+                key:         \"rs\".to_string(),\n+                label:       \"Rust\".to_string(),\n+                description: None,\n+                preview:     None,\n             },\n             InterviewOption {\n-                key:   \"ts\".to_string(),\n-                label: \"TypeScript\".to_string(),\n+                key:         \"ts\".to_string(),\n+                label:       \"TypeScript\".to_string(),\n+                description: None,\n+                preview:     None,\n             },\n             InterviewOption {\n-                key:   \"py\".to_string(),\n-                label: \"Python\".to_string(),\n+                key:         \"py\".to_string(),\n+                label:       \"Python\".to_string(),\n+                description: None,\n+                preview:     None,\n             },\n         ];\n         let blocks = question_to_blocks(\"run-1\", \"q-3\", &q, None);\n@@ -566,12 +617,16 @@ mod tests {\n         let mut q = Question::new(\"Select features:\", QuestionType::MultiSelect);\n         q.options = vec![\n             InterviewOption {\n-                key:   \"a\".to_string(),\n-                label: \"Auth\".to_string(),\n+                key:         \"a\".to_string(),\n+                label:       \"Auth\".to_string(),\n+                description: None,\n+                preview:     None,\n             },\n             InterviewOption {\n-                key:   \"b\".to_string(),\n-                label: \"Billing\".to_string(),\n+                key:         \"b\".to_string(),\n+                label:       \"Billing\".to_string(),\n+                description: None,\n+                preview:     None,\n             },\n         ];\n         let blocks = question_to_blocks(\"run-1\", \"q-5\", &q, None);\n@@ -602,4 +657,22 @@ mod tests {\n                 .contains(\"\\\"qid\\\":\\\"q-5\\\"\")\n         );\n     }\n+\n+    #[test]\n+    fn option_descriptions_are_rendered_and_preview_is_not_special_cased() {\n+        let mut q = Question::new(\"Pick one:\", QuestionType::MultipleChoice);\n+        q.options = vec![InterviewOption {\n+            key:         \"ship\".to_string(),\n+            label:       \"Ship\".to_string(),\n+            description: Some(\"Deploy <now>\".to_string()),\n+            preview:     Some(\"preview should not render\".to_string()),\n+        }];\n+\n+        let blocks_value: Value =\n+            serde_json::to_value(question_to_blocks(\"run-1\", \"q-6\", &q, None)).unwrap();\n+        let text = blocks_value.to_string();\n+\n+        assert!(text.contains(\"Deploy &lt;now&gt;\"));\n+        assert!(!text.contains(\"preview should not render\"));\n+    }\n }\ndiff --git a/lib/crates/fabro-store/src/run_state.rs b/lib/crates/fabro-store/src/run_state.rs\nindex 0da6e957c..246784d40 100644\n--- a/lib/crates/fabro-store/src/run_state.rs\n+++ b/lib/crates/fabro-store/src/run_state.rs\n@@ -1757,12 +1757,16 @@ mod tests {\n                     question_type:   \"multiple_choice\".to_string(),\n                     options:         vec![\n                         InterviewOption {\n-                            key:   \"approve\".to_string(),\n-                            label: \"Approve\".to_string(),\n+                            key:         \"approve\".to_string(),\n+                            label:       \"Approve\".to_string(),\n+                            description: Some(\"Ship it\".to_string()),\n+                            preview:     Some(\"deploy --prod\".to_string()),\n                         },\n                         InterviewOption {\n-                            key:   \"revise\".to_string(),\n-                            label: \"Revise\".to_string(),\n+                            key:         \"revise\".to_string(),\n+                            label:       \"Revise\".to_string(),\n+                            description: None,\n+                            preview:     None,\n                         },\n                     ],\n                     allow_freeform:  true,\n@@ -1781,6 +1785,14 @@ mod tests {\n         assert_eq!(pending.question.stage, \"gate\");\n         assert_eq!(pending.question.question_type, QuestionType::MultipleChoice);\n         assert_eq!(pending.question.options.len(), 2);\n+        assert_eq!(\n+            pending.question.options[0].description.as_deref(),\n+            Some(\"Ship it\")\n+        );\n+        assert_eq!(\n+            pending.question.options[0].preview.as_deref(),\n+            Some(\"deploy --prod\")\n+        );\n         assert!(pending.question.allow_freeform);\n         assert_eq!(pending.question.timeout_seconds, Some(30.0));\n         assert_eq!(\ndiff --git a/lib/crates/fabro-types/src/run_event/misc.rs b/lib/crates/fabro-types/src/run_event/misc.rs\nindex b2280c243..3f88cdbd3 100644\n--- a/lib/crates/fabro-types/src/run_event/misc.rs\n+++ b/lib/crates/fabro-types/src/run_event/misc.rs\n@@ -4,10 +4,14 @@ use serde_json::Value;\n use super::ExecOutputTail;\n use crate::{CommandTermination, PullRequestLink};\n \n-#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]\n+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, Default)]\n pub struct InterviewOption {\n-    pub key:   String,\n-    pub label: String,\n+    pub key:         String,\n+    pub label:       String,\n+    #[serde(default, skip_serializing_if = \"Option::is_none\")]\n+    pub description: Option<String>,\n+    #[serde(default, skip_serializing_if = \"Option::is_none\")]\n+    pub preview:     Option<String>,\n }\n \n #[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]\ndiff --git a/lib/crates/fabro-workflow/src/handler/agent.rs b/lib/crates/fabro-workflow/src/handler/agent.rs\nindex 24feb5650..9bfb742b5 100644\n--- a/lib/crates/fabro-workflow/src/handler/agent.rs\n+++ b/lib/crates/fabro-workflow/src/handler/agent.rs\n@@ -11,6 +11,7 @@ use super::{EngineServices, Handler, NodeTimeoutPolicy};\n use crate::context::{Context, WorkflowContext, keys};\n use crate::error::Error;\n use crate::event::{Emitter, Event, StageScope};\n+use crate::interview_runtime::WorkflowAgentQuestionRuntime;\n use crate::outcome::{\n     BilledModelUsage, FailureCategory, FailureDetail, Outcome, OutcomeExt, StageOutcome,\n };\n@@ -27,14 +28,15 @@ pub enum CodergenResult {\n }\n \n pub struct CodergenRunRequest<'a> {\n-    pub node:         &'a Node,\n-    pub prompt:       &'a str,\n-    pub context:      &'a Context,\n-    pub thread_id:    Option<&'a str>,\n-    pub emitter:      &'a Arc<Emitter>,\n-    pub sandbox:      &'a Arc<dyn Sandbox>,\n-    pub tool_hooks:   Option<Arc<dyn fabro_agent::ToolHookCallback>>,\n-    pub cancel_token: CancellationToken,\n+    pub node:               &'a Node,\n+    pub prompt:             &'a str,\n+    pub context:            &'a Context,\n+    pub thread_id:          Option<&'a str>,\n+    pub emitter:            &'a Arc<Emitter>,\n+    pub sandbox:            &'a Arc<dyn Sandbox>,\n+    pub tool_hooks:         Option<Arc<dyn fabro_agent::ToolHookCallback>>,\n+    pub cancel_token:       CancellationToken,\n+    pub agent_tool_runtime: fabro_agent::AgentToolRuntime,\n }\n \n pub struct OneShotRequest<'a> {\n@@ -258,6 +260,15 @@ impl Handler for AgentHandler {\n             .or_else(|| Some(services.run.provider_id.to_string()));\n         let prompt_model = node.model().map(String::from);\n         let stage_scope = StageScope::for_handler(context, &node.id);\n+        let agent_tool_runtime = fabro_agent::AgentToolRuntime::with_question_runtime(Arc::new(\n+            WorkflowAgentQuestionRuntime::new(\n+                Arc::clone(&services.interviewer),\n+                Arc::clone(&services.run.emitter),\n+                stage_scope.clone(),\n+                node.id.clone(),\n+                Arc::clone(&services.run.interview_blocker),\n+            ),\n+        ));\n         services.run.emitter.emit_scoped(\n             &Event::Prompt {\n                 stage:    node.id.clone(),\n@@ -299,6 +310,7 @@ impl Handler for AgentHandler {\n                         sandbox: &services.run.sandbox,\n                         tool_hooks,\n                         cancel_token: services.run.cancel_token(),\n+                        agent_tool_runtime: agent_tool_runtime.clone(),\n                     })\n                     .await;\n                 match result {\ndiff --git a/lib/crates/fabro-workflow/src/handler/fan_in.rs b/lib/crates/fabro-workflow/src/handler/fan_in.rs\nindex 455b27b26..3a712ec0f 100644\n--- a/lib/crates/fabro-workflow/src/handler/fan_in.rs\n+++ b/lib/crates/fabro-workflow/src/handler/fan_in.rs\n@@ -269,6 +269,7 @@ async fn llm_evaluate(\n             sandbox,\n             tool_hooks: None,\n             cancel_token,\n+            agent_tool_runtime: fabro_agent::AgentToolRuntime::default(),\n         })\n         .await\n     {\ndiff --git a/lib/crates/fabro-workflow/src/handler/human.rs b/lib/crates/fabro-workflow/src/handler/human.rs\nindex a46a6e525..77b83454a 100644\n--- a/lib/crates/fabro-workflow/src/handler/human.rs\n+++ b/lib/crates/fabro-workflow/src/handler/human.rs\n@@ -1,13 +1,12 @@\n use std::path::Path;\n use std::str::FromStr;\n use std::sync::Arc;\n-use std::sync::atomic::{AtomicUsize, Ordering};\n use std::time::Instant;\n \n use async_trait::async_trait;\n use fabro_graphviz::graph::{Graph, Node};\n use fabro_interview::{Answer, AnswerValue, Interviewer, Question, ask_with_timeout};\n-use fabro_types::{BlockedReason, InterviewOption, Principal, QuestionType, SystemActorKind};\n+use fabro_types::{InterviewOption, Principal, QuestionType, SystemActorKind};\n use ulid::Ulid;\n \n use super::{EngineServices, Handler, NodeTimeoutPolicy};\n@@ -112,8 +111,10 @@ fn build_human_gate_question(\n     question.options = choices\n         .iter()\n         .map(|choice| InterviewOption {\n-            key:   choice.key.clone(),\n-            label: choice.label.clone(),\n+            key:         choice.key.clone(),\n+            label:       choice.label.clone(),\n+            description: None,\n+            preview:     None,\n         })\n         .collect();\n     question.allow_freeform = freeform_target.is_some();\n@@ -138,61 +139,10 @@ fn build_human_gate_question(\n     })\n }\n \n-/// Refcount of open interviews for this handler's run. Emits `run.blocked`\n-/// exactly once when the count transitions 0→1, and `run.unblocked` exactly\n-/// once when it transitions back to 0. Internal to `HumanHandler`; shared\n-/// across concurrent `execute` calls fanned out by `ParallelHandler`.\n-struct BlockedStateTracker {\n-    unresolved_interviews: AtomicUsize,\n-}\n-\n-impl BlockedStateTracker {\n-    fn new() -> Self {\n-        Self {\n-            unresolved_interviews: AtomicUsize::new(0),\n-        }\n-    }\n-\n-    fn interview_started(&self, emitter: &Emitter) {\n-        if self.unresolved_interviews.fetch_add(1, Ordering::AcqRel) == 0 {\n-            emitter.emit(&Event::RunBlocked {\n-                blocked_reason: BlockedReason::HumanInputRequired,\n-            });\n-        }\n-    }\n-\n-    fn interview_resolved(&self, emitter: &Emitter) {\n-        // Guard against unmatched resolves (e.g., tests that over-resolve) so\n-        // the counter cannot underflow. `compare_exchange_weak` loops until we\n-        // either observe zero (and bail) or successfully decrement.\n-        let mut current = self.unresolved_interviews.load(Ordering::Acquire);\n-        loop {\n-            if current == 0 {\n-                return;\n-            }\n-            match self.unresolved_interviews.compare_exchange_weak(\n-                current,\n-                current - 1,\n-                Ordering::AcqRel,\n-                Ordering::Acquire,\n-            ) {\n-                Ok(_) => {\n-                    if current == 1 {\n-                        emitter.emit(&Event::RunUnblocked);\n-                    }\n-                    return;\n-                }\n-                Err(observed) => current = observed,\n-            }\n-        }\n-    }\n-}\n-\n /// Blocks until a human selects an option derived from outgoing edges.\n pub struct HumanHandler {\n     interviewer: Arc<dyn Interviewer>,\n     emitter:     Option<Arc<Emitter>>,\n-    tracker:     BlockedStateTracker,\n }\n \n impl HumanHandler {\n@@ -200,7 +150,6 @@ impl HumanHandler {\n         Self {\n             interviewer,\n             emitter: None,\n-            tracker: BlockedStateTracker::new(),\n         }\n     }\n \n@@ -295,8 +244,10 @@ impl Handler for HumanHandler {\n                     .options\n                     .iter()\n                     .map(|option| InterviewOption {\n-                        key:   option.key.clone(),\n-                        label: option.label.clone(),\n+                        key:         option.key.clone(),\n+                        label:       option.label.clone(),\n+                        description: option.description.clone(),\n+                        preview:     option.preview.clone(),\n                     })\n                     .collect(),\n                 allow_freeform:  question.allow_freeform,\n@@ -305,8 +256,10 @@ impl Handler for HumanHandler {\n             },\n             &stage_scope,\n         );\n-        self.tracker\n-            .interview_started(services.run.emitter.as_ref());\n+        let interview_guard = services\n+            .run\n+            .interview_blocker\n+            .block(Arc::clone(&services.run.emitter));\n         let interview_start = Instant::now();\n         let answer_submission = ask_with_timeout(self.interviewer.as_ref(), question).await;\n         let answer_actor = answer_submission.actor.clone();\n@@ -327,8 +280,7 @@ impl Handler for HumanHandler {\n                 },\n                 &stage_scope,\n             );\n-            self.tracker\n-                .interview_resolved(services.run.emitter.as_ref());\n+            interview_guard.resolve();\n             let default_choice = node\n                 .attrs\n                 .get(\"human.default_choice\")\n@@ -371,8 +323,7 @@ impl Handler for HumanHandler {\n                 },\n                 &stage_scope,\n             );\n-            self.tracker\n-                .interview_resolved(services.run.emitter.as_ref());\n+            interview_guard.resolve();\n             return Ok(unanswered_human_gate(\n                 \"human interaction interrupted before an answer was provided\",\n             ));\n@@ -389,8 +340,7 @@ impl Handler for HumanHandler {\n                 },\n                 &stage_scope,\n             );\n-            self.tracker\n-                .interview_resolved(services.run.emitter.as_ref());\n+            interview_guard.resolve();\n             return Ok(unanswered_human_gate(\"human skipped interaction\"));\n         }\n \n@@ -406,8 +356,7 @@ impl Handler for HumanHandler {\n             },\n             &stage_scope,\n         );\n-        self.tracker\n-            .interview_resolved(services.run.emitter.as_ref());\n+        interview_guard.resolve();\n \n         // Try fixed-choice match\n         if let Some(selected) = find_choice_match(&answer, &choices) {\n@@ -894,8 +843,10 @@ mod tests {\n     async fn wait_human_emits_blocked_then_unblocked_around_interview() {\n         let interviewer = Arc::new(CallbackInterviewer::new(|_| {\n             Answer::selected(\"A\", InterviewOption {\n-                key:   \"A\".to_string(),\n-                label: \"Approve\".to_string(),\n+                key:         \"A\".to_string(),\n+                label:       \"Approve\".to_string(),\n+                description: None,\n+                preview:     None,\n             })\n         }));\n         let handler = HumanHandler::new(interviewer);\n@@ -1128,9 +1079,10 @@ mod tests {\n \n     #[test]\n     fn blocked_state_tracker_emits_once_across_parallel_interview_races() {\n-        let tracker = BlockedStateTracker::new();\n+        let blocker = Arc::new(crate::interview_runtime::RunInterviewBlocker::new());\n         let emitter = Arc::new(Emitter::new(fabro_types::fixtures::RUN_1));\n         let event_names = Arc::new(Mutex::new(Vec::new()));\n+        let guards = Arc::new(Mutex::new(Vec::new()));\n \n         emitter.on_event({\n             let event_names = Arc::clone(&event_names);\n@@ -1148,22 +1100,25 @@ mod tests {\n \n         std::thread::scope(|scope| {\n             for _ in 0..8 {\n-                let tracker = &tracker;\n+                let blocker = Arc::clone(&blocker);\n                 let emitter = Arc::clone(&emitter);\n-                scope.spawn(move || tracker.interview_started(emitter.as_ref()));\n+                let guards = Arc::clone(&guards);\n+                scope.spawn(move || {\n+                    guards.lock().unwrap().push(blocker.block(emitter));\n+                });\n             }\n         });\n \n         std::thread::scope(|scope| {\n             for _ in 0..8 {\n-                let tracker = &tracker;\n-                let emitter = Arc::clone(&emitter);\n-                scope.spawn(move || tracker.interview_resolved(emitter.as_ref()));\n+                let guards = Arc::clone(&guards);\n+                scope.spawn(move || {\n+                    let guard = guards.lock().unwrap().pop().unwrap();\n+                    guard.resolve();\n+                });\n             }\n         });\n \n-        tracker.interview_resolved(emitter.as_ref());\n-\n         assert_eq!(event_names.lock().unwrap().as_slice(), [\n             \"run.blocked\",\n             \"run.unblocked\"\ndiff --git a/lib/crates/fabro-workflow/src/handler/llm/acp.rs b/lib/crates/fabro-workflow/src/handler/llm/acp.rs\nindex 40929cc97..316882de6 100644\n--- a/lib/crates/fabro-workflow/src/handler/llm/acp.rs\n+++ b/lib/crates/fabro-workflow/src/handler/llm/acp.rs\n@@ -449,14 +449,15 @@ mod tests {\n         let context = Context::new();\n         let result = backend\n             .run(CodergenRunRequest {\n-                node:         &node,\n-                prompt:       \"write hello\",\n-                context:      &context,\n-                thread_id:    None,\n-                emitter:      &emitter,\n-                sandbox:      &sandbox,\n-                tool_hooks:   None,\n-                cancel_token: CancellationToken::new(),\n+                node:               &node,\n+                prompt:             \"write hello\",\n+                context:            &context,\n+                thread_id:          None,\n+                emitter:            &emitter,\n+                sandbox:            &sandbox,\n+                tool_hooks:         None,\n+                cancel_token:       CancellationToken::new(),\n+                agent_tool_runtime: fabro_agent::AgentToolRuntime::default(),\n             })\n             .await\n             .unwrap();\n@@ -516,14 +517,15 @@ mod tests {\n         let context = Context::new();\n         let result = backend\n             .run(CodergenRunRequest {\n-                node:         &node,\n-                prompt:       \"write hello\",\n-                context:      &context,\n-                thread_id:    None,\n-                emitter:      &emitter,\n-                sandbox:      &sandbox,\n-                tool_hooks:   None,\n-                cancel_token: CancellationToken::new(),\n+                node:               &node,\n+                prompt:             \"write hello\",\n+                context:            &context,\n+                thread_id:          None,\n+                emitter:            &emitter,\n+                sandbox:            &sandbox,\n+                tool_hooks:         None,\n+                cancel_token:       CancellationToken::new(),\n+                agent_tool_runtime: fabro_agent::AgentToolRuntime::default(),\n             })\n             .await\n             .unwrap();\n@@ -563,14 +565,15 @@ mod tests {\n         let context = Context::new();\n         let result = backend\n             .run(CodergenRunRequest {\n-                node:         &node,\n-                prompt:       \"write hello\",\n-                context:      &context,\n-                thread_id:    None,\n-                emitter:      &emitter,\n-                sandbox:      &sandbox,\n-                tool_hooks:   None,\n-                cancel_token: CancellationToken::new(),\n+                node:               &node,\n+                prompt:             \"write hello\",\n+                context:            &context,\n+                thread_id:          None,\n+                emitter:            &emitter,\n+                sandbox:            &sandbox,\n+                tool_hooks:         None,\n+                cancel_token:       CancellationToken::new(),\n+                agent_tool_runtime: fabro_agent::AgentToolRuntime::default(),\n             })\n             .await\n             .unwrap();\n@@ -601,14 +604,15 @@ mod tests {\n         let context = Context::new();\n         let result = backend\n             .run(CodergenRunRequest {\n-                node:         &node,\n-                prompt:       \"write hello\",\n-                context:      &context,\n-                thread_id:    None,\n-                emitter:      &emitter,\n-                sandbox:      &sandbox_dyn,\n-                tool_hooks:   None,\n-                cancel_token: CancellationToken::new(),\n+                node:               &node,\n+                prompt:             \"write hello\",\n+                context:            &context,\n+                thread_id:          None,\n+                emitter:            &emitter,\n+                sandbox:            &sandbox_dyn,\n+                tool_hooks:         None,\n+                cancel_token:       CancellationToken::new(),\n+                agent_tool_runtime: fabro_agent::AgentToolRuntime::default(),\n             })\n             .await;\n         assert!(result.is_err());\n@@ -650,14 +654,15 @@ mod tests {\n         let context = Context::new();\n         let result = backend\n             .run(CodergenRunRequest {\n-                node:         &node,\n-                prompt:       \"cancel\",\n-                context:      &context,\n-                thread_id:    None,\n-                emitter:      &emitter,\n-                sandbox:      &sandbox,\n-                tool_hooks:   None,\n-                cancel_token: CancellationToken::new(),\n+                node:               &node,\n+                prompt:             \"cancel\",\n+                context:            &context,\n+                thread_id:          None,\n+                emitter:            &emitter,\n+                sandbox:            &sandbox,\n+                tool_hooks:         None,\n+                cancel_token:       CancellationToken::new(),\n+                agent_tool_runtime: fabro_agent::AgentToolRuntime::default(),\n             })\n             .await;\n         let Err(err) = result else {\n@@ -703,14 +708,15 @@ mod tests {\n         let context = Context::new();\n         backend\n             .run(CodergenRunRequest {\n-                node:         &node,\n-                prompt:       \"write hello\",\n-                context:      &context,\n-                thread_id:    None,\n-                emitter:      &emitter,\n-                sandbox:      &sandbox,\n-                tool_hooks:   None,\n-                cancel_token: CancellationToken::new(),\n+                node:               &node,\n+                prompt:             \"write hello\",\n+                context:            &context,\n+                thread_id:          None,\n+                emitter:            &emitter,\n+                sandbox:            &sandbox,\n+                tool_hooks:         None,\n+                cancel_token:       CancellationToken::new(),\n+                agent_tool_runtime: fabro_agent::AgentToolRuntime::default(),\n             })\n             .await\n             .unwrap();\n@@ -744,14 +750,15 @@ mod tests {\n         let context = Context::new();\n         let result = backend\n             .run(CodergenRunRequest {\n-                node:         &node,\n-                prompt:       \"write hello\",\n-                context:      &context,\n-                thread_id:    None,\n-                emitter:      &emitter,\n-                sandbox:      &sandbox_dyn,\n-                tool_hooks:   None,\n-                cancel_token: CancellationToken::new(),\n+                node:               &node,\n+                prompt:             \"write hello\",\n+                context:            &context,\n+                thread_id:          None,\n+                emitter:            &emitter,\n+                sandbox:            &sandbox_dyn,\n+                tool_hooks:         None,\n+                cancel_token:       CancellationToken::new(),\n+                agent_tool_runtime: fabro_agent::AgentToolRuntime::default(),\n             })\n             .await;\n         let Err(err) = result else {\n@@ -796,14 +803,15 @@ mod tests {\n         let context = Context::new();\n         let result = backend\n             .run(CodergenRunRequest {\n-                node:         &node,\n-                prompt:       \"write hello\",\n-                context:      &context,\n-                thread_id:    None,\n-                emitter:      &emitter,\n-                sandbox:      &sandbox_dyn,\n-                tool_hooks:   None,\n-                cancel_token: CancellationToken::new(),\n+                node:               &node,\n+                prompt:             \"write hello\",\n+                context:            &context,\n+                thread_id:          None,\n+                emitter:            &emitter,\n+                sandbox:            &sandbox_dyn,\n+                tool_hooks:         None,\n+                cancel_token:       CancellationToken::new(),\n+                agent_tool_runtime: fabro_agent::AgentToolRuntime::default(),\n             })\n             .await;\n         let Err(err) = result else {\ndiff --git a/lib/crates/fabro-workflow/src/handler/llm/api.rs b/lib/crates/fabro-workflow/src/handler/llm/api.rs\nindex 38f20a5d9..41c6ca1aa 100644\n--- a/lib/crates/fabro-workflow/src/handler/llm/api.rs\n+++ b/lib/crates/fabro-workflow/src/handler/llm/api.rs\n@@ -7,7 +7,7 @@ use fabro_agent::tool_registry::{RegisteredTool, ToolContext, ToolRegistry};\n use fabro_agent::{\n     AgentEvent, AgentProfile, AnthropicProfile, CompletionCoordinator, GeminiProfile,\n     Message as AgentMessage, OpenAiProfile, Sandbox, Session, SessionOptions, StaticEnvProvider,\n-    ToolEnvProvider,\n+    ToolEnvProvider, register_question_tools,\n };\n use fabro_auth::{CredentialSource, EnvCredentialSource};\n use fabro_graphviz::graph::{AttrValue, Node};\n@@ -733,6 +733,7 @@ impl AgentApiBackend {\n         });\n \n         profile.register_subagent_tools(manager, factory, 0);\n+        register_question_tools(provider.profile_kind, profile.tool_registry_mut());\n         if let Some(services) = fabro_run_tools {\n             register_fabro_run_tools(profile.tool_registry_mut(), &services);\n         }\n@@ -971,6 +972,7 @@ impl CodergenBackend for AgentApiBackend {\n         let sandbox = request.sandbox;\n         let tool_hooks = request.tool_hooks;\n         let cancel_token = request.cancel_token;\n+        let agent_tool_runtime = request.agent_tool_runtime;\n \n         let fidelity = context.fidelity();\n         let reuse_key = if fidelity == Fidelity::Full {\n@@ -1081,7 +1083,9 @@ impl CodergenBackend for AgentApiBackend {\n                         return Err(err);\n                     }\n                 }\n-                session.process_input(prompt).await\n+                session\n+                    .process_input_with_runtime(prompt, agent_tool_runtime.clone())\n+                    .await\n             }\n             Err(err) => Err(err),\n         };\n@@ -1205,7 +1209,10 @@ impl CodergenBackend for AgentApiBackend {\n                                 return Err(err);\n                             }\n                         }\n-                        match session.process_input(prompt).await {\n+                        match session\n+                            .process_input_with_runtime(prompt, agent_tool_runtime.clone())\n+                            .await\n+                        {\n                             Ok(()) => {\n                                 succeeded = true;\n                                 break;\ndiff --git a/lib/crates/fabro-workflow/src/handler/manager_loop.rs b/lib/crates/fabro-workflow/src/handler/manager_loop.rs\nindex cf6fcaf8e..d7e1dff7b 100644\n--- a/lib/crates/fabro-workflow/src/handler/manager_loop.rs\n+++ b/lib/crates/fabro-workflow/src/handler/manager_loop.rs\n@@ -231,6 +231,7 @@ impl Handler for SubWorkflowHandler {\n \n         let parent_run = Arc::clone(&services.run);\n         let registry = Arc::clone(&services.registry);\n+        let interviewer = Arc::clone(&services.interviewer);\n         let base_env = services.base_env.clone();\n         let github_token = services.github_token.clone();\n         let inputs = services.inputs.clone();\n@@ -269,6 +270,7 @@ impl Handler for SubWorkflowHandler {\n                 engine:        Arc::new(EngineServices {\n                     run: child_run,\n                     registry,\n+                    interviewer,\n                     git_state: std::sync::RwLock::new(None),\n                     base_env,\n                     github_token,\ndiff --git a/lib/crates/fabro-workflow/src/handler/parallel.rs b/lib/crates/fabro-workflow/src/handler/parallel.rs\nindex 4ec7fbcd8..58cbe93f5 100644\n--- a/lib/crates/fabro-workflow/src/handler/parallel.rs\n+++ b/lib/crates/fabro-workflow/src/handler/parallel.rs\n@@ -304,6 +304,7 @@ impl Handler for ParallelHandler {\n         for setup in branch_setups {\n             let parent_run = Arc::clone(&services.run);\n             let registry = Arc::clone(&services.registry);\n+            let interviewer = Arc::clone(&services.interviewer);\n             let base_env = services.base_env.clone();\n             let github_token = services.github_token.clone();\n             let inputs = services.inputs.clone();\n@@ -375,6 +376,7 @@ impl Handler for ParallelHandler {\n                 let branch_services = EngineServices {\n                     run: parent_run.with_sandbox(Arc::clone(&setup.sandbox)),\n                     registry: Arc::clone(&registry),\n+                    interviewer,\n                     git_state: std::sync::RwLock::new(None),\n                     base_env: base_env.clone(),\n                     github_token: github_token.clone(),\ndiff --git a/lib/crates/fabro-workflow/src/interview_runtime.rs b/lib/crates/fabro-workflow/src/interview_runtime.rs\nnew file mode 100644\nindex 000000000..e844de8b3\n--- /dev/null\n+++ b/lib/crates/fabro-workflow/src/interview_runtime.rs\n@@ -0,0 +1,632 @@\n+use std::sync::Arc;\n+use std::sync::atomic::{AtomicUsize, Ordering};\n+use std::time::Instant;\n+\n+use async_trait::async_trait;\n+use fabro_agent::{\n+    AgentQuestion, AgentQuestionAnswer, AgentQuestionAnswerStatus, AgentQuestionRuntime,\n+};\n+use fabro_interview::{Answer, AnswerSubmission, AnswerValue, Interviewer, Question};\n+use fabro_types::{BlockedReason, InterviewOption, Principal, SystemActorKind};\n+use futures::future;\n+use tokio_util::sync::CancellationToken;\n+use ulid::Ulid;\n+\n+use crate::event::{Emitter, Event, StageScope};\n+use crate::millis_u64;\n+\n+/// Run-scoped refcount for unresolved human input. Emits `run.blocked` on the\n+/// first unresolved human/agent interview and `run.unblocked` after the last\n+/// one resolves.\n+pub(crate) struct RunInterviewBlocker {\n+    unresolved_interviews: AtomicUsize,\n+}\n+\n+impl RunInterviewBlocker {\n+    #[must_use]\n+    pub(crate) fn new() -> Self {\n+        Self {\n+            unresolved_interviews: AtomicUsize::new(0),\n+        }\n+    }\n+\n+    pub(crate) fn block(self: &Arc<Self>, emitter: Arc<Emitter>) -> RunInterviewGuard {\n+        if self.unresolved_interviews.fetch_add(1, Ordering::AcqRel) == 0 {\n+            emitter.emit(&Event::RunBlocked {\n+                blocked_reason: BlockedReason::HumanInputRequired,\n+            });\n+        }\n+        RunInterviewGuard {\n+            blocker: Arc::clone(self),\n+            emitter,\n+            resolved: false,\n+        }\n+    }\n+\n+    fn resolved(&self, emitter: &Emitter) {\n+        let mut current = self.unresolved_interviews.load(Ordering::Acquire);\n+        loop {\n+            if current == 0 {\n+                return;\n+            }\n+            match self.unresolved_interviews.compare_exchange_weak(\n+                current,\n+                current - 1,\n+                Ordering::AcqRel,\n+                Ordering::Acquire,\n+            ) {\n+                Ok(_) => {\n+                    if current == 1 {\n+                        emitter.emit(&Event::RunUnblocked);\n+                    }\n+                    return;\n+                }\n+                Err(observed) => current = observed,\n+            }\n+        }\n+    }\n+}\n+\n+impl Default for RunInterviewBlocker {\n+    fn default() -> Self {\n+        Self::new()\n+    }\n+}\n+\n+pub(crate) struct RunInterviewGuard {\n+    blocker:  Arc<RunInterviewBlocker>,\n+    emitter:  Arc<Emitter>,\n+    resolved: bool,\n+}\n+\n+impl RunInterviewGuard {\n+    pub(crate) fn resolve(mut self) {\n+        if !self.resolved {\n+            self.blocker.resolved(self.emitter.as_ref());\n+            self.resolved = true;\n+        }\n+    }\n+}\n+\n+impl Drop for RunInterviewGuard {\n+    fn drop(&mut self) {\n+        if !self.resolved {\n+            self.blocker.resolved(self.emitter.as_ref());\n+            self.resolved = true;\n+        }\n+    }\n+}\n+\n+pub(crate) struct WorkflowAgentQuestionRuntime {\n+    interviewer: Arc<dyn Interviewer>,\n+    emitter:     Arc<Emitter>,\n+    stage_scope: StageScope,\n+    stage_id:    String,\n+    blocker:     Arc<RunInterviewBlocker>,\n+}\n+\n+impl WorkflowAgentQuestionRuntime {\n+    #[must_use]\n+    pub(crate) fn new(\n+        interviewer: Arc<dyn Interviewer>,\n+        emitter: Arc<Emitter>,\n+        stage_scope: StageScope,\n+        stage_id: impl Into<String>,\n+        blocker: Arc<RunInterviewBlocker>,\n+    ) -> Self {\n+        Self {\n+            interviewer,\n+            emitter,\n+            stage_scope,\n+            stage_id: stage_id.into(),\n+            blocker,\n+        }\n+    }\n+}\n+\n+struct PreparedQuestion {\n+    agent_question: AgentQuestion,\n+    question:       Question,\n+}\n+\n+struct PendingAgentQuestionBatch {\n+    emitter:     Arc<Emitter>,\n+    stage_scope: StageScope,\n+    stage_id:    String,\n+    questions:   Vec<(String, String)>,\n+    started_at:  Instant,\n+    guard:       Option<RunInterviewGuard>,\n+}\n+\n+impl PendingAgentQuestionBatch {\n+    fn new(\n+        emitter: Arc<Emitter>,\n+        stage_scope: StageScope,\n+        stage_id: String,\n+        prepared: &[PreparedQuestion],\n+        guard: RunInterviewGuard,\n+        started_at: Instant,\n+    ) -> Self {\n+        Self {\n+            emitter,\n+            stage_scope,\n+            stage_id,\n+            questions: prepared\n+                .iter()\n+                .map(|prepared_question| {\n+                    (\n+                        prepared_question.question.id.clone(),\n+                        prepared_question.question.text.clone(),\n+                    )\n+                })\n+                .collect(),\n+            started_at,\n+            guard: Some(guard),\n+        }\n+    }\n+\n+    fn resolve(mut self) {\n+        if let Some(guard) = self.guard.take() {\n+            guard.resolve();\n+        }\n+    }\n+}\n+\n+impl Drop for PendingAgentQuestionBatch {\n+    fn drop(&mut self) {\n+        if self.guard.is_none() {\n+            return;\n+        }\n+        let duration_ms = millis_u64(self.started_at.elapsed());\n+        for (question_id, question) in &self.questions {\n+            self.emitter.emit_scoped(\n+                &Event::InterviewInterrupted {\n+                    actor: Some(Principal::System {\n+                        system_kind: SystemActorKind::Engine,\n+                    }),\n+                    question_id: question_id.clone(),\n+                    question: question.clone(),\n+                    stage: self.stage_id.clone(),\n+                    reason: \"interrupted\".to_string(),\n+                    duration_ms,\n+                },\n+                &self.stage_scope,\n+            );\n+        }\n+        if let Some(guard) = self.guard.take() {\n+            guard.resolve();\n+        }\n+    }\n+}\n+\n+#[async_trait]\n+impl AgentQuestionRuntime for WorkflowAgentQuestionRuntime {\n+    async fn ask_questions(\n+        &self,\n+        tool_call_id: &str,\n+        questions: Vec<AgentQuestion>,\n+        cancel_token: CancellationToken,\n+    ) -> Result<Vec<AgentQuestionAnswer>, String> {\n+        if questions.is_empty() {\n+            return Ok(Vec::new());\n+        }\n+\n+        let prepared = questions\n+            .into_iter()\n+            .enumerate()\n+            .map(|(index, question)| self.prepare_question(tool_call_id, index, question))\n+            .collect::<Vec<_>>();\n+\n+        for prepared_question in &prepared {\n+            let question = &prepared_question.question;\n+            self.emitter.emit_scoped(\n+                &Event::InterviewStarted {\n+                    question_id:     question.id.clone(),\n+                    question:        question.text.clone(),\n+                    stage:           self.stage_id.clone(),\n+                    question_type:   question.question_type.to_string(),\n+                    options:         question.options.clone(),\n+                    allow_freeform:  question.allow_freeform,\n+                    timeout_seconds: None,\n+                    context_display: question.context_display.clone(),\n+                },\n+                &self.stage_scope,\n+            );\n+        }\n+\n+        let interview_start = Instant::now();\n+        let cleanup = PendingAgentQuestionBatch::new(\n+            Arc::clone(&self.emitter),\n+            self.stage_scope.clone(),\n+            self.stage_id.clone(),\n+            &prepared,\n+            self.blocker.block(Arc::clone(&self.emitter)),\n+            interview_start,\n+        );\n+        let ask_all = future::join_all(\n+            prepared\n+                .iter()\n+                .map(|prepared_question| self.interviewer.ask(prepared_question.question.clone())),\n+        );\n+        tokio::pin!(ask_all);\n+\n+        let answers = tokio::select! {\n+            submissions = &mut ask_all => Some(submissions),\n+            () = cancel_token.cancelled() => None,\n+        };\n+\n+        let results = match answers {\n+            Some(submissions) => prepared\n+                .iter()\n+                .zip(submissions)\n+                .map(|(prepared_question, submission)| {\n+                    self.emit_submission_event(\n+                        prepared_question,\n+                        &submission,\n+                        millis_u64(interview_start.elapsed()),\n+                    );\n+                    answer_from_submission(&prepared_question.agent_question, &submission)\n+                })\n+                .collect::<Vec<_>>(),\n+            None => prepared\n+                .iter()\n+                .map(|prepared_question| {\n+                    self.emit_interrupted(\n+                        prepared_question,\n+                        Some(Principal::System {\n+                            system_kind: SystemActorKind::Engine,\n+                        }),\n+                        \"interrupted\",\n+                        millis_u64(interview_start.elapsed()),\n+                    );\n+                    AgentQuestionAnswer {\n+                        original_id:       prepared_question.agent_question.original_id.clone(),\n+                        original_question: prepared_question\n+                            .agent_question\n+                            .original_question\n+                            .clone(),\n+                        answers:           Vec::new(),\n+                        status:            AgentQuestionAnswerStatus::Interrupted,\n+                    }\n+                })\n+                .collect::<Vec<_>>(),\n+        };\n+\n+        cleanup.resolve();\n+        Ok(results)\n+    }\n+}\n+\n+impl WorkflowAgentQuestionRuntime {\n+    fn prepare_question(\n+        &self,\n+        tool_call_id: &str,\n+        index: usize,\n+        agent_question: AgentQuestion,\n+    ) -> PreparedQuestion {\n+        let mut question = Question::new(agent_question.text.clone(), agent_question.question_type);\n+        question.id = internal_question_id(&self.stage_scope, tool_call_id, index);\n+        question.options.clone_from(&agent_question.options);\n+        question.allow_freeform = agent_question.allow_freeform;\n+        question.stage.clone_from(&self.stage_id);\n+        question.metadata.insert(\n+            \"agent.tool_call_id\".to_string(),\n+            serde_json::json!(tool_call_id),\n+        );\n+        question.metadata.insert(\n+            \"agent.original_question\".to_string(),\n+            serde_json::json!(agent_question.original_question),\n+        );\n+        if let Some(original_id) = &agent_question.original_id {\n+            question.metadata.insert(\n+                \"agent.original_id\".to_string(),\n+                serde_json::json!(original_id),\n+            );\n+        }\n+        if let Some(header) = &agent_question.header {\n+            question\n+                .metadata\n+                .insert(\"agent.header\".to_string(), serde_json::json!(header));\n+        }\n+        PreparedQuestion {\n+            agent_question,\n+            question,\n+        }\n+    }\n+\n+    fn emit_submission_event(\n+        &self,\n+        prepared: &PreparedQuestion,\n+        submission: &AnswerSubmission,\n+        duration_ms: u64,\n+    ) {\n+        match submission.answer.value {\n+            AnswerValue::Timeout => self.emitter.emit_scoped(\n+                &Event::InterviewTimeout {\n+                    actor: Some(Principal::System {\n+                        system_kind: SystemActorKind::Timeout,\n+                    }),\n+                    question_id: prepared.question.id.clone(),\n+                    question: prepared.question.text.clone(),\n+                    stage: self.stage_id.clone(),\n+                    duration_ms,\n+                },\n+                &self.stage_scope,\n+            ),\n+            AnswerValue::Interrupted => self.emit_interrupted(\n+                prepared,\n+                Some(submission.actor.clone()),\n+                \"interrupted\",\n+                duration_ms,\n+            ),\n+            AnswerValue::Cancelled => self.emit_interrupted(\n+                prepared,\n+                Some(submission.actor.clone()),\n+                \"cancelled\",\n+                duration_ms,\n+            ),\n+            _ => self.emitter.emit_scoped(\n+                &Event::InterviewCompleted {\n+                    actor: Some(submission.actor.clone()),\n+                    question_id: prepared.question.id.clone(),\n+                    question: prepared.question.text.clone(),\n+                    answer: answer_labels(&prepared.question.options, &submission.answer)\n+                        .join(\", \"),\n+                    duration_ms,\n+                },\n+                &self.stage_scope,\n+            ),\n+        }\n+    }\n+\n+    fn emit_interrupted(\n+        &self,\n+        prepared: &PreparedQuestion,\n+        actor: Option<Principal>,\n+        reason: &str,\n+        duration_ms: u64,\n+    ) {\n+        self.emitter.emit_scoped(\n+            &Event::InterviewInterrupted {\n+                actor,\n+                question_id: prepared.question.id.clone(),\n+                question: prepared.question.text.clone(),\n+                stage: self.stage_id.clone(),\n+                reason: reason.to_string(),\n+                duration_ms,\n+            },\n+            &self.stage_scope,\n+        );\n+    }\n+}\n+\n+fn answer_from_submission(\n+    agent_question: &AgentQuestion,\n+    submission: &AnswerSubmission,\n+) -> AgentQuestionAnswer {\n+    let status = match &submission.answer.value {\n+        AnswerValue::Cancelled => AgentQuestionAnswerStatus::Cancelled,\n+        AnswerValue::Interrupted => AgentQuestionAnswerStatus::Interrupted,\n+        AnswerValue::Skipped => AgentQuestionAnswerStatus::Skipped,\n+        AnswerValue::Timeout => AgentQuestionAnswerStatus::Timeout,\n+        _ => AgentQuestionAnswerStatus::Answered,\n+    };\n+    let answers = if status == AgentQuestionAnswerStatus::Answered {\n+        answer_labels(&agent_question.options, &submission.answer)\n+    } else {\n+        Vec::new()\n+    };\n+    AgentQuestionAnswer {\n+        original_id: agent_question.original_id.clone(),\n+        original_question: agent_question.original_question.clone(),\n+        answers,\n+        status,\n+    }\n+}\n+\n+fn answer_labels(options: &[InterviewOption], answer: &Answer) -> Vec<String> {\n+    match &answer.value {\n+        AnswerValue::Selected(key) => {\n+            vec![option_label(options, key, answer.selected_option.as_ref())]\n+        }\n+        AnswerValue::MultiSelected(keys) => keys\n+            .iter()\n+            .map(|key| option_label(options, key, None))\n+            .collect(),\n+        AnswerValue::Text(text) => vec![text.clone()],\n+        AnswerValue::Yes => vec![\"yes\".to_string()],\n+        AnswerValue::No => vec![\"no\".to_string()],\n+        AnswerValue::Cancelled => vec![\"cancelled\".to_string()],\n+        AnswerValue::Interrupted => vec![\"interrupted\".to_string()],\n+        AnswerValue::Skipped => vec![\"skipped\".to_string()],\n+        AnswerValue::Timeout => vec![\"timeout\".to_string()],\n+    }\n+}\n+\n+fn option_label(\n+    options: &[InterviewOption],\n+    key: &str,\n+    selected_option: Option<&InterviewOption>,\n+) -> String {\n+    selected_option\n+        .filter(|option| option.key == key)\n+        .or_else(|| options.iter().find(|option| option.key == key))\n+        .map_or_else(|| key.to_string(), |option| option.label.clone())\n+}\n+\n+fn internal_question_id(scope: &StageScope, tool_call_id: &str, index: usize) -> String {\n+    format!(\n+        \"agentq-{}-v{}-{}-{}-{}\",\n+        slug(&scope.node_id),\n+        scope.visit,\n+        slug(tool_call_id),\n+        index + 1,\n+        Ulid::new()\n+    )\n+}\n+\n+fn slug(value: &str) -> String {\n+    let mut out = value\n+        .chars()\n+        .filter_map(|ch| {\n+            if ch.is_ascii_alphanumeric() {\n+                Some(ch.to_ascii_lowercase())\n+            } else if matches!(ch, '-' | '_') {\n+                Some(ch)\n+            } else {\n+                None\n+            }\n+        })\n+        .take(48)\n+        .collect::<String>();\n+    if out.is_empty() {\n+        out.push('x');\n+    }\n+    out\n+}\n+\n+#[cfg(test)]\n+mod tests {\n+    use fabro_interview::ControlInterviewer;\n+    use fabro_types::{EventBody, RunId};\n+\n+    use super::*;\n+\n+    #[test]\n+    fn answer_labels_return_user_facing_labels_in_submission_order() {\n+        let options = vec![\n+            InterviewOption {\n+                key: \"a\".to_string(),\n+                label: \"Alpha\".to_string(),\n+                ..InterviewOption::default()\n+            },\n+            InterviewOption {\n+                key: \"b\".to_string(),\n+                label: \"Beta\".to_string(),\n+                ..InterviewOption::default()\n+            },\n+        ];\n+        let answer = Answer::multi_selected(vec![\"b\".to_string(), \"a\".to_string()]);\n+\n+        assert_eq!(answer_labels(&options, &answer), vec![\"Beta\", \"Alpha\"]);\n+    }\n+\n+    #[test]\n+    fn internal_question_id_includes_stage_visit_and_tool_call_context() {\n+        let scope = StageScope {\n+            node_id:            \"Review Changes\".to_string(),\n+            visit:              3,\n+            parallel_group_id:  None,\n+            parallel_branch_id: None,\n+        };\n+\n+        let id = internal_question_id(&scope, \"call_123\", 1);\n+\n+        assert!(id.starts_with(\"agentq-reviewchanges-v3-call_123-2-\"));\n+    }\n+\n+    #[tokio::test]\n+    async fn batch_questions_are_all_started_before_run_is_blocked_and_return_labels() {\n+        let interviewer = Arc::new(ControlInterviewer::new());\n+        let emitter = Arc::new(Emitter::new(RunId::new()));\n+        let events = Arc::new(std::sync::Mutex::new(Vec::new()));\n+        emitter.on_event({\n+            let events = Arc::clone(&events);\n+            move |event| events.lock().unwrap().push(event.clone())\n+        });\n+        let runtime = WorkflowAgentQuestionRuntime::new(\n+            interviewer.clone(),\n+            Arc::clone(&emitter),\n+            StageScope {\n+                node_id:            \"ask\".to_string(),\n+                visit:              1,\n+                parallel_group_id:  None,\n+                parallel_branch_id: None,\n+            },\n+            \"ask\",\n+            Arc::new(RunInterviewBlocker::new()),\n+        );\n+        let option = InterviewOption {\n+            key:         \"ship\".to_string(),\n+            label:       \"Ship it\".to_string(),\n+            description: Some(\"Deploy\".to_string()),\n+            preview:     Some(\"preview\".to_string()),\n+        };\n+\n+        let ask = tokio::spawn(async move {\n+            runtime\n+                .ask_questions(\n+                    \"call_1\",\n+                    vec![\n+                        AgentQuestion {\n+                            original_id:       Some(\"q1\".to_string()),\n+                            original_question: \"First?\".to_string(),\n+                            header:            None,\n+                            text:              \"First?\".to_string(),\n+                            question_type:     fabro_types::QuestionType::MultipleChoice,\n+                            options:           vec![option.clone()],\n+                            allow_freeform:    true,\n+                        },\n+                        AgentQuestion {\n+                            original_id:       Some(\"q2\".to_string()),\n+                            original_question: \"Second?\".to_string(),\n+                            header:            None,\n+                            text:              \"Second?\".to_string(),\n+                            question_type:     fabro_types::QuestionType::MultipleChoice,\n+                            options:           vec![option.clone()],\n+                            allow_freeform:    true,\n+                        },\n+                    ],\n+                    CancellationToken::new(),\n+                )\n+                .await\n+                .unwrap()\n+        });\n+\n+        tokio::task::yield_now().await;\n+        let question_ids = {\n+            let events = events.lock().unwrap();\n+            assert!(matches!(events[0].body, EventBody::InterviewStarted(_)));\n+            assert!(matches!(events[1].body, EventBody::InterviewStarted(_)));\n+            assert!(matches!(events[2].body, EventBody::RunBlocked(_)));\n+            events\n+                .iter()\n+                .filter_map(|event| match &event.body {\n+                    EventBody::InterviewStarted(props) => Some(props.question_id.clone()),\n+                    _ => None,\n+                })\n+                .collect::<Vec<_>>()\n+        };\n+\n+        for question_id in question_ids {\n+            let option = InterviewOption {\n+                key: \"ship\".to_string(),\n+                label: \"Ship it\".to_string(),\n+                ..InterviewOption::default()\n+            };\n+            interviewer\n+                .submit(\n+                    &question_id,\n+                    AnswerSubmission::system(\n+                        Answer::selected(\"ship\", option),\n+                        SystemActorKind::Engine,\n+                    ),\n+                )\n+                .await\n+                .unwrap();\n+        }\n+\n+        let answers = ask.await.unwrap();\n+\n+        assert_eq!(answers.len(), 2);\n+        assert_eq!(answers[0].answers, vec![\"Ship it\"]);\n+        assert_eq!(answers[1].answers, vec![\"Ship it\"]);\n+        assert!(\n+            events\n+                .lock()\n+                .unwrap()\n+                .iter()\n+                .any(|event| matches!(event.body, EventBody::RunUnblocked(_)))\n+        );\n+    }\n+}\ndiff --git a/lib/crates/fabro-workflow/src/lib.rs b/lib/crates/fabro-workflow/src/lib.rs\nindex 194a0ca2c..42f5c7f43 100644\n--- a/lib/crates/fabro-workflow/src/lib.rs\n+++ b/lib/crates/fabro-workflow/src/lib.rs\n@@ -298,6 +298,7 @@ pub mod github_token_source;\n pub(crate) mod graph;\n pub mod handler;\n mod hook_context;\n+mod interview_runtime;\n #[allow(\n     dead_code,\n     reason = \"The lifecycle module remains crate-visible for tests and pending integrations.\"\ndiff --git a/lib/crates/fabro-workflow/src/pipeline/initialize.rs b/lib/crates/fabro-workflow/src/pipeline/initialize.rs\nindex 953f39d2d..3165f7855 100644\n--- a/lib/crates/fabro-workflow/src/pipeline/initialize.rs\n+++ b/lib/crates/fabro-workflow/src/pipeline/initialize.rs\n@@ -675,6 +675,7 @@ pub async fn initialize(\n     let engine = Arc::new(EngineServices {\n         run: Arc::clone(&run_services),\n         registry,\n+        interviewer: Arc::clone(&options.interviewer),\n         git_state: std::sync::RwLock::new(None),\n         base_env,\n         github_token,\ndiff --git a/lib/crates/fabro-workflow/src/services.rs b/lib/crates/fabro-workflow/src/services.rs\nindex 72c5499f9..9210998d3 100644\n--- a/lib/crates/fabro-workflow/src/services.rs\n+++ b/lib/crates/fabro-workflow/src/services.rs\n@@ -9,6 +9,7 @@ use fabro_auth::CredentialSource;\n #[cfg(test)]\n use fabro_auth::ResolvedCredentials;\n use fabro_hooks::{HookContext, HookDecision, HookExecutionContext, HookRunner};\n+use fabro_interview::Interviewer;\n use fabro_model::{Catalog, ProviderId};\n use fabro_types::{ManifestPath, RunId};\n use tokio_util::sync::CancellationToken;\n@@ -16,6 +17,7 @@ use tokio_util::sync::CancellationToken;\n use crate::event::Emitter;\n use crate::github_token_source::GitHubTokenSource;\n use crate::handler::HandlerRegistry;\n+use crate::interview_runtime::RunInterviewBlocker;\n use crate::run_metadata::{RunMetadataRuntime, RunMetadataWriterHandle};\n use crate::runtime_store::RunStoreHandle;\n use crate::sandbox_git::GitState;\n@@ -91,19 +93,20 @@ pub struct FabroRunToolServices {\n /// does NOT count as cancellation.\n #[derive(Clone)]\n pub struct RunServices {\n-    pub run_store:               RunStoreHandle,\n-    pub emitter:                 Arc<Emitter>,\n-    pub sandbox:                 Arc<dyn Sandbox>,\n-    pub hook_runner:             Option<Arc<HookRunner>>,\n-    pub locations:               RunLocations,\n-    pub(crate) cancel_token:     CancellationToken,\n-    pub provider_id:             ProviderId,\n-    pub model:                   String,\n-    pub llm_source:              Arc<dyn CredentialSource>,\n-    pub catalog:                 Arc<Catalog>,\n-    pub(crate) sandbox_git:      Arc<SandboxGitRuntime>,\n-    pub(crate) metadata_runtime: Arc<RunMetadataRuntime>,\n-    pub(crate) metadata_writer:  Option<RunMetadataWriterHandle>,\n+    pub run_store:                RunStoreHandle,\n+    pub emitter:                  Arc<Emitter>,\n+    pub sandbox:                  Arc<dyn Sandbox>,\n+    pub hook_runner:              Option<Arc<HookRunner>>,\n+    pub locations:                RunLocations,\n+    pub(crate) cancel_token:      CancellationToken,\n+    pub provider_id:              ProviderId,\n+    pub model:                    String,\n+    pub llm_source:               Arc<dyn CredentialSource>,\n+    pub catalog:                  Arc<Catalog>,\n+    pub(crate) sandbox_git:       Arc<SandboxGitRuntime>,\n+    pub(crate) metadata_runtime:  Arc<RunMetadataRuntime>,\n+    pub(crate) metadata_writer:   Option<RunMetadataWriterHandle>,\n+    pub(crate) interview_blocker: Arc<RunInterviewBlocker>,\n }\n \n impl RunServices {\n@@ -137,6 +140,7 @@ impl RunServices {\n             sandbox_git,\n             metadata_runtime,\n             metadata_writer,\n+            interview_blocker: Arc::new(RunInterviewBlocker::new()),\n         })\n     }\n \n@@ -224,6 +228,7 @@ impl RunServices {\n pub struct EngineServices {\n     pub run:              Arc<RunServices>,\n     pub registry:         Arc<HandlerRegistry>,\n+    pub interviewer:      Arc<dyn Interviewer>,\n     /// Git state for the current run. Set via `set_git_state` at the start of\n     /// `execute` and read by parallel/fan-in handlers.\n     pub(crate) git_state: std::sync::RwLock<Option<Arc<GitState>>>,\n@@ -330,6 +335,7 @@ impl EngineServices {\n                 None,\n             ),\n             registry:        Arc::new(HandlerRegistry::new(Box::new(start::StartHandler))),\n+            interviewer:     Arc::new(fabro_interview::AutoApproveInterviewer::engine()),\n             git_state:       std::sync::RwLock::new(None),\n             base_env:        HashMap::new(),\n             github_token:    None,\ndiff --git a/lib/crates/fabro-workflow/src/test_support.rs b/lib/crates/fabro-workflow/src/test_support.rs\nindex aa9c8117a..43b0c27cf 100644\n--- a/lib/crates/fabro-workflow/src/test_support.rs\n+++ b/lib/crates/fabro-workflow/src/test_support.rs\n@@ -7,6 +7,7 @@ use std::time::Duration;\n use fabro_agent::Sandbox;\n use fabro_auth::{CredentialSource, EnvCredentialSource};\n use fabro_graphviz::graph::Graph as GvGraph;\n+use fabro_interview::AutoApproveInterviewer;\n use fabro_model::Catalog;\n use fabro_store::{ArtifactStore, Database, RunProjection};\n use object_store::local::LocalFileSystem;\n@@ -175,6 +176,7 @@ async fn initialized(\n                     None,\n                 ),\n                 registry:        Arc::new(registry),\n+                interviewer:     Arc::new(AutoApproveInterviewer::engine()),\n                 git_state:       std::sync::RwLock::new(None),\n                 base_env:        options.env,\n                 github_token:    None,\ndiff --git a/lib/packages/fabro-api-client/src/.openapi-generator/FILES b/lib/packages/fabro-api-client/src/.openapi-generator/FILES\nindex 653eea117..278e89ec9 100644\n--- a/lib/packages/fabro-api-client/src/.openapi-generator/FILES\n+++ b/lib/packages/fabro-api-client/src/.openapi-generator/FILES\n@@ -24,7 +24,6 @@ index.ts\n models/agent-permissions.ts\n models/aggregate-billing-totals.ts\n models/aggregate-billing.ts\n-models/api-question-option.ts\n models/api-question.ts\n models/append-event-response.ts\n models/approval-mode.ts\n@@ -410,6 +409,10 @@ models/system-resources-response.ts\n models/system-run-counts.ts\n models/timeline-entry-response.ts\n models/tls-mode.ts\n+models/todo-list-kind.ts\n+models/todo-list-projection.ts\n+models/todo-projection.ts\n+models/todo-status.ts\n models/update-run-parent-request.ts\n models/update-run-request.ts\n models/user-response.ts\n@@ -425,4 +428,4 @@ models/workflow-ref.ts\n models/workflow-reference.ts\n models/workflow-schedule-summary.ts\n models/workflow-settings.ts\n-models/write-blob-response.ts\n\\ No newline at end of file\n+models/write-blob-response.ts\ndiff --git a/lib/packages/fabro-api-client/src/models/api-question-option.ts b/lib/packages/fabro-api-client/src/models/api-question-option.ts\ndeleted file mode 100644\nindex ddbe5577d..000000000\n--- a/lib/packages/fabro-api-client/src/models/api-question-option.ts\n+++ /dev/null\n@@ -1,29 +0,0 @@\n-/* tslint:disable */\n-/* eslint-disable */\n-/**\n- * Fabro Run API\n- * HTTP API for managing Fabro workflow run executions.\n- *\n- * The version of the OpenAPI document: 0.1.0\n- *\n- *\n- * NOTE: This class is auto generated by OpenAPI Generator (https://openapi-generator.tech).\n- * https://openapi-generator.tech\n- * Do not edit the class manually.\n- */\n-\n-\n-\n-/**\n- * A selectable option for a multiple-choice or multi-select question.\n- */\n-export interface ApiQuestionOption {\n-    /**\n-     * Machine-readable option key used when submitting an answer.\n-     */\n-    'key': string;\n-    /**\n-     * Human-readable label displayed to the user.\n-     */\n-    'label': string;\n-}\ndiff --git a/lib/packages/fabro-api-client/src/models/api-question.ts b/lib/packages/fabro-api-client/src/models/api-question.ts\nindex 0cf9e785b..74908ff63 100644\n--- a/lib/packages/fabro-api-client/src/models/api-question.ts\n+++ b/lib/packages/fabro-api-client/src/models/api-question.ts\n@@ -15,7 +15,7 @@\n \n // May contain unused imports in some cases\n // @ts-ignore\n-import type { ApiQuestionOption } from './api-question-option';\n+import type { InterviewOption } from './interview-option';\n // May contain unused imports in some cases\n // @ts-ignore\n import type { QuestionType } from './question-type';\n@@ -40,7 +40,7 @@ export interface ApiQuestion {\n     /**\n      * Available options for selection-based questions. Empty for freeform questions.\n      */\n-    'options': Array<ApiQuestionOption>;\n+    'options': Array<InterviewOption>;\n     /**\n      * Whether the user may provide freeform text in addition to selecting options.\n      */\ndiff --git a/lib/packages/fabro-api-client/src/models/index.ts b/lib/packages/fabro-api-client/src/models/index.ts\nindex 23a07a598..c2b6db816 100644\n--- a/lib/packages/fabro-api-client/src/models/index.ts\n+++ b/lib/packages/fabro-api-client/src/models/index.ts\n@@ -2,7 +2,6 @@ export * from './agent-permissions';\n export * from './aggregate-billing';\n export * from './aggregate-billing-totals';\n export * from './api-question';\n-export * from './api-question-option';\n export * from './append-event-response';\n export * from './approval-mode';\n export * from './artifact-batch-upload-entry';\n@@ -405,4 +404,4 @@ export * from './workflow-ref';\n export * from './workflow-reference';\n export * from './workflow-schedule-summary';\n export * from './workflow-settings';\n-export * from './write-blob-response';\n\\ No newline at end of file\n+export * from './write-blob-response';\ndiff --git a/lib/packages/fabro-api-client/src/models/interview-option.ts b/lib/packages/fabro-api-client/src/models/interview-option.ts\nindex 4215b382e..16c536f80 100644\n--- a/lib/packages/fabro-api-client/src/models/interview-option.ts\n+++ b/lib/packages/fabro-api-client/src/models/interview-option.ts\n@@ -18,6 +18,20 @@\n  * Option stored with an interview question in the event log.\n  */\n export interface InterviewOption {\n+    /**\n+     * Machine-readable option key used when submitting an answer.\n+     */\n     'key': string;\n+    /**\n+     * Human-readable label displayed to the user.\n+     */\n     'label': string;\n+    /**\n+     * Optional untrusted model-authored option description for display.\n+     */\n+    'description'?: string | null;\n+    /**\n+     * Optional untrusted model-authored option preview captured for clients.\n+     */\n+    'preview'?: string | null;\n }\ndiff --git a/lib/packages/fabro-api-client/src/models/run-projection.ts b/lib/packages/fabro-api-client/src/models/run-projection.ts\nindex 0505c0cd0..0ab7b799f 100644\n--- a/lib/packages/fabro-api-client/src/models/run-projection.ts\n+++ b/lib/packages/fabro-api-client/src/models/run-projection.ts\n@@ -87,4 +87,4 @@ export interface RunProjection {\n      * Map from StageId (`node_id@visit`) to stage projection data.\n      */\n     'stages': { [key: string]: StageProjection; };\n-}\n\\ No newline at end of file\n+}\n",
+        "summary": {
+          "files_changed": 42,
+          "additions": 2059,
+          "deletions": 326
+        }
+      }
+    },
+    {
+      "seq": 0,
+      "checkpoint": {
+        "timestamp": "2026-05-23T17:36:14.594548Z",
+        "current_node": "simplify_opus",
+        "completed_nodes": [
+          "start",
+          "toolchain",
+          "preflight_compile",
+          "preflight_lint",
+          "implement",
+          "simplify_opus"
+        ],
+        "node_retries": {},
+        "context_values": {
           "internal.fidelity": "compact",
           "response.implement": "Implemented the mid-stage agent interview tools and metadata plumbing.\n\nSummary of changes:\n- Extended canonical `InterviewOption` with optional `description` and `preview`.\n- Updated OpenAPI and regenerated the TypeScript API client:\n  - `ApiQuestion.options` now uses `InterviewOption`.\n  - Removed stale `ApiQuestionOption` export/file.\n- Added model-native agent question tools:\n  - OpenAI: `request_user_input`\n  - Anthropic: `AskUserQuestion`\n  - Gemini: no question tool\n- Added per-turn `AgentToolRuntime` so cached sessions receive current workflow/stage interview context dynamically.\n- Enforced question-tool round rules:\n  - question tool runs alone\n  - mixed tool peers get model-visible errors\n  - multiple separate question-tool calls execute only the first and tell the model to batch\n- Added shared run-level interview blocker/guard used by both human gates and agent questions.\n- Added workflow runtime for agent-originated batched questions:\n  - emits/registers all questions first\n  - blocks once per batch\n  - waits concurrently\n  - emits completion/interruption/timeout events per question\n  - cleans up pending questions on cancellation/drop\n- Preserved answer endpoints/paths through existing pending interview projection and submission flow.\n- Updated web/Slack display:\n  - web shows option descriptions\n  - Slack renders option descriptions where practical\n  - preview is captured/exposed but not specially rendered\n- Added/updated tests across agent, workflow, API, store, Slack, and web.\n\nValidation run:\n- `cargo check -q -p fabro-agent -p fabro-workflow -p fabro-api -p fabro-server -p fabro-slack -p fabro-store -p fabro-cli --tests` ✅\n- `cargo +nightly-2026-04-14 fmt --check --all` ✅\n- `cargo +nightly-2026-04-14 clippy -q -p fabro-agent -p fabro-workflow -p fabro-api -p fabro-server -p fabro-slack -p fabro-store -p fabro-cli --all-targets -- -D warnings` ✅\n- `cargo nextest run -p fabro-agent -p fabro-workflow -p fabro-api ...focused question/API tests...` ✅\n- `cargo nextest run -p fabro-interview` ✅\n- `cargo nextest run -p fabro-slack -p fabro-store --lib --tests` ✅\n- `cd apps/fabro-web && bun test ./app/components/interview-dock.test.tsx ./app/components/stage-renderers/helpers.test.ts` ✅\n- `cd apps/fabro-web && bun run typecheck` ✅\n- `cd lib/packages/fabro-api-client && bun run typecheck` ✅\n\nNoted failures:\n- `cargo nextest run -p fabro-server ...` fails on `server::tests::get_graph_returns_svg`; rerun of the single test also fails. The error is the graph renderer subprocess returning nextest output (`running 0 tests`) instead of SVG, unrelated to the interview changes.\n- Full `cd apps/fabro-web && bun test` still has existing unrelated failures in run-detail/lifecycle tests; focused changed web tests pass and typecheck passes.",
           "failure_signature": "",
           "failure_class": "",
-          "graph.rankdir": "LR",
           "thread.toolchain.current_node": "preflight_compile",
           "command.output": "blob://sha256/12ae32cb1ec02d01eda3581b127c1fee3b0dc53572ed6baf239721a03d82e126",
-          "internal.thread_id": "preflight_lint",
+          "internal.thread_id": "implement",
           "thread.preflight_lint.current_node": "implement",
-          "last_stage": "implement",
+          "internal.retry_count.simplify_opus": 0,
+          "internal.retry_count.preflight_lint": 0,
+          "internal.retry_count.preflight_compile": 0,
+          "response.simplify_opus": "All clean. Final summary:\n\n## Cleanup Applied\n\nAggregated findings from three reviews and applied 7 focused fixes (3 files, net −31 lines):\n\n**Efficiency**\n- `bounded_display_field` now truncates in a single `char_indices().nth(...)` pass instead of `chars().count()` + `chars().take()`.\n- Dropped redundant `Ulid::new()` from `internal_question_id` — slug + stage visit + tool-call-id + index is already unique.\n\n**Code quality**\n- De-duped the access-denial / hook-block / normal-execution event triple via shared `emit_tool_call_started` and `emit_tool_call_result` helpers in `tool_execution.rs`, dropping ~25 lines of copy-pasted emit code.\n- Cleaned up dead `unwrap_or_else` fallbacks in `format_anthropic_answers` (infallible string serialization).\n- Extracted repeated \"root agent session\" error string into `ROOT_SESSION_REQUIRED_ERROR` constant.\n- Renamed `option_label` to `label_for_key`, dropping the redundant `selected_option.filter(key)` defensive check. Preserved the `selected_option`-first lookup for `Selected` (control-protocol path supplies `Selected(key)` with `selected_option: None`).\n- Removed the unused `Default for RunInterviewBlocker` impl.\n- De-duped `RunInterviewGuard::resolve` and `Drop` via a shared `resolve_in_place` method.\n\n**Findings noted but not applied** (with rationale):\n- Lifting `agent_tool_runtime` from a parameter into a task-local set once at the top: large cross-crate refactor; skipped to keep scope bounded.\n- Unifying OpenAI/Anthropic question tools into one schema: schemas intentionally differ (`id`/`header` required for OpenAI, `multiSelect` and `preview` for Anthropic); separation matches provider contracts.\n- `AgentToolRuntime` one-field wrapper: keeping a struct allows future expansion at no extra friction.\n- `register_question_tools` moving into `AgentProfile` trait: cross-crate refactor; current explicit dispatch is fine for two profiles.\n- Cleaning `ControlInterviewer.pending` on agent-batch cancel: needs a new `Interviewer` API; flagged as the most substantial follow-up.\n- `..InterviewOption::default()` adoption in tests: high-churn, low-value compared to other items in this pass.\n- The `mem::forget` suggestion for `RunInterviewGuard::resolve` is buggy — it would leak the `Arc<RunInterviewBlocker>` and `Arc<Emitter>` refs. Used a shared method instead.\n\n**Verification**\n- `cargo nextest run -p fabro-agent -p fabro-workflow -p fabro-interview -p fabro-server`: all tests pass except 3 pre-existing flaky SVG-render tests (confirmed by stashing changes and reproducing the same failures on the pre-cleanup tree).\n- `cargo +nightly-2026-04-14 clippy --workspace --all-targets -- -D warnings`: clean.\n- `cargo +nightly-2026-04-14 fmt --check --all`: clean.",
+          "internal.retry_count.start": 0,
+          "internal.run_id": "01KSATQNAXG41FHKV0QH5N1QGC",
+          "thread.start.current_node": "toolchain",
+          "graph.goal": "# Mid-Stage Agent Interview Tools\n\n## Summary\n\nAdd model-native question tools that let agents pause mid-stage and ask the human for input through Fabro's existing interview system.\n\nOpenAI-profile agents get `request_user_input`; Anthropic-profile agents get `AskUserQuestion`. When either tool is called, Fabro creates pending interview questions, surfaces them through the existing web/API/Slack paths, waits for answers, then returns provider-shaped tool results so the model can continue the same stage.\n\n## Key Changes\n\n- Extend the existing interview contract without type sprawl:\n  - Add optional `description` and `preview` fields to the canonical `fabro_types::InterviewOption`; reuse that type through `fabro-api` replacements instead of introducing `AgentQuestionOption`, API-only aliases, or adapter-only duplicate types.\n  - Update OpenAPI, generated Rust/TypeScript clients, event conversion, projection, Slack/web mappers, and the existing `with_replacement(\"InterviewOption\", \"fabro_types::InterviewOption\", ...)` parity tests.\n  - Treat both fields as untrusted model-authored display data. Store and expose them after enforcing bounded lengths; truncate or reject oversized values consistently before persistence.\n  - Initial UI behavior: display `description` under option labels where practical. Capture and expose `preview`, but do not render preview content specially in web or Slack v1.\n\n- Add a shared run-level interview runtime:\n  - Move the private human-node blocked-state refcount into a reusable run-level guard used by both `HumanHandler` and agent question tools, so `RunUnblocked` is emitted only when all human and agent interviews for the run are resolved.\n  - Runtime accepts the interviewer, workflow emitter, stage scope, stage id, tool call id, and normalized questions.\n  - Support batch asks as a first-class operation: emit/register all questions first, mark the run blocked once, await all answers concurrently, then emit completion/timeout/interrupted events per question and unblock when the batch resolves.\n  - Batch support applies only to multiple `questions[]` inside one question-tool call. Do not aggregate multiple separate question-tool calls from the same model round.\n  - Generate safe internal question IDs with a ULID/UUID plus stage visit/tool-call context; store original model question IDs/text in question metadata for provider result mapping.\n\n- Add provider-specific agent tools:\n  - `request_user_input` for `AgentProfileKind::OpenAi`.\n    - Accept Codex-compatible schema: `questions[]` with `id`, `header`, `question`, and `options[] { label, description }`.\n    - Normalize each question to `QuestionType::MultipleChoice` with `allow_freeform: true`.\n    - Return JSON text matching Codex shape, keyed by the original model question ID: `{\"answers\":{\"id\":{\"answers\":[\"...\"]}}}`.\n  - `AskUserQuestion` for `AgentProfileKind::Anthropic`.\n    - Accept Claude-compatible schema: `questions[]` with `question`, `header`, `options[] { label, description, preview? }`, and `multiSelect`.\n    - Normalize single-select to `MultipleChoice`, multi-select to `MultiSelect`, always with `allow_freeform: true`.\n    - Return Claude-style tool result text keyed by the original question text: `User has answered your questions: \"...question...\"=\"answer\". You can now continue...`.\n  - Answer formatting for both tools returns user-facing option labels to the model. Preserve internal option keys for validation and event storage. For multi-select, preserve the submission order supplied by the answer path.\n\n- Thread workflow interview context into agent tool execution:\n  - Add an explicit per-turn agent tool runtime context passed into `process_input` or an adjacent `process_input_with_runtime` API. It carries the interviewer, workflow emitter, stage scope/id, shared block guard, and provider answer formatter.\n  - Do not capture stage-specific interview handles in the profile registry or cached session construction; cached full-fidelity sessions must receive the current turn's stage context dynamically.\n  - Child/subagent sessions must not expose these question tools. If somehow called outside the root session, return a model-visible error.\n  - Question tools must execute alone in a model tool round. If a round contains one question tool plus any other tool call, execute the question tool and return model-visible error results for the non-question peers, preserving tool-call/tool-result ordering. If a round contains multiple separate question-tool calls, execute only the first and return model-visible error results for the later question-tool calls instructing the model to combine questions into one `questions[]` batch.\n  - Agent-originated questions have no per-question timeout in v1 because the provider schemas do not include timeout. They rely on existing stage timeout, wall-clock timeout, cancellation, and interruption behavior.\n\n- Preserve existing answer paths:\n  - Do not add a new answer endpoint.\n  - Continue using `GET /runs/{id}/questions` and `POST /runs/{id}/questions/{qid}/answer`.\n  - Keep `ControlInterviewer`, web `InterviewDock`, Slack blocks, and run projection as the delivery mechanism.\n\n## Test Plan\n\n- Unit tests for schema parsing and normalization:\n  - Codex request with descriptions maps to Fabro multiple-choice questions and returns answers by model question ID.\n  - Claude request with `multiSelect: true` maps to `MultiSelect` and returns comma-separated answer text.\n  - Batched Codex and Claude requests surface all questions as pending before awaiting answers, then return one result with every answer mapped to the original model ID/text.\n  - Optional `preview` and `description` survive event, projection, API conversion, OpenAPI replacement tests, and TypeScript client generation.\n  - Oversized `description`/`preview` values are bounded before persistence and never rendered as trusted HTML.\n\n- Workflow and agent tests:\n  - OpenAI-profile session advertises `request_user_input`; Anthropic-profile session advertises `AskUserQuestion`; Gemini advertises neither.\n  - Subagent profiles do not advertise the question tools.\n  - Root agent can ask a question and resume after the answer.\n  - Subagent or missing interview context returns a clear tool error.\n  - Cached full-fidelity session emits interview events against the current stage, not the original cached stage.\n  - A mixed tool round containing a human-question tool plus another tool preserves all required tool results and rejects the peer calls with model-visible errors.\n  - A round with multiple separate question-tool calls executes only the first and rejects later question-tool calls with model-visible errors.\n\n- Server, projection, and UI tests:\n  - `InterviewStarted` with option metadata appears in pending questions.\n  - Submitting valid selected, multi-selected, and freeform answers unblocks the waiting tool.\n  - Duplicate answer submission remains rejected through existing accepted-question logic.\n  - Parallel human gate plus agent question keeps the run blocked until both are answered.\n  - Pause, cancel, and interrupt while an agent question is waiting resolve pending questions consistently and do not leave the run blocked.\n  - Stage timeout or wall-clock timeout while an agent question is waiting interrupts the batch; no per-question timeout event is expected unless a future schema adds timeout.\n  - Slack answer submissions work for agent-originated questions using the same pending interview transport.\n\n- Run checks:\n  - `cargo nextest run -p fabro-interview -p fabro-workflow -p fabro-server -p fabro-agent`\n  - `cd apps/fabro-web && bun test && bun run typecheck`\n  - Regenerate and verify OpenAPI-derived Rust and TypeScript clients after schema changes.\n\n## Assumptions\n\n- This feature is only for in-process/API-backed agent sessions, not ACP external agents in v1.\n- `preview` is stored and exposed but not rendered specially in the first implementation.\n- Human-question tools are available only during root agent execution inside a workflow run with an active interviewer.\n- Existing interview events remain the source of truth for pending questions; no separate agent-question event family is added.\n- The implementation should prefer extending existing interview structs and replacement mappings over adding parallel API DTOs or conversion-only aliases.\n",
+          "internal.work_dir": "/home/daytona/workspace/fabro",
+          "last_stage": "simplify_opus",
+          "graph.rankdir": "LR",
           "internal.retry_count.implement": 0,
-          "current_node": "implement",
+          "current_node": "simplify_opus",
           "internal.node_visit_count": 1,
           "internal.retry_count.toolchain": 0,
           "thread.preflight_compile.current_node": "preflight_lint",
-          "internal.retry_count.preflight_compile": 0,
-          "internal.retry_count.start": 0,
-          "internal.retry_count.preflight_lint": 0,
-          "internal.run_id": "01KSATQNAXG41FHKV0QH5N1QGC",
-          "thread.start.current_node": "toolchain",
+          "last_response": "All clean. Final summary:\n\n## Cleanup Applied\n\nAggregated findings from three reviews and applied 7 focused fixes (3 files, net −31 lines):\n\n**Efficiency**\n- `bounded_display_field` now truncates in",
+          "thread.implement.current_node": "simplify_opus",
           "graph.model_stylesheet": "\n            * { model: claude-opus-4-7; }\n        ",
-          "last_response": "Implemented the mid-stage agent interview tools and metadata plumbing.\n\nSummary of changes:\n- Extended canonical `InterviewOption` with optional `description` and `preview`.\n- Updated OpenAPI and rege",
           "outcome": "succeeded"
         },
         "node_outcomes": {
@@ -842,12 +969,50 @@
             },
             "notes": "Script completed: cargo +nightly-2026-04-14 clippy -q --workspace --all-targets -- -D warnings 2>&1",
             "usage": null
+          },
+          "simplify_opus": {
+            "status": "succeeded",
+            "context_updates": {
+              "last_response": "All clean. Final summary:\n\n## Cleanup Applied\n\nAggregated findings from three reviews and applied 7 focused fixes (3 files, net −31 lines):\n\n**Efficiency**\n- `bounded_display_field` now truncates in",
+              "response.simplify_opus": "All clean. Final summary:\n\n## Cleanup Applied\n\nAggregated findings from three reviews and applied 7 focused fixes (3 files, net −31 lines):\n\n**Efficiency**\n- `bounded_display_field` now truncates in a single `char_indices().nth(...)` pass instead of `chars().count()` + `chars().take()`.\n- Dropped redundant `Ulid::new()` from `internal_question_id` — slug + stage visit + tool-call-id + index is already unique.\n\n**Code quality**\n- De-duped the access-denial / hook-block / normal-execution event triple via shared `emit_tool_call_started` and `emit_tool_call_result` helpers in `tool_execution.rs`, dropping ~25 lines of copy-pasted emit code.\n- Cleaned up dead `unwrap_or_else` fallbacks in `format_anthropic_answers` (infallible string serialization).\n- Extracted repeated \"root agent session\" error string into `ROOT_SESSION_REQUIRED_ERROR` constant.\n- Renamed `option_label` to `label_for_key`, dropping the redundant `selected_option.filter(key)` defensive check. Preserved the `selected_option`-first lookup for `Selected` (control-protocol path supplies `Selected(key)` with `selected_option: None`).\n- Removed the unused `Default for RunInterviewBlocker` impl.\n- De-duped `RunInterviewGuard::resolve` and `Drop` via a shared `resolve_in_place` method.\n\n**Findings noted but not applied** (with rationale):\n- Lifting `agent_tool_runtime` from a parameter into a task-local set once at the top: large cross-crate refactor; skipped to keep scope bounded.\n- Unifying OpenAI/Anthropic question tools into one schema: schemas intentionally differ (`id`/`header` required for OpenAI, `multiSelect` and `preview` for Anthropic); separation matches provider contracts.\n- `AgentToolRuntime` one-field wrapper: keeping a struct allows future expansion at no extra friction.\n- `register_question_tools` moving into `AgentProfile` trait: cross-crate refactor; current explicit dispatch is fine for two profiles.\n- Cleaning `ControlInterviewer.pending` on agent-batch cancel: needs a new `Interviewer` API; flagged as the most substantial follow-up.\n- `..InterviewOption::default()` adoption in tests: high-churn, low-value compared to other items in this pass.\n- The `mem::forget` suggestion for `RunInterviewGuard::resolve` is buggy — it would leak the `Arc<RunInterviewBlocker>` and `Arc<Emitter>` refs. Used a shared method instead.\n\n**Verification**\n- `cargo nextest run -p fabro-agent -p fabro-workflow -p fabro-interview -p fabro-server`: all tests pass except 3 pre-existing flaky SVG-render tests (confirmed by stashing changes and reproducing the same failures on the pre-cleanup tree).\n- `cargo +nightly-2026-04-14 clippy --workspace --all-targets -- -D warnings`: clean.\n- `cargo +nightly-2026-04-14 fmt --check --all`: clean.",
+              "last_stage": "simplify_opus"
+            },
+            "notes": "Stage completed: simplify_opus",
+            "usage": {
+              "input": {
+                "usage": {
+                  "model": {
+                    "provider": "anthropic",
+                    "model_id": "claude-opus-4-7"
+                  },
+                  "tokens": {
+                    "input_tokens": 112522,
+                    "output_tokens": 28696,
+                    "reasoning_tokens": 0,
+                    "cache_read_tokens": 6453275,
+                    "cache_write_tokens": 664565
+                  }
+                },
+                "facts": {
+                  "algorithm": "anthropic",
+                  "cache_write_5m_tokens": 664565,
+                  "cache_write_1h_tokens": 0
+                }
+              },
+              "total_usd_micros": 8660178
+            },
+            "files_touched": [
+              "/home/daytona/workspace/fabro/lib/crates/fabro-agent/src/question_tools.rs",
+              "/home/daytona/workspace/fabro/lib/crates/fabro-agent/src/tool_execution.rs",
+              "/home/daytona/workspace/fabro/lib/crates/fabro-workflow/src/interview_runtime.rs"
+            ]
           }
         },
-        "next_node_id": "simplify_opus",
+        "next_node_id": "simplify_gpt",
         "node_visits": {
           "preflight_compile": 1,
           "implement": 1,
+          "simplify_opus": 1,
           "preflight_lint": 1,
           "start": 1,
           "toolchain": 1
@@ -877,6 +1042,20 @@
   "superseded_by": null,
   "pending_interviews": {},
   "todos_by_list": {
+    "anthropic_tasks:e2c3ef75-6a08-43e1-938a-e9df250e3568": {
+      "kind": "anthropic_tasks",
+      "list_id": "anthropic_tasks:e2c3ef75-6a08-43e1-938a-e9df250e3568",
+      "items": [
+        {
+          "id": "1",
+          "status": "completed",
+          "order": 0,
+          "subject": "Apply review fixes from three agents",
+          "description": "Apply concrete fixes identified by the reuse, quality, and efficiency reviews:\n1. Simplify RunInterviewGuard::resolve using mem::forget\n2. Single-pass bounded_display_field\n3. Remove dead unwrap_or_else in format_anthropic_answers\n4. Reuse access-denial emit path for error_tool_result_with_events\n5. Drop redundant defensive lookup in option_label\n6. Simplify AgentToolRuntime wrapper (one-field) to plain Option<Arc<dyn>>\n7. Simplify PendingAgentQuestionBatch\n8. Drop redundant Ulid suffix on internal_question_id\n9. Adopt ..InterviewOption::default() in modified test sites\n10. De-dup repeated \"root agent session\" error string\n11. Move agent metadata keys into constants",
+          "active_form": "Applying review fixes"
+        }
+      ]
+    },
     "openai_plan:2eb27b9b-fdd2-4013-8eb9-2cf31c483981": {
       "kind": "openai_plan",
       "list_id": "openai_plan:2eb27b9b-fdd2-4013-8eb9-2cf31c483981",
@@ -937,7 +1116,12 @@
       "first_event_seq": 50,
       "prompt": null,
       "response": null,
-      "completion": null,
+      "completion": {
+        "outcome": "succeeded",
+        "notes": "Stage completed: implement",
+        "failure_reason": null,
+        "timestamp": "2026-05-23T17:14:03.543746Z"
+      },
       "provider_used": {
         "mode": "agent",
         "provider": "openai",
@@ -950,6 +1134,12 @@
       "output": null,
       "started_at": "2026-05-23T16:34:35.522394Z",
       "handler": "agent",
+      "timing": {
+        "wall_time_ms": 2368006,
+        "inference_time_ms": 0,
+        "tool_time_ms": 0,
+        "active_time_ms": 0
+      },
       "usage": {
         "input_tokens": 483262,
         "output_tokens": 43089,
@@ -963,7 +1153,7 @@
         "provider": "openai",
         "model_id": "gpt-5.5"
       },
-      "state": "running"
+      "state": "succeeded"
     },
     "toolchain@1": {
       "first_event_seq": 20,
@@ -1047,6 +1237,38 @@
       },
       "state": "succeeded"
     },
+    "simplify_opus@1": {
+      "first_event_seq": 1027,
+      "prompt": null,
+      "response": null,
+      "completion": null,
+      "provider_used": {
+        "mode": "agent",
+        "provider": "anthropic",
+        "model": "claude-opus-4-7"
+      },
+      "diff": null,
+      "script_invocation": null,
+      "script_timing": null,
+      "parallel_results": null,
+      "output": null,
+      "started_at": "2026-05-23T17:14:07.821973Z",
+      "handler": "agent",
+      "usage": {
+        "input_tokens": 112763,
+        "output_tokens": 29805,
+        "total_tokens": 7384871,
+        "reasoning_tokens": 0,
+        "cache_read_tokens": 6577290,
+        "cache_write_tokens": 665013,
+        "total_usd_micros": 8660178
+      },
+      "model": {
+        "provider": "anthropic",
+        "model_id": "claude-opus-4-7"
+      },
+      "state": "running"
+    },
     "preflight_compile@1": {
       "first_event_seq": 30,
       "prompt": null,
diff --git a/stages/005-implement@1/diff.patch b/stages/005-implement@1/diff.patch
new file mode 100644
index 000000000..30ed37f1f
--- /dev/null
+++ b/stages/005-implement@1/diff.patch
@@ -0,0 +1,3518 @@
+diff --git a/apps/fabro-web/app/components/interview-dock.test.tsx b/apps/fabro-web/app/components/interview-dock.test.tsx
+index 34a58269f..6fb48bfb9 100644
+--- a/apps/fabro-web/app/components/interview-dock.test.tsx
++++ b/apps/fabro-web/app/components/interview-dock.test.tsx
+@@ -133,6 +133,27 @@ describe("InterviewDock", () => {
+     expect(buttons.Revise).toBeDefined();
+   });
+ 
++  test("multiple choice renders option descriptions as display text", () => {
++    const question = makeQuestion({
++      question_type: QuestionType.MULTIPLE_CHOICE,
++      options: [
++        {
++          key: "A",
++          label: "[A] Approve",
++          description: "Deploy the current patch",
++          preview: "<b>not rendered specially</b>",
++        },
++      ],
++    });
++    const tree = render(
++      <InterviewDock runId="run-1" questions={[question]} />,
++    );
++    const text = textContent(tree.toJSON());
++    expect(text).toContain("Approve");
++    expect(text).toContain("Deploy the current patch");
++    expect(text).not.toContain("<b>not rendered specially</b>");
++  });
++
+   test("freeform question renders a textarea and disables send when empty", () => {
+     const question = makeQuestion({
+       question_type: QuestionType.FREEFORM,
+diff --git a/apps/fabro-web/app/components/interview-dock.tsx b/apps/fabro-web/app/components/interview-dock.tsx
+index 35eb2fcf5..967044907 100644
+--- a/apps/fabro-web/app/components/interview-dock.tsx
++++ b/apps/fabro-web/app/components/interview-dock.tsx
+@@ -15,7 +15,7 @@ import {
+ import { QuestionType } from "@qltysh/fabro-api-client";
+ import type {
+   ApiQuestion,
+-  ApiQuestionOption,
++  InterviewOption,
+ } from "@qltysh/fabro-api-client";
+ 
+ import {
+@@ -270,7 +270,7 @@ function ChoiceBody({
+   submitting,
+   onSubmit,
+ }: {
+-  options: ApiQuestionOption[];
++  options: InterviewOption[];
+   allowFreeform: boolean;
+   submitting: boolean;
+   onSubmit: (answer: SubmitInterviewAnswer) => Promise<void>;
+@@ -287,7 +287,7 @@ function ChoiceBody({
+               onClick={() => void onSubmit({ kind: "selected", option_key: option.key })}
+               className={CHOICE_BUTTON}
+             >
+-              {displayLabel(option.label)}
++              <OptionLabel option={option} />
+             </button>
+           ))}
+         </div>
+@@ -314,7 +314,7 @@ function MultiSelectBody({
+   submitting,
+   onSubmit,
+ }: {
+-  options: ApiQuestionOption[];
++  options: InterviewOption[];
+   submitting: boolean;
+   onSubmit: (answer: SubmitInterviewAnswer) => Promise<void>;
+ }) {
+@@ -346,7 +346,7 @@ function MultiSelectBody({
+               className={isSelected ? CHOICE_BUTTON_SELECTED : CHOICE_BUTTON}
+             >
+               {isSelected && <CheckIcon className="size-3.5" aria-hidden="true" />}
+-              {displayLabel(option.label)}
++              <OptionLabel option={option} />
+             </button>
+           );
+         })}
+@@ -455,6 +455,19 @@ function FreeformBody({
+   );
+ }
+ 
++function OptionLabel({ option }: { option: InterviewOption }) {
++  return (
++    <span className="text-left">
++      <span className="block">{displayLabel(option.label)}</span>
++      {option.description && (
++        <span className="mt-0.5 block text-xs/5 font-normal text-fg-muted">
++          {option.description}
++        </span>
++      )}
++    </span>
++  );
++}
++
+ function Spinner() {
+   return <ArrowPathIcon className="size-4 animate-spin" aria-hidden="true" />;
+ }
+diff --git a/apps/fabro-web/app/components/stage-renderers/helpers.test.ts b/apps/fabro-web/app/components/stage-renderers/helpers.test.ts
+index 6c03541c0..8426ae141 100644
+--- a/apps/fabro-web/app/components/stage-renderers/helpers.test.ts
++++ b/apps/fabro-web/app/components/stage-renderers/helpers.test.ts
+@@ -79,6 +79,36 @@ describe("parseHumanInterviewPairs", () => {
+     expect(pairs[0].resolution).toBeNull();
+   });
+ 
++  test("preserves option description and preview metadata from started events", () => {
++    const events: EventEnvelope[] = [
++      envelope(1, {
++        event: "interview.started",
++        properties: {
++          question_id: "q-1",
++          question: "Pick a path",
++          question_type: "multiple_choice",
++          options: [
++            {
++              key: "ship",
++              label: "Ship",
++              description: "Deploy the current patch",
++              preview: "diff preview",
++            },
++          ],
++        },
++      }),
++    ];
++
++    const pairs = parseHumanInterviewPairs(events);
++
++    expect(pairs[0].question.options[0]).toEqual({
++      key: "ship",
++      label: "Ship",
++      description: "Deploy the current patch",
++      preview: "diff preview",
++    });
++  });
++
+   test("captures timeout and interrupted resolutions", () => {
+     const events: EventEnvelope[] = [
+       envelope(1, {
+diff --git a/apps/fabro-web/app/components/stage-renderers/helpers.ts b/apps/fabro-web/app/components/stage-renderers/helpers.ts
+index 561779a55..676e372ee 100644
+--- a/apps/fabro-web/app/components/stage-renderers/helpers.ts
++++ b/apps/fabro-web/app/components/stage-renderers/helpers.ts
+@@ -5,6 +5,8 @@ import { getArray, getNumber, getObject, getString, type UnknownRecord } from ".
+ export interface InterviewOption {
+   key: string;
+   label: string;
++  description?: string | null;
++  preview?: string | null;
+ }
+ 
+ export interface HumanQuestion {
+@@ -54,7 +56,14 @@ function parseInterviewOptions(value: unknown): InterviewOption[] {
+     const record = item as UnknownRecord;
+     const key = getString(record, "key");
+     const label = getString(record, "label");
+-    if (key && label) out.push({ key, label });
++    if (key && label) {
++      const option: InterviewOption = { key, label };
++      const description = getString(record, "description");
++      const preview = getString(record, "preview");
++      if (description !== null) option.description = description;
++      if (preview !== null) option.preview = preview;
++      out.push(option);
++    }
+   }
+   return out;
+ }
+diff --git a/apps/fabro-web/app/components/stage-renderers/human-qa.tsx b/apps/fabro-web/app/components/stage-renderers/human-qa.tsx
+index 50dc48cdd..37464ce12 100644
+--- a/apps/fabro-web/app/components/stage-renderers/human-qa.tsx
++++ b/apps/fabro-web/app/components/stage-renderers/human-qa.tsx
+@@ -184,7 +184,14 @@ function QuestionBlock({
+                 <span className="inline-flex size-5 shrink-0 items-center justify-center rounded bg-overlay-strong font-mono text-[11px] text-fg-2">
+                   {option.key}
+                 </span>
+-                <span className="text-sm text-fg-3">{option.label}</span>
++                <span className="min-w-0 text-sm text-fg-3">
++                  <span>{option.label}</span>
++                  {option.description && (
++                    <span className="mt-0.5 block text-xs/5 text-fg-muted">
++                      {option.description}
++                    </span>
++                  )}
++                </span>
+               </li>
+             ))}
+             {question.allowFreeform && (
+diff --git a/docs/public/api-reference/fabro-api.yaml b/docs/public/api-reference/fabro-api.yaml
+index 44bfe394c..933a7228e 100644
+--- a/docs/public/api-reference/fabro-api.yaml
++++ b/docs/public/api-reference/fabro-api.yaml
+@@ -6941,22 +6941,6 @@ components:
+         id:
+           type: string
+ 
+-    ApiQuestionOption:
+-      description: A selectable option for a multiple-choice or multi-select question.
+-      type: object
+-      required:
+-        - key
+-        - label
+-      properties:
+-        key:
+-          type: string
+-          description: Machine-readable option key used when submitting an answer.
+-          example: option_a
+-        label:
+-          type: string
+-          description: Human-readable label displayed to the user.
+-          example: Accept changes
+-
+     ApiQuestion:
+       description: A pending human-in-the-loop question generated by a workflow stage.
+       type: object
+@@ -6986,7 +6970,7 @@ components:
+           type: array
+           description: Available options for selection-based questions. Empty for freeform questions.
+           items:
+-            $ref: "#/components/schemas/ApiQuestionOption"
++            $ref: "#/components/schemas/InterviewOption"
+         allow_freeform:
+           type: boolean
+           description: Whether the user may provide freeform text in addition to selecting options.
+@@ -7758,8 +7742,16 @@ components:
+       properties:
+         key:
+           type: string
++          description: Machine-readable option key used when submitting an answer.
+         label:
+           type: string
++          description: Human-readable label displayed to the user.
++        description:
++          type: ["string", "null"]
++          description: Optional untrusted model-authored option description for display.
++        preview:
++          type: ["string", "null"]
++          description: Optional untrusted model-authored option preview captured for clients.
+ 
+     InterviewQuestionRecord:
+       description: Storage shape of an interview question recorded in the event log.
+@@ -11651,4 +11643,4 @@ components:
+         login:
+           type: string
+           description: User's login identifier (e.g. GitHub username).
+-          example: octocat
+\ No newline at end of file
++          example: octocat
+diff --git a/lib/crates/fabro-agent/src/lib.rs b/lib/crates/fabro-agent/src/lib.rs
+index 04a1c6fdf..010004694 100644
+--- a/lib/crates/fabro-agent/src/lib.rs
++++ b/lib/crates/fabro-agent/src/lib.rs
+@@ -15,6 +15,7 @@ pub mod loop_detection;
+ pub mod mcp_integration;
+ pub mod memory;
+ pub mod profiles;
++pub mod question_tools;
+ pub mod read_before_write_sandbox;
+ pub mod sandbox;
+ pub mod session;
+@@ -46,6 +47,11 @@ pub use local_sandbox::LocalSandbox;
+ pub use loop_detection::detect_loop;
+ pub use memory::{MemoryDocument, discover_memory};
+ pub use profiles::{AnthropicProfile, EnvContext, GeminiProfile, OpenAiProfile};
++pub use question_tools::{
++    ANTHROPIC_ASK_USER_QUESTION_TOOL, AgentQuestion, AgentQuestionAnswer,
++    AgentQuestionAnswerStatus, AgentQuestionRuntime, AgentToolRuntime,
++    OPENAI_REQUEST_USER_INPUT_TOOL, register_question_tools,
++};
+ pub use read_before_write_sandbox::ReadBeforeWriteSandbox;
+ pub use sandbox::{
+     CommandOutputCallback, DirEntry, ExecResult, ExecStreamingResult, GrepOptions, Sandbox,
+diff --git a/lib/crates/fabro-agent/src/question_tools.rs b/lib/crates/fabro-agent/src/question_tools.rs
+new file mode 100644
+index 000000000..fd9c7deda
+--- /dev/null
++++ b/lib/crates/fabro-agent/src/question_tools.rs
+@@ -0,0 +1,586 @@
++//! Model-native tools that let a root workflow agent ask the human for input.
++
++use std::collections::BTreeMap;
++use std::future::Future;
++use std::sync::Arc;
++
++use async_trait::async_trait;
++use fabro_llm::types::ToolDefinition;
++use fabro_model::AgentProfileKind;
++use fabro_types::{InterviewOption, QuestionType};
++use serde::Deserialize;
++use serde_json::json;
++use tokio_util::sync::CancellationToken;
++
++use crate::tool_registry::{RegisteredTool, ToolContext, ToolRegistry};
++
++tokio::task_local! {
++    static CURRENT_AGENT_TOOL_RUNTIME: AgentToolRuntime;
++}
++
++pub const OPENAI_REQUEST_USER_INPUT_TOOL: &str = "request_user_input";
++pub const ANTHROPIC_ASK_USER_QUESTION_TOOL: &str = "AskUserQuestion";
++
++pub const OPTION_DESCRIPTION_MAX_CHARS: usize = 2_000;
++pub const OPTION_PREVIEW_MAX_CHARS: usize = 4_000;
++
++#[derive(Clone, Default)]
++pub struct AgentToolRuntime {
++    question_runtime: Option<Arc<dyn AgentQuestionRuntime>>,
++}
++
++impl AgentToolRuntime {
++    #[must_use]
++    pub fn new() -> Self {
++        Self::default()
++    }
++
++    #[must_use]
++    pub fn with_question_runtime(runtime: Arc<dyn AgentQuestionRuntime>) -> Self {
++        Self {
++            question_runtime: Some(runtime),
++        }
++    }
++
++    #[must_use]
++    pub fn question_runtime(&self) -> Option<Arc<dyn AgentQuestionRuntime>> {
++        self.question_runtime.clone()
++    }
++}
++
++pub async fn scope_agent_tool_runtime<F>(runtime: AgentToolRuntime, future: F) -> F::Output
++where
++    F: Future,
++{
++    CURRENT_AGENT_TOOL_RUNTIME.scope(runtime, future).await
++}
++
++fn current_agent_tool_runtime() -> AgentToolRuntime {
++    CURRENT_AGENT_TOOL_RUNTIME
++        .try_with(Clone::clone)
++        .unwrap_or_default()
++}
++
++#[derive(Debug, Clone, PartialEq, Eq)]
++pub struct AgentQuestion {
++    pub original_id:       Option<String>,
++    pub original_question: String,
++    pub header:            Option<String>,
++    pub text:              String,
++    pub question_type:     QuestionType,
++    pub options:           Vec<InterviewOption>,
++    pub allow_freeform:    bool,
++}
++
++#[derive(Debug, Clone, Copy, PartialEq, Eq)]
++pub enum AgentQuestionAnswerStatus {
++    Answered,
++    Cancelled,
++    Interrupted,
++    Skipped,
++    Timeout,
++}
++
++#[derive(Debug, Clone, PartialEq, Eq)]
++pub struct AgentQuestionAnswer {
++    pub original_id:       Option<String>,
++    pub original_question: String,
++    pub answers:           Vec<String>,
++    pub status:            AgentQuestionAnswerStatus,
++}
++
++#[async_trait]
++pub trait AgentQuestionRuntime: Send + Sync {
++    async fn ask_questions(
++        &self,
++        tool_call_id: &str,
++        questions: Vec<AgentQuestion>,
++        cancel_token: CancellationToken,
++    ) -> Result<Vec<AgentQuestionAnswer>, String>;
++}
++
++#[derive(Debug, Deserialize)]
++struct OpenAiQuestionToolArgs {
++    questions: Vec<OpenAiQuestion>,
++}
++
++#[derive(Debug, Deserialize)]
++struct OpenAiQuestion {
++    id:       String,
++    header:   String,
++    question: String,
++    #[serde(default)]
++    options:  Vec<OpenAiOption>,
++}
++
++#[derive(Debug, Deserialize)]
++struct OpenAiOption {
++    label:       String,
++    #[serde(default)]
++    description: Option<String>,
++}
++
++#[derive(Debug, Deserialize)]
++struct AnthropicQuestionToolArgs {
++    questions: Vec<AnthropicQuestion>,
++}
++
++#[derive(Debug, Deserialize)]
++#[serde(rename_all = "camelCase")]
++struct AnthropicQuestion {
++    question:     String,
++    #[serde(default)]
++    header:       Option<String>,
++    #[serde(default)]
++    options:      Vec<AnthropicOption>,
++    #[serde(default)]
++    multi_select: bool,
++}
++
++#[derive(Debug, Deserialize)]
++struct AnthropicOption {
++    label:       String,
++    #[serde(default)]
++    description: Option<String>,
++    #[serde(default)]
++    preview:     Option<String>,
++}
++
++#[must_use]
++pub fn is_question_tool(name: &str) -> bool {
++    matches!(
++        name,
++        OPENAI_REQUEST_USER_INPUT_TOOL | ANTHROPIC_ASK_USER_QUESTION_TOOL
++    )
++}
++
++pub fn register_question_tools(profile_kind: AgentProfileKind, registry: &mut ToolRegistry) {
++    match profile_kind {
++        AgentProfileKind::OpenAi => registry.register(make_openai_question_tool()),
++        AgentProfileKind::Anthropic => registry.register(make_anthropic_question_tool()),
++        AgentProfileKind::Gemini => {}
++    }
++}
++
++fn make_openai_question_tool() -> RegisteredTool {
++    RegisteredTool {
++        definition: ToolDefinition {
++            name:        OPENAI_REQUEST_USER_INPUT_TOOL.to_string(),
++            description: "Ask the human one or more questions and wait for their answers before continuing this stage.".to_string(),
++            parameters:  json!({
++                "type": "object",
++                "required": ["questions"],
++                "properties": {
++                    "questions": {
++                        "type": "array",
++                        "minItems": 1,
++                        "items": {
++                            "type": "object",
++                            "required": ["id", "header", "question", "options"],
++                            "properties": {
++                                "id": { "type": "string" },
++                                "header": { "type": "string" },
++                                "question": { "type": "string" },
++                                "options": {
++                                    "type": "array",
++                                    "items": {
++                                        "type": "object",
++                                        "required": ["label"],
++                                        "properties": {
++                                            "label": { "type": "string" },
++                                            "description": { "type": "string" }
++                                        }
++                                    }
++                                }
++                            }
++                        }
++                    }
++                }
++            }),
++        },
++        executor:   Arc::new(|args, ctx| {
++            Box::pin(async move {
++                let parsed: OpenAiQuestionToolArgs = parse_tool_args(args)?;
++                let questions = normalize_openai_questions(parsed)?;
++                let answers = execute_question_tool(ctx, questions).await?;
++                format_openai_answers(&answers)
++            })
++        }),
++    }
++}
++
++fn make_anthropic_question_tool() -> RegisteredTool {
++    RegisteredTool {
++        definition: ToolDefinition {
++            name:        ANTHROPIC_ASK_USER_QUESTION_TOOL.to_string(),
++            description: "Ask the human one or more questions and wait for their answers before continuing this stage.".to_string(),
++            parameters:  json!({
++                "type": "object",
++                "required": ["questions"],
++                "properties": {
++                    "questions": {
++                        "type": "array",
++                        "minItems": 1,
++                        "items": {
++                            "type": "object",
++                            "required": ["question", "options", "multiSelect"],
++                            "properties": {
++                                "question": { "type": "string" },
++                                "header": { "type": "string" },
++                                "options": {
++                                    "type": "array",
++                                    "items": {
++                                        "type": "object",
++                                        "required": ["label"],
++                                        "properties": {
++                                            "label": { "type": "string" },
++                                            "description": { "type": "string" },
++                                            "preview": { "type": "string" }
++                                        }
++                                    }
++                                },
++                                "multiSelect": { "type": "boolean" }
++                            }
++                        }
++                    }
++                }
++            }),
++        },
++        executor:   Arc::new(|args, ctx| {
++            Box::pin(async move {
++                let parsed: AnthropicQuestionToolArgs = parse_tool_args(args)?;
++                let questions = normalize_anthropic_questions(parsed)?;
++                let answers = execute_question_tool(ctx, questions).await?;
++                format_anthropic_answers(&answers)
++            })
++        }),
++    }
++}
++
++fn parse_tool_args<T: for<'de> Deserialize<'de>>(args: serde_json::Value) -> Result<T, String> {
++    serde_json::from_value(args).map_err(|err| format!("invalid question tool arguments: {err}"))
++}
++
++async fn execute_question_tool(
++    ctx: ToolContext,
++    questions: Vec<AgentQuestion>,
++) -> Result<Vec<AgentQuestionAnswer>, String> {
++    let session_id = ctx.session_id.as_deref().ok_or_else(|| {
++        "human-question tools are available only during a root workflow agent session".to_string()
++    })?;
++    let root_session_id = ctx.root_session_id.as_deref().ok_or_else(|| {
++        "human-question tools are available only during a root workflow agent session".to_string()
++    })?;
++    if session_id != root_session_id {
++        return Err(
++            "human-question tools are only available to the root agent; subagents must report back to their parent".to_string(),
++        );
++    }
++    let tool_call_id = ctx
++        .tool_call_id
++        .as_deref()
++        .ok_or_else(|| "human-question tool call is missing a provider tool_call_id".to_string())?;
++    let runtime = current_agent_tool_runtime().question_runtime().ok_or_else(|| {
++        "human-question tools are available only inside a workflow run with an active interviewer".to_string()
++    })?;
++    runtime
++        .ask_questions(tool_call_id, questions, ctx.cancel.clone())
++        .await
++}
++
++fn normalize_openai_questions(args: OpenAiQuestionToolArgs) -> Result<Vec<AgentQuestion>, String> {
++    if args.questions.is_empty() {
++        return Err("questions must contain at least one question".to_string());
++    }
++    args.questions
++        .into_iter()
++        .map(|question| {
++            let original_question = question.question.trim().to_string();
++            Ok(AgentQuestion {
++                original_id: Some(non_empty(&question.id, "question id")?),
++                text: display_text(Some(question.header.as_str()), &question.question),
++                header: Some(question.header),
++                original_question,
++                question_type: QuestionType::MultipleChoice,
++                options: options_from_openai(question.options),
++                allow_freeform: true,
++            })
++        })
++        .collect()
++}
++
++fn normalize_anthropic_questions(
++    args: AnthropicQuestionToolArgs,
++) -> Result<Vec<AgentQuestion>, String> {
++    if args.questions.is_empty() {
++        return Err("questions must contain at least one question".to_string());
++    }
++    args.questions
++        .into_iter()
++        .map(|question| {
++            let original_question = non_empty(&question.question, "question")?;
++            Ok(AgentQuestion {
++                original_id: None,
++                text: display_text(question.header.as_deref(), &question.question),
++                header: question.header,
++                original_question,
++                question_type: if question.multi_select {
++                    QuestionType::MultiSelect
++                } else {
++                    QuestionType::MultipleChoice
++                },
++                options: options_from_anthropic(question.options),
++                allow_freeform: true,
++            })
++        })
++        .collect()
++}
++
++fn options_from_openai(options: Vec<OpenAiOption>) -> Vec<InterviewOption> {
++    options
++        .into_iter()
++        .enumerate()
++        .map(|(idx, option)| InterviewOption {
++            key:         option_key(idx),
++            label:       option.label,
++            description: option
++                .description
++                .map(|value| bounded_display_field(&value, OPTION_DESCRIPTION_MAX_CHARS)),
++            preview:     None,
++        })
++        .collect()
++}
++
++fn options_from_anthropic(options: Vec<AnthropicOption>) -> Vec<InterviewOption> {
++    options
++        .into_iter()
++        .enumerate()
++        .map(|(idx, option)| InterviewOption {
++            key:         option_key(idx),
++            label:       option.label,
++            description: option
++                .description
++                .map(|value| bounded_display_field(&value, OPTION_DESCRIPTION_MAX_CHARS)),
++            preview:     option
++                .preview
++                .map(|value| bounded_display_field(&value, OPTION_PREVIEW_MAX_CHARS)),
++        })
++        .collect()
++}
++
++fn option_key(idx: usize) -> String {
++    format!("option_{}", idx + 1)
++}
++
++fn non_empty(value: &str, field: &str) -> Result<String, String> {
++    let trimmed = value.trim();
++    if trimmed.is_empty() {
++        Err(format!("{field} must not be empty"))
++    } else {
++        Ok(trimmed.to_string())
++    }
++}
++
++fn display_text(header: Option<&str>, question: &str) -> String {
++    let header = header.map(str::trim).filter(|value| !value.is_empty());
++    let question = question.trim();
++    match (header, question.is_empty()) {
++        (Some(header), false) => format!("{header}\n\n{question}"),
++        (Some(header), true) => header.to_string(),
++        (None, false) => question.to_string(),
++        (None, true) => String::new(),
++    }
++}
++
++#[must_use]
++pub fn bounded_display_field(value: &str, max_chars: usize) -> String {
++    if value.chars().count() <= max_chars {
++        return value.to_string();
++    }
++    value.chars().take(max_chars).collect()
++}
++
++fn ensure_all_answered(answers: &[AgentQuestionAnswer]) -> Result<(), String> {
++    if let Some(answer) = answers
++        .iter()
++        .find(|answer| answer.status != AgentQuestionAnswerStatus::Answered)
++    {
++        return Err(format!(
++            "human-question request ended before the user answered `{}`: {}",
++            answer.original_question,
++            answer_status_label(answer.status)
++        ));
++    }
++    Ok(())
++}
++
++fn answer_status_label(status: AgentQuestionAnswerStatus) -> &'static str {
++    match status {
++        AgentQuestionAnswerStatus::Answered => "answered",
++        AgentQuestionAnswerStatus::Cancelled => "cancelled",
++        AgentQuestionAnswerStatus::Interrupted => "interrupted",
++        AgentQuestionAnswerStatus::Skipped => "skipped",
++        AgentQuestionAnswerStatus::Timeout => "timed out",
++    }
++}
++
++fn format_openai_answers(answers: &[AgentQuestionAnswer]) -> Result<String, String> {
++    ensure_all_answered(answers)?;
++    let mut answer_map = BTreeMap::new();
++    for answer in answers {
++        let Some(original_id) = answer.original_id.as_ref() else {
++            return Err(
++                "OpenAI question answer is missing the original model question id".to_string(),
++            );
++        };
++        answer_map.insert(original_id.clone(), json!({ "answers": answer.answers }));
++    }
++    serde_json::to_string(&json!({ "answers": answer_map }))
++        .map_err(|err| format!("failed to serialize answers: {err}"))
++}
++
++fn format_anthropic_answers(answers: &[AgentQuestionAnswer]) -> Result<String, String> {
++    ensure_all_answered(answers)?;
++    let pairs = answers
++        .iter()
++        .map(|answer| {
++            let question = serde_json::to_string(&answer.original_question)
++                .unwrap_or_else(|_| "\"question\"".to_string());
++            let answer_text = serde_json::to_string(&answer.answers.join(", "))
++                .unwrap_or_else(|_| "\"\"".to_string());
++            format!("{question}={answer_text}")
++        })
++        .collect::<Vec<_>>()
++        .join(", ");
++    Ok(format!(
++        "User has answered your questions: {pairs}. You can now continue with the task."
++    ))
++}
++
++#[cfg(test)]
++mod tests {
++    use super::*;
++
++    fn answered(
++        original_id: Option<&str>,
++        question: &str,
++        answers: &[&str],
++    ) -> AgentQuestionAnswer {
++        AgentQuestionAnswer {
++            original_id:       original_id.map(str::to_string),
++            original_question: question.to_string(),
++            answers:           answers.iter().map(|value| (*value).to_string()).collect(),
++            status:            AgentQuestionAnswerStatus::Answered,
++        }
++    }
++
++    #[test]
++    fn openai_request_with_descriptions_normalizes_to_multiple_choice() {
++        let args: OpenAiQuestionToolArgs = serde_json::from_value(json!({
++            "questions": [{
++                "id": "q1",
++                "header": "Decision",
++                "question": "Which path?",
++                "options": [{ "label": "Ship", "description": "Deploy now" }]
++            }]
++        }))
++        .unwrap();
++
++        let questions = normalize_openai_questions(args).unwrap();
++
++        assert_eq!(questions.len(), 1);
++        assert_eq!(questions[0].original_id.as_deref(), Some("q1"));
++        assert_eq!(questions[0].question_type, QuestionType::MultipleChoice);
++        assert!(questions[0].allow_freeform);
++        assert_eq!(questions[0].text, "Decision\n\nWhich path?");
++        assert_eq!(questions[0].options[0].key, "option_1");
++        assert_eq!(questions[0].options[0].label, "Ship");
++        assert_eq!(
++            questions[0].options[0].description.as_deref(),
++            Some("Deploy now")
++        );
++    }
++
++    #[test]
++    fn anthropic_multiselect_preserves_preview_and_formats_comma_joined_answers() {
++        let args: AnthropicQuestionToolArgs = serde_json::from_value(json!({
++            "questions": [{
++                "header": "Pick features",
++                "question": "Which features?",
++                "multiSelect": true,
++                "options": [{
++                    "label": "Auth",
++                    "description": "Login support",
++                    "preview": "auth diff"
++                }]
++            }]
++        }))
++        .unwrap();
++
++        let questions = normalize_anthropic_questions(args).unwrap();
++
++        assert_eq!(questions[0].question_type, QuestionType::MultiSelect);
++        assert_eq!(
++            questions[0].options[0].preview.as_deref(),
++            Some("auth diff")
++        );
++        let text =
++            format_anthropic_answers(&[answered(None, "Which features?", &["Auth", "Billing"])])
++                .unwrap();
++        assert!(text.contains("\"Which features?\"=\"Auth, Billing\""));
++    }
++
++    #[test]
++    fn openai_answers_are_keyed_by_original_model_question_id() {
++        let text = format_openai_answers(&[
++            answered(Some("first"), "First?", &["Yes"]),
++            answered(Some("second"), "Second?", &["No"]),
++        ])
++        .unwrap();
++
++        assert_eq!(
++            serde_json::from_str::<serde_json::Value>(&text).unwrap(),
++            json!({
++                "answers": {
++                    "first": { "answers": ["Yes"] },
++                    "second": { "answers": ["No"] }
++                }
++            })
++        );
++    }
++
++    #[test]
++    fn option_description_and_preview_are_bounded() {
++        let long = "x".repeat(OPTION_PREVIEW_MAX_CHARS + 10);
++
++        assert_eq!(
++            bounded_display_field(&long, OPTION_DESCRIPTION_MAX_CHARS)
++                .chars()
++                .count(),
++            OPTION_DESCRIPTION_MAX_CHARS
++        );
++        assert_eq!(
++            bounded_display_field(&long, OPTION_PREVIEW_MAX_CHARS)
++                .chars()
++                .count(),
++            OPTION_PREVIEW_MAX_CHARS
++        );
++    }
++
++    #[test]
++    fn question_tool_registration_is_profile_specific() {
++        let mut openai = ToolRegistry::new();
++        register_question_tools(AgentProfileKind::OpenAi, &mut openai);
++        assert!(openai.get(OPENAI_REQUEST_USER_INPUT_TOOL).is_some());
++        assert!(openai.get(ANTHROPIC_ASK_USER_QUESTION_TOOL).is_none());
++
++        let mut anthropic = ToolRegistry::new();
++        register_question_tools(AgentProfileKind::Anthropic, &mut anthropic);
++        assert!(anthropic.get(ANTHROPIC_ASK_USER_QUESTION_TOOL).is_some());
++        assert!(anthropic.get(OPENAI_REQUEST_USER_INPUT_TOOL).is_none());
++
++        let mut gemini = ToolRegistry::new();
++        register_question_tools(AgentProfileKind::Gemini, &mut gemini);
++        assert!(gemini.names().is_empty());
++    }
++}
+diff --git a/lib/crates/fabro-agent/src/session.rs b/lib/crates/fabro-agent/src/session.rs
+index 2d0d0b4e8..0d4fd749f 100644
+--- a/lib/crates/fabro-agent/src/session.rs
++++ b/lib/crates/fabro-agent/src/session.rs
+@@ -32,6 +32,7 @@ use crate::history::History;
+ use crate::loop_detection::detect_loop;
+ use crate::memory::{BUDGET_BYTES, MemoryDocument, discover_memory};
+ use crate::profiles::EnvContext;
++use crate::question_tools::AgentToolRuntime;
+ use crate::sandbox::Sandbox;
+ use crate::skills::{
+     ExpandedInput, Skill, default_skill_dirs, discover_skills, expand_skill, make_use_skill_tool,
+@@ -1119,6 +1120,15 @@ impl Session {
+     }
+ 
+     pub async fn process_input(&mut self, input: &str) -> Result<(), Error> {
++        self.process_input_with_runtime(input, AgentToolRuntime::default())
++            .await
++    }
++
++    pub async fn process_input_with_runtime(
++        &mut self,
++        input: &str,
++        agent_tool_runtime: AgentToolRuntime,
++    ) -> Result<(), Error> {
+         if self.state == SessionState::Closed {
+             return Err(Error::SessionClosed);
+         }
+@@ -1142,7 +1152,7 @@ impl Session {
+         });
+ 
+         // Process the initial input, then drain any followups
+-        let mut result = self.run_single_input(input).await;
++        let mut result = self.run_single_input(input, &agent_tool_runtime).await;
+ 
+         if result.is_ok() {
+             loop {
+@@ -1152,7 +1162,7 @@ impl Session {
+                     .expect("followup queue lock poisoned")
+                     .pop_front();
+                 let Some(followup) = followup else { break };
+-                result = self.run_single_input(&followup).await;
++                result = self.run_single_input(&followup, &agent_tool_runtime).await;
+                 if result.is_err() {
+                     break;
+                 }
+@@ -1172,7 +1182,11 @@ impl Session {
+         result
+     }
+ 
+-    async fn run_single_input(&mut self, input: &str) -> Result<(), Error> {
++    async fn run_single_input(
++        &mut self,
++        input: &str,
++        agent_tool_runtime: &AgentToolRuntime,
++    ) -> Result<(), Error> {
+         const STREAM_CONSUME_RETRIES: usize = 3;
+ 
+         if self.state == SessionState::Closed {
+@@ -1623,6 +1637,7 @@ impl Session {
+                 &self.id,
+                 &self.root_session_id,
+                 self.tool_env_provider.as_ref(),
++                agent_tool_runtime,
+             )
+             .await;
+             composite_watcher.abort();
+diff --git a/lib/crates/fabro-agent/src/tool_execution.rs b/lib/crates/fabro-agent/src/tool_execution.rs
+index 939e6a4e9..3015b6147 100644
+--- a/lib/crates/fabro-agent/src/tool_execution.rs
++++ b/lib/crates/fabro-agent/src/tool_execution.rs
+@@ -7,6 +7,7 @@ use tracing::debug;
+ 
+ use crate::config::{SessionOptions, ToolHookCallback, ToolHookDecision};
+ use crate::event::{Emitter, SessionBoundEmitter};
++use crate::question_tools::{self, AgentToolRuntime, is_question_tool};
+ use crate::sandbox::Sandbox;
+ use crate::session::ToolEnvProvider;
+ use crate::tool_registry::{AgentEventEmitter, RegisteredTool, ToolContext, ToolRegistry};
+@@ -31,7 +32,25 @@ pub async fn execute_tool_calls(
+     session_id: &str,
+     root_session_id: &str,
+     tool_env_provider: Option<&Arc<dyn ToolEnvProvider>>,
++    agent_tool_runtime: &AgentToolRuntime,
+ ) -> Vec<ToolResult> {
++    if tool_calls.iter().any(|tc| is_question_tool(&tc.name)) {
++        return execute_question_tool_round(
++            tool_calls,
++            registry,
++            env,
++            tool_hooks,
++            cancel_token,
++            config,
++            emitter,
++            session_id,
++            root_session_id,
++            tool_env_provider,
++            agent_tool_runtime,
++        )
++        .await;
++    }
++
+     if parallel && tool_calls.len() > 1 {
+         execute_tool_calls_parallel(
+             tool_calls,
+@@ -44,6 +63,7 @@ pub async fn execute_tool_calls(
+             session_id,
+             root_session_id,
+             tool_env_provider,
++            agent_tool_runtime,
+         )
+         .await
+     } else {
+@@ -58,6 +78,7 @@ pub async fn execute_tool_calls(
+             session_id,
+             root_session_id,
+             tool_env_provider,
++            agent_tool_runtime,
+         )
+         .await
+     }
+@@ -78,6 +99,7 @@ async fn execute_tool_calls_sequential(
+     session_id: &str,
+     root_session_id: &str,
+     tool_env_provider: Option<&Arc<dyn ToolEnvProvider>>,
++    agent_tool_runtime: &AgentToolRuntime,
+ ) -> Vec<ToolResult> {
+     let mut results = Vec::new();
+     for tc in tool_calls {
+@@ -86,7 +108,7 @@ async fn execute_tool_calls_sequential(
+             continue;
+         }
+ 
+-        let result = execute_and_emit_one_tool(
++        let result = execute_and_emit_one_tool_with_runtime(
+             tc,
+             registry,
+             env.clone(),
+@@ -97,6 +119,7 @@ async fn execute_tool_calls_sequential(
+             session_id,
+             root_session_id,
+             tool_env_provider,
++            agent_tool_runtime,
+         )
+         .await;
+         results.push(result);
+@@ -119,8 +142,10 @@ async fn execute_tool_calls_parallel(
+     session_id: &str,
+     root_session_id: &str,
+     tool_env_provider: Option<&Arc<dyn ToolEnvProvider>>,
++    agent_tool_runtime: &AgentToolRuntime,
+ ) -> Vec<ToolResult> {
+     let tool_env_provider = tool_env_provider.cloned();
++    let agent_tool_runtime = agent_tool_runtime.clone();
+     let futures: Vec<_> = tool_calls
+         .iter()
+         .map(|tc| {
+@@ -133,6 +158,7 @@ async fn execute_tool_calls_parallel(
+             let root_session_id = root_session_id.to_owned();
+             let tool_hooks = tool_hooks.cloned();
+             let tool_env_provider = tool_env_provider.clone();
++            let agent_tool_runtime = agent_tool_runtime.clone();
+             let access_denial = config.tool_access_denial_reason(&tc.name);
+             // Look up the tool before spawning since ToolRegistry is not Send.
+             let registered_tool = if access_denial.is_none() {
+@@ -153,6 +179,7 @@ async fn execute_tool_calls_parallel(
+                     &session_id,
+                     &root_session_id,
+                     tool_env_provider.as_ref(),
++                    &agent_tool_runtime,
+                 )
+                 .await
+             }
+@@ -162,6 +189,99 @@ async fn execute_tool_calls_parallel(
+     future::join_all(futures).await
+ }
+ 
++#[allow(
++    clippy::too_many_arguments,
++    reason = "Question-tool round handling needs the same execution context as normal tool dispatch."
++)]
++async fn execute_question_tool_round(
++    tool_calls: &[ToolCall],
++    registry: &ToolRegistry,
++    env: Arc<dyn Sandbox>,
++    tool_hooks: Option<&Arc<dyn ToolHookCallback>>,
++    cancel_token: &CancellationToken,
++    config: &SessionOptions,
++    emitter: &Emitter,
++    session_id: &str,
++    root_session_id: &str,
++    tool_env_provider: Option<&Arc<dyn ToolEnvProvider>>,
++    agent_tool_runtime: &AgentToolRuntime,
++) -> Vec<ToolResult> {
++    let first_question_index = tool_calls
++        .iter()
++        .position(|tc| is_question_tool(&tc.name))
++        .expect("question-tool round should contain a question tool");
++    let mut results = Vec::with_capacity(tool_calls.len());
++
++    for (index, tc) in tool_calls.iter().enumerate() {
++        if cancel_token.is_cancelled() {
++            results.push(ToolResult::error(tc.id.clone(), "Cancelled"));
++            continue;
++        }
++
++        if index == first_question_index {
++            results.push(
++                execute_and_emit_one_tool_with_runtime(
++                    tc,
++                    registry,
++                    env.clone(),
++                    tool_hooks,
++                    cancel_token.child_token(),
++                    config,
++                    emitter,
++                    session_id,
++                    root_session_id,
++                    tool_env_provider,
++                    agent_tool_runtime,
++                )
++                .await,
++            );
++        } else if is_question_tool(&tc.name) {
++            results.push(error_tool_result_with_events(
++                tc,
++                emitter,
++                session_id,
++                config,
++                "Only one human-question tool call may be used in a tool round. Combine all questions into a single questions[] batch and call the question tool once.",
++            ));
++        } else {
++            results.push(error_tool_result_with_events(
++                tc,
++                emitter,
++                session_id,
++                config,
++                "This tool call was not executed because human-question tools must run alone in a tool round. Retry non-question tools in a later round after the user answers.",
++            ));
++        }
++    }
++
++    results
++}
++
++fn error_tool_result_with_events(
++    tc: &ToolCall,
++    emitter: &Emitter,
++    session_id: &str,
++    config: &SessionOptions,
++    message: &str,
++) -> ToolResult {
++    emitter.emit(session_id.to_owned(), AgentEvent::ToolCallStarted {
++        tool_name:    tc.name.clone(),
++        tool_call_id: tc.id.clone(),
++        arguments:    tc.arguments.clone(),
++    });
++    let result = ToolResult::error(&tc.id, message);
++    emitter.emit(session_id.to_owned(), AgentEvent::ToolCallOutputDelta {
++        delta: result.content.to_string(),
++    });
++    emitter.emit(session_id.to_owned(), AgentEvent::ToolCallCompleted {
++        tool_name:    tc.name.clone(),
++        tool_call_id: tc.id.clone(),
++        output:       result.content.clone(),
++        is_error:     true,
++    });
++    truncate_tool_result(&result, &tc.name, config)
++}
++
+ /// Execute a single tool call with event emission and output truncation.
+ #[allow(
+     clippy::too_many_arguments,
+@@ -178,6 +298,39 @@ pub async fn execute_and_emit_one_tool(
+     session_id: &str,
+     root_session_id: &str,
+     tool_env_provider: Option<&Arc<dyn ToolEnvProvider>>,
++) -> ToolResult {
++    execute_and_emit_one_tool_with_runtime(
++        tc,
++        registry,
++        env,
++        tool_hooks,
++        cancel_token,
++        config,
++        emitter,
++        session_id,
++        root_session_id,
++        tool_env_provider,
++        &AgentToolRuntime::default(),
++    )
++    .await
++}
++
++#[allow(
++    clippy::too_many_arguments,
++    reason = "Single-tool execution needs the tool, runtime handles, and emission context."
++)]
++async fn execute_and_emit_one_tool_with_runtime(
++    tc: &ToolCall,
++    registry: &ToolRegistry,
++    env: Arc<dyn Sandbox>,
++    tool_hooks: Option<&Arc<dyn ToolHookCallback>>,
++    cancel_token: CancellationToken,
++    config: &SessionOptions,
++    emitter: &Emitter,
++    session_id: &str,
++    root_session_id: &str,
++    tool_env_provider: Option<&Arc<dyn ToolEnvProvider>>,
++    agent_tool_runtime: &AgentToolRuntime,
+ ) -> ToolResult {
+     let access_denial = config.tool_access_denial_reason(&tc.name);
+     let registered_tool = if access_denial.is_none() {
+@@ -197,6 +350,7 @@ pub async fn execute_and_emit_one_tool(
+         session_id,
+         root_session_id,
+         tool_env_provider,
++        agent_tool_runtime,
+     )
+     .await
+ }
+@@ -219,6 +373,7 @@ async fn execute_and_emit_one_tool_with_lookup(
+     session_id: &str,
+     root_session_id: &str,
+     tool_env_provider: Option<&Arc<dyn ToolEnvProvider>>,
++    agent_tool_runtime: &AgentToolRuntime,
+ ) -> ToolResult {
+     emitter.emit(session_id.to_owned(), AgentEvent::ToolCallStarted {
+         tool_name:    tc.name.clone(),
+@@ -276,6 +431,7 @@ async fn execute_and_emit_one_tool_with_lookup(
+         session_id,
+         root_session_id,
+         tool_env_provider,
++        agent_tool_runtime,
+     )
+     .await;
+ 
+@@ -329,6 +485,7 @@ async fn execute_one_tool(
+     session_id: &str,
+     root_session_id: &str,
+     tool_env_provider: Option<&Arc<dyn ToolEnvProvider>>,
++    agent_tool_runtime: &AgentToolRuntime,
+ ) -> ToolResult {
+     match registered_tool {
+         Some(tool) => {
+@@ -355,7 +512,10 @@ async fn execute_one_tool(
+                 tool_call_id: Some(tc.id.clone()),
+                 agent_event_emitter,
+             };
+-            match (tool.executor)(tc.arguments.clone(), ctx).await {
++            let execution = (tool.executor)(tc.arguments.clone(), ctx);
++            match question_tools::scope_agent_tool_runtime(agent_tool_runtime.clone(), execution)
++                .await
++            {
+                 Ok(output) => ToolResult::success(&tc.id, serde_json::json!(output)),
+                 Err(err) => ToolResult::error(&tc.id, err),
+             }
+@@ -418,9 +578,11 @@ pub fn validate_tool_args(
+ #[cfg(test)]
+ mod tests {
+     use std::collections::HashMap;
+-    use std::sync::Mutex;
++    use std::sync::{Arc, Mutex};
+ 
++    use async_trait::async_trait;
+     use fabro_llm::types::{ToolCall, ToolDefinition};
++    use fabro_model::AgentProfileKind;
+ 
+     use super::*;
+     use crate::config::{
+@@ -428,6 +590,10 @@ mod tests {
+     };
+     use crate::event::Emitter;
+     use crate::local_sandbox::LocalSandbox;
++    use crate::question_tools::{
++        AgentQuestion, AgentQuestionAnswer, AgentQuestionAnswerStatus, AgentQuestionRuntime,
++        AgentToolRuntime, register_question_tools,
++    };
+     use crate::read_before_write_sandbox::ReadBeforeWriteSandbox;
+     use crate::test_support::MutableMockSandbox;
+     use crate::tool_registry::{RegisteredTool, ToolContext, ToolRegistry};
+@@ -505,6 +671,125 @@ mod tests {
+         }
+     }
+ 
++    struct StubQuestionRuntime;
++
++    #[async_trait]
++    impl AgentQuestionRuntime for StubQuestionRuntime {
++        async fn ask_questions(
++            &self,
++            _tool_call_id: &str,
++            questions: Vec<AgentQuestion>,
++            _cancel_token: CancellationToken,
++        ) -> Result<Vec<AgentQuestionAnswer>, String> {
++            Ok(questions
++                .into_iter()
++                .map(|question| AgentQuestionAnswer {
++                    original_id:       question.original_id,
++                    original_question: question.original_question,
++                    answers:           vec!["Ship".to_string()],
++                    status:            AgentQuestionAnswerStatus::Answered,
++                })
++                .collect())
++        }
++    }
++
++    #[tokio::test]
++    async fn question_tool_round_rejects_non_question_peers_and_preserves_order() {
++        let mut registry = ToolRegistry::new();
++        register_question_tools(AgentProfileKind::OpenAi, &mut registry);
++        registry.register(make_echo_tool());
++        let tool_calls = vec![
++            make_tool_call(
++                "request_user_input",
++                "call_question",
++                serde_json::json!({
++                    "questions": [{
++                        "id": "q1",
++                        "header": "Decision",
++                        "question": "Ship it?",
++                        "options": [{ "label": "Ship" }]
++                    }]
++                }),
++            ),
++            make_tool_call("echo", "call_echo", serde_json::json!({"text": "hello"})),
++        ];
++        let runtime = AgentToolRuntime::with_question_runtime(Arc::new(StubQuestionRuntime));
++
++        let results = execute_tool_calls(
++            &tool_calls,
++            true,
++            &registry,
++            Arc::new(LocalSandbox::new(std::env::current_dir().unwrap())),
++            None,
++            &CancellationToken::new(),
++            &SessionOptions::default(),
++            &Emitter::new(),
++            "root",
++            "root",
++            None,
++            &runtime,
++        )
++        .await;
++
++        assert_eq!(results.len(), 2);
++        assert_eq!(results[0].tool_call_id, "call_question");
++        assert!(!results[0].is_error);
++        assert_eq!(results[1].tool_call_id, "call_echo");
++        assert!(results[1].is_error);
++        assert!(
++            results[1]
++                .content
++                .as_str()
++                .unwrap()
++                .contains("human-question tools must run alone")
++        );
++    }
++
++    #[tokio::test]
++    async fn multiple_question_tool_calls_execute_only_first() {
++        let mut registry = ToolRegistry::new();
++        register_question_tools(AgentProfileKind::OpenAi, &mut registry);
++        let question_args = serde_json::json!({
++            "questions": [{
++                "id": "q1",
++                "header": "Decision",
++                "question": "Ship it?",
++                "options": [{ "label": "Ship" }]
++            }]
++        });
++        let tool_calls = vec![
++            make_tool_call("request_user_input", "call_first", question_args.clone()),
++            make_tool_call("request_user_input", "call_second", question_args),
++        ];
++        let runtime = AgentToolRuntime::with_question_runtime(Arc::new(StubQuestionRuntime));
++
++        let results = execute_tool_calls(
++            &tool_calls,
++            true,
++            &registry,
++            Arc::new(LocalSandbox::new(std::env::current_dir().unwrap())),
++            None,
++            &CancellationToken::new(),
++            &SessionOptions::default(),
++            &Emitter::new(),
++            "root",
++            "root",
++            None,
++            &runtime,
++        )
++        .await;
++
++        assert!(!results[0].is_error);
++        assert!(results[1].is_error);
++        assert!(
++            results[1]
++                .content
++                .as_str()
++                .unwrap()
++                .contains("Combine all questions into a single questions[] batch")
++        );
++    }
++
+     struct MockHookCallback {
+         pre_decision:       ToolHookDecision,
+         post_calls:         Arc<Mutex<Vec<(String, String, String)>>>,
+diff --git a/lib/crates/fabro-api/tests/interview_option_round_trip.rs b/lib/crates/fabro-api/tests/interview_option_round_trip.rs
+index 40fb51b23..06f4cff82 100644
+--- a/lib/crates/fabro-api/tests/interview_option_round_trip.rs
++++ b/lib/crates/fabro-api/tests/interview_option_round_trip.rs
+@@ -13,7 +13,9 @@ fn interview_option_reuses_canonical_type() {
+ fn interview_option_round_trips_representative_json() {
+     let value = json!({
+         "key": "approve",
+-        "label": "Approve"
++        "label": "Approve",
++        "description": "Approve the proposed changes.",
++        "preview": "diff --stat output"
+     });
+ 
+     let option: InterviewOption = serde_json::from_value(value.clone()).unwrap();
+diff --git a/lib/crates/fabro-api/tests/interview_question_record_round_trip.rs b/lib/crates/fabro-api/tests/interview_question_record_round_trip.rs
+index 8e8f743b6..14338c4b4 100644
+--- a/lib/crates/fabro-api/tests/interview_question_record_round_trip.rs
++++ b/lib/crates/fabro-api/tests/interview_question_record_round_trip.rs
+@@ -17,7 +17,12 @@ fn interview_question_record_round_trips_representative_json() {
+         "stage": "gate",
+         "question_type": "multiple_choice",
+         "options": [
+-            { "key": "approve", "label": "Approve" },
++            {
++                "key": "approve",
++                "label": "Approve",
++                "description": "Deploy now",
++                "preview": "deploy --prod"
++            },
+             { "key": "reject", "label": "Reject" }
+         ],
+         "allow_freeform": true,
+diff --git a/lib/crates/fabro-api/tests/pending_interview_record_round_trip.rs b/lib/crates/fabro-api/tests/pending_interview_record_round_trip.rs
+index 9ce7c4c9c..dae81fdc3 100644
+--- a/lib/crates/fabro-api/tests/pending_interview_record_round_trip.rs
++++ b/lib/crates/fabro-api/tests/pending_interview_record_round_trip.rs
+@@ -18,7 +18,12 @@ fn pending_interview_record_round_trips_populated_question() {
+             "stage": "gate",
+             "question_type": "multiple_choice",
+             "options": [
+-                { "key": "approve", "label": "Approve" },
++                {
++                    "key": "approve",
++                    "label": "Approve",
++                    "description": "Deploy now",
++                    "preview": "deploy --prod"
++                },
+                 { "key": "reject", "label": "Reject" }
+             ],
+             "allow_freeform": true,
+diff --git a/lib/crates/fabro-api/tests/run_projection_round_trip.rs b/lib/crates/fabro-api/tests/run_projection_round_trip.rs
+index 00aeb7ae2..64a00df91 100644
+--- a/lib/crates/fabro-api/tests/run_projection_round_trip.rs
++++ b/lib/crates/fabro-api/tests/run_projection_round_trip.rs
+@@ -54,7 +54,12 @@ fn run_projection_round_trips_populated_projection() {
+                     "stage": "gate",
+                     "question_type": "multiple_choice",
+                     "options": [
+-                        { "key": "approve", "label": "Approve" },
++                        {
++                            "key": "approve",
++                            "label": "Approve",
++                            "description": "Deploy now",
++                            "preview": "deploy --prod"
++                        },
+                         { "key": "reject", "label": "Reject" }
+                     ],
+                     "allow_freeform": true,
+diff --git a/lib/crates/fabro-cli/src/commands/run/attach.rs b/lib/crates/fabro-cli/src/commands/run/attach.rs
+index d66249aad..c8ad2fff2 100644
+--- a/lib/crates/fabro-cli/src/commands/run/attach.rs
++++ b/lib/crates/fabro-cli/src/commands/run/attach.rs
+@@ -426,8 +426,10 @@ fn api_question_to_question(question: &types::ApiQuestion) -> Question {
+         .options
+         .iter()
+         .map(|option| InterviewOption {
+-            key:   option.key.clone(),
+-            label: option.label.clone(),
++            key:         option.key.clone(),
++            label:       option.label.clone(),
++            description: option.description.clone(),
++            preview:     option.preview.clone(),
+         })
+         .collect();
+     converted.allow_freeform = question.allow_freeform;
+@@ -966,8 +968,10 @@ mod tests {
+     fn invalid_multiple_choice_without_freeform_is_user_correctable() {
+         let mut question = Question::new("Pick one.", QuestionType::MultipleChoice);
+         question.options = vec![InterviewOption {
+-            key:   "A".to_string(),
+-            label: "Approve".to_string(),
++            key:         "A".to_string(),
++            label:       "Approve".to_string(),
++            description: None,
++            preview:     None,
+         }];
+ 
+         let response = parse_choice_response(&question, PromptRead::Line("bogus".to_string()));
+@@ -979,8 +983,10 @@ mod tests {
+     fn unmatched_multiple_choice_with_freeform_remains_text() {
+         let mut question = Question::new("Pick one.", QuestionType::MultipleChoice);
+         question.options = vec![InterviewOption {
+-            key:   "A".to_string(),
+-            label: "Approve".to_string(),
++            key:         "A".to_string(),
++            label:       "Approve".to_string(),
++            description: None,
++            preview:     None,
+         }];
+         question.allow_freeform = true;
+ 
+@@ -1000,12 +1006,16 @@ mod tests {
+         let mut question = Question::new("Pick many.", QuestionType::MultiSelect);
+         question.options = vec![
+             InterviewOption {
+-                key:   "A".to_string(),
+-                label: "Approve".to_string(),
++                key:         "A".to_string(),
++                label:       "Approve".to_string(),
++                description: None,
++                preview:     None,
+             },
+             InterviewOption {
+-                key:   "N".to_string(),
+-                label: "Notify".to_string(),
++                key:         "N".to_string(),
++                label:       "Notify".to_string(),
++                description: None,
++                preview:     None,
+             },
+         ];
+ 
+diff --git a/lib/crates/fabro-interview/src/auto_approve.rs b/lib/crates/fabro-interview/src/auto_approve.rs
+index faeba9a7b..66a323548 100644
+--- a/lib/crates/fabro-interview/src/auto_approve.rs
++++ b/lib/crates/fabro-interview/src/auto_approve.rs
+@@ -72,12 +72,16 @@ mod tests {
+         let mut q = Question::new("Choose:", QuestionType::MultipleChoice);
+         q.options = vec![
+             InterviewOption {
+-                key:   "A".to_string(),
+-                label: "Alpha".to_string(),
++                key:         "A".to_string(),
++                label:       "Alpha".to_string(),
++                description: None,
++                preview:     None,
+             },
+             InterviewOption {
+-                key:   "B".to_string(),
+-                label: "Beta".to_string(),
++                key:         "B".to_string(),
++                label:       "Beta".to_string(),
++                description: None,
++                preview:     None,
+             },
+         ];
+         let answer = interviewer.ask(q).await.answer;
+@@ -85,8 +89,10 @@ mod tests {
+         assert_eq!(
+             answer.selected_option,
+             Some(InterviewOption {
+-                key:   "A".to_string(),
+-                label: "Alpha".to_string(),
++                key:         "A".to_string(),
++                label:       "Alpha".to_string(),
++                description: None,
++                preview:     None,
+             })
+         );
+     }
+diff --git a/lib/crates/fabro-interview/src/console.rs b/lib/crates/fabro-interview/src/console.rs
+index 437025041..0005b1aca 100644
+--- a/lib/crates/fabro-interview/src/console.rs
++++ b/lib/crates/fabro-interview/src/console.rs
+@@ -296,12 +296,16 @@ mod tests {
+     fn find_matching_option_by_key() {
+         let options = vec![
+             InterviewOption {
+-                key:   "A".to_string(),
+-                label: "Approve".to_string(),
++                key:         "A".to_string(),
++                label:       "Approve".to_string(),
++                description: None,
++                preview:     None,
+             },
+             InterviewOption {
+-                key:   "R".to_string(),
+-                label: "Reject".to_string(),
++                key:         "R".to_string(),
++                label:       "Reject".to_string(),
++                description: None,
++                preview:     None,
+             },
+         ];
+         let result = find_matching_option("A", &options);
+@@ -313,8 +317,10 @@ mod tests {
+     #[test]
+     fn find_matching_option_by_key_case_insensitive() {
+         let options = vec![InterviewOption {
+-            key:   "Y".to_string(),
+-            label: "Yes".to_string(),
++            key:         "Y".to_string(),
++            label:       "Yes".to_string(),
++            description: None,
++            preview:     None,
+         }];
+         let result = find_matching_option("y", &options);
+         assert!(result.is_some());
+@@ -324,12 +330,16 @@ mod tests {
+     fn find_matching_option_by_index() {
+         let options = vec![
+             InterviewOption {
+-                key:   "A".to_string(),
+-                label: "Alpha".to_string(),
++                key:         "A".to_string(),
++                label:       "Alpha".to_string(),
++                description: None,
++                preview:     None,
+             },
+             InterviewOption {
+-                key:   "B".to_string(),
+-                label: "Beta".to_string(),
++                key:         "B".to_string(),
++                label:       "Beta".to_string(),
++                description: None,
++                preview:     None,
+             },
+         ];
+         let result = find_matching_option("2", &options);
+@@ -341,8 +351,10 @@ mod tests {
+     #[test]
+     fn find_matching_option_no_match() {
+         let options = vec![InterviewOption {
+-            key:   "A".to_string(),
+-            label: "Alpha".to_string(),
++            key:         "A".to_string(),
++            label:       "Alpha".to_string(),
++            description: None,
++            preview:     None,
+         }];
+         let result = find_matching_option("zzz", &options);
+         assert!(result.is_none());
+@@ -351,8 +363,10 @@ mod tests {
+     #[test]
+     fn find_matching_option_index_out_of_range() {
+         let options = vec![InterviewOption {
+-            key:   "A".to_string(),
+-            label: "Alpha".to_string(),
++            key:         "A".to_string(),
++            label:       "Alpha".to_string(),
++            description: None,
++            preview:     None,
+         }];
+         let result = find_matching_option("5", &options);
+         assert!(result.is_none());
+@@ -362,8 +376,10 @@ mod tests {
+     fn non_tty_multiple_choice_eof_returns_interrupted() {
+         let mut question = Question::new("Approve?", QuestionType::MultipleChoice);
+         question.options = vec![InterviewOption {
+-            key:   "A".to_string(),
+-            label: "Approve".to_string(),
++            key:         "A".to_string(),
++            label:       "Approve".to_string(),
++            description: None,
++            preview:     None,
+         }];
+ 
+         let answer = parse_non_tty_choice_response(&question, PromptRead::Eof);
+diff --git a/lib/crates/fabro-interview/src/lib.rs b/lib/crates/fabro-interview/src/lib.rs
+index c242cf14a..4c050c4f7 100644
+--- a/lib/crates/fabro-interview/src/lib.rs
++++ b/lib/crates/fabro-interview/src/lib.rs
+@@ -301,8 +301,10 @@ mod tests {
+     #[test]
+     fn answer_selected() {
+         let opt = InterviewOption {
+-            key:   "A".to_string(),
+-            label: "Approve".to_string(),
++            key:         "A".to_string(),
++            label:       "Approve".to_string(),
++            description: None,
++            preview:     None,
+         };
+         let a = Answer::selected("A", opt.clone());
+         assert_eq!(a.value, AnswerValue::Selected("A".to_string()));
+@@ -319,12 +321,16 @@ mod tests {
+     #[test]
+     fn question_option_eq() {
+         let a = InterviewOption {
+-            key:   "Y".to_string(),
+-            label: "Yes".to_string(),
++            key:         "Y".to_string(),
++            label:       "Yes".to_string(),
++            description: None,
++            preview:     None,
+         };
+         let b = InterviewOption {
+-            key:   "Y".to_string(),
+-            label: "Yes".to_string(),
++            key:         "Y".to_string(),
++            label:       "Yes".to_string(),
++            description: None,
++            preview:     None,
+         };
+         assert_eq!(a, b);
+     }
+diff --git a/lib/crates/fabro-server/src/demo/mod.rs b/lib/crates/fabro-server/src/demo/mod.rs
+index a71d9c748..cf533295b 100644
+--- a/lib/crates/fabro-server/src/demo/mod.rs
++++ b/lib/crates/fabro-server/src/demo/mod.rs
+@@ -1677,13 +1677,17 @@ mod runs {
+                 stage:           "review".into(),
+                 question_type:   QuestionType::YesNo,
+                 options:         vec![
+-                    ApiQuestionOption {
+-                        key:   "yes".into(),
+-                        label: "Yes".into(),
++                    InterviewOption {
++                        key:         "yes".into(),
++                        label:       "Yes".into(),
++                        description: None,
++                        preview:     None,
+                     },
+-                    ApiQuestionOption {
+-                        key:   "no".into(),
+-                        label: "No".into(),
++                    InterviewOption {
++                        key:         "no".into(),
++                        label:       "No".into(),
++                        description: None,
++                        preview:     None,
+                     },
+                 ],
+                 allow_freeform:  false,
+@@ -1696,13 +1700,17 @@ mod runs {
+                 stage:           "migration".into(),
+                 question_type:   QuestionType::MultipleChoice,
+                 options:         vec![
+-                    ApiQuestionOption {
+-                        key:   "incremental".into(),
+-                        label: "Incremental migration".into(),
++                    InterviewOption {
++                        key:         "incremental".into(),
++                        label:       "Incremental migration".into(),
++                        description: None,
++                        preview:     None,
+                     },
+-                    ApiQuestionOption {
+-                        key:   "big_bang".into(),
+-                        label: "Big-bang rewrite".into(),
++                    InterviewOption {
++                        key:         "big_bang".into(),
++                        label:       "Big-bang rewrite".into(),
++                        description: None,
++                        preview:     None,
+                     },
+                 ],
+                 allow_freeform:  true,
+diff --git a/lib/crates/fabro-server/src/server.rs b/lib/crates/fabro-server/src/server.rs
+index 9bbf93cff..701697a2c 100644
+--- a/lib/crates/fabro-server/src/server.rs
++++ b/lib/crates/fabro-server/src/server.rs
+@@ -22,10 +22,10 @@ use base64::Engine as _;
+ use base64::engine::general_purpose::STANDARD as BASE64_STANDARD;
+ use bytes::Bytes;
+ pub use fabro_api::types::{
+-    AggregateBilling, AggregateBillingTotals, ApiQuestion, ApiQuestionOption, AppendEventResponse,
+-    ArtifactEntry, ArtifactListResponse, BillingByModel, BillingStageRef,
+-    CloseRunPullRequestResponse, CompletionContentPart, CompletionMessage, CompletionMessageRole,
+-    CompletionResponse, CompletionToolChoiceMode, CompletionUsage, CreateCompletionRequest,
++    AggregateBilling, AggregateBillingTotals, ApiQuestion, AppendEventResponse, ArtifactEntry,
++    ArtifactListResponse, BillingByModel, BillingStageRef, CloseRunPullRequestResponse,
++    CompletionContentPart, CompletionMessage, CompletionMessageRole, CompletionResponse,
++    CompletionToolChoiceMode, CompletionUsage, CreateCompletionRequest,
+     CreateRunPullRequestRequest, CreateSecretRequest, DeleteRunResponse, DeleteRunSandbox,
+     DeleteSecretRequest, DiskUsageResponse, DiskUsageRunRow, DiskUsageSummaryRow, ForkRequest,
+     ForkResponse, LinkRunPullRequestRequest, MergeRunPullRequestRequest,
+@@ -2894,14 +2894,7 @@ fn api_question_from_interview_record(question: &InterviewQuestionRecord) -> Api
+         text:            question.text.clone(),
+         stage:           question.stage.clone(),
+         question_type:   question.question_type,
+-        options:         question
+-            .options
+-            .iter()
+-            .map(|option| ApiQuestionOption {
+-                key:   option.key.clone(),
+-                label: option.label.clone(),
+-            })
+-            .collect(),
++        options:         question.options.clone(),
+         allow_freeform:  question.allow_freeform,
+         timeout_seconds: question.timeout_seconds,
+         context_display: question.context_display.clone(),
+diff --git a/lib/crates/fabro-server/src/server/tests.rs b/lib/crates/fabro-server/src/server/tests.rs
+index a9ad98fd5..5829dfccc 100644
+--- a/lib/crates/fabro-server/src/server/tests.rs
++++ b/lib/crates/fabro-server/src/server/tests.rs
+@@ -4861,8 +4861,10 @@ async fn submit_pending_interview_answer_rejects_invalid_answer_shape() {
+             stage:           "gate".to_string(),
+             question_type:   QuestionType::MultipleChoice,
+             options:         vec![fabro_types::run_event::InterviewOption {
+-                key:   "approve".to_string(),
+-                label: "Approve".to_string(),
++                key:         "approve".to_string(),
++                label:       "Approve".to_string(),
++                description: None,
++                preview:     None,
+             }],
+             allow_freeform:  false,
+             timeout_seconds: None,
+@@ -4948,8 +4950,10 @@ fn answer_from_typed_selected_request_validates_and_attaches_option() {
+         stage:           "gate".to_string(),
+         question_type:   QuestionType::MultipleChoice,
+         options:         vec![fabro_types::run_event::InterviewOption {
+-            key:   "approve".to_string(),
+-            label: "Approve".to_string(),
++            key:         "approve".to_string(),
++            label:       "Approve".to_string(),
++            description: None,
++            preview:     None,
+         }],
+         allow_freeform:  false,
+         timeout_seconds: None,
+@@ -4979,12 +4983,16 @@ fn answer_from_typed_multi_selected_request_validates_option_keys() {
+         question_type:   QuestionType::MultiSelect,
+         options:         vec![
+             fabro_types::run_event::InterviewOption {
+-                key:   "approve".to_string(),
+-                label: "Approve".to_string(),
++                key:         "approve".to_string(),
++                label:       "Approve".to_string(),
++                description: None,
++                preview:     None,
+             },
+             fabro_types::run_event::InterviewOption {
+-                key:   "notify".to_string(),
+-                label: "Notify".to_string(),
++                key:         "notify".to_string(),
++                label:       "Notify".to_string(),
++                description: None,
++                preview:     None,
+             },
+         ],
+         allow_freeform:  false,
+diff --git a/lib/crates/fabro-slack/src/blocks.rs b/lib/crates/fabro-slack/src/blocks.rs
+index 83a94543b..7632ab97e 100644
+--- a/lib/crates/fabro-slack/src/blocks.rs
++++ b/lib/crates/fabro-slack/src/blocks.rs
+@@ -153,6 +153,32 @@ fn lead_blocks(question: &Question, run_web_url: Option<&str>) -> Vec<Value> {
+     blocks
+ }
+ 
++fn option_descriptions_section(question: &Question) -> Option<Value> {
++    let rows = question
++        .options
++        .iter()
++        .filter_map(|option| {
++            let description = option.description.as_deref()?.trim();
++            if description.is_empty() {
++                return None;
++            }
++            Some(format!(
++                "• *{}* — {}",
++                escape_slack_controls(&option.label),
++                escape_slack_controls(description)
++            ))
++        })
++        .collect::<Vec<_>>();
++    if rows.is_empty() {
++        return None;
++    }
++    Some(text_block(&truncate_to_limit(
++        &rows.join("\n"),
++        SLACK_SECTION_TEXT_LIMIT,
++        HEADER_TRUNCATION_SUFFIX,
++    )))
++}
++
+ pub fn answered_blocks(question_text: &str, answer_text: &str) -> Vec<Value> {
+     vec![text_block(&format!(
+         "~{}~\n*Answer:* {}",
+@@ -168,6 +194,9 @@ pub fn question_to_blocks(
+     run_web_url: Option<&str>,
+ ) -> Vec<Value> {
+     let mut blocks = lead_blocks(question, run_web_url);
++    if let Some(descriptions) = option_descriptions_section(question) {
++        blocks.push(descriptions);
++    }
+ 
+     match question.question_type {
+         QuestionType::YesNo | QuestionType::Confirmation => {
+@@ -212,10 +241,26 @@ pub fn question_to_blocks(
+                 .options
+                 .iter()
+                 .map(|opt| {
+-                    json!({
++                    let mut option = json!({
+                         "text": { "type": "plain_text", "text": opt.label },
+                         "value": opt.key
+-                    })
++                    });
++                    if let Some(description) = opt
++                        .description
++                        .as_deref()
++                        .map(str::trim)
++                        .filter(|value| !value.is_empty())
++                    {
++                        option["description"] = json!({
++                            "type": "plain_text",
++                            "text": truncate_to_limit(
++                                description,
++                                75,
++                                HEADER_TRUNCATION_SUFFIX
++                            )
++                        });
++                    }
++                    option
+                 })
+                 .collect();
+             blocks.push(json!({
+@@ -293,16 +338,22 @@ mod tests {
+         let mut q = Question::new("Pick a language:", QuestionType::MultipleChoice);
+         q.options = vec![
+             InterviewOption {
+-                key:   "rs".to_string(),
+-                label: "Rust".to_string(),
++                key:         "rs".to_string(),
++                label:       "Rust".to_string(),
++                description: None,
++                preview:     None,
+             },
+             InterviewOption {
+-                key:   "ts".to_string(),
+-                label: "TypeScript".to_string(),
++                key:         "ts".to_string(),
++                label:       "TypeScript".to_string(),
++                description: None,
++                preview:     None,
+             },
+             InterviewOption {
+-                key:   "py".to_string(),
+-                label: "Python".to_string(),
++                key:         "py".to_string(),
++                label:       "Python".to_string(),
++                description: None,
++                preview:     None,
+             },
+         ];
+         let blocks = question_to_blocks("run-1", "q-3", &q, None);
+@@ -566,12 +617,16 @@ mod tests {
+         let mut q = Question::new("Select features:", QuestionType::MultiSelect);
+         q.options = vec![
+             InterviewOption {
+-                key:   "a".to_string(),
+-                label: "Auth".to_string(),
++                key:         "a".to_string(),
++                label:       "Auth".to_string(),
++                description: None,
++                preview:     None,
+             },
+             InterviewOption {
+-                key:   "b".to_string(),
+-                label: "Billing".to_string(),
++                key:         "b".to_string(),
++                label:       "Billing".to_string(),
++                description: None,
++                preview:     None,
+             },
+         ];
+         let blocks = question_to_blocks("run-1", "q-5", &q, None);
+@@ -602,4 +657,22 @@ mod tests {
+                 .contains("\"qid\":\"q-5\"")
+         );
+     }
++
++    #[test]
++    fn option_descriptions_are_rendered_and_preview_is_not_special_cased() {
++        let mut q = Question::new("Pick one:", QuestionType::MultipleChoice);
++        q.options = vec![InterviewOption {
++            key:         "ship".to_string(),
++            label:       "Ship".to_string(),
++            description: Some("Deploy <now>".to_string()),
++            preview:     Some("preview should not render".to_string()),
++        }];
++
++        let blocks_value: Value =
++            serde_json::to_value(question_to_blocks("run-1", "q-6", &q, None)).unwrap();
++        let text = blocks_value.to_string();
++
++        assert!(text.contains("Deploy &lt;now&gt;"));
++        assert!(!text.contains("preview should not render"));
++    }
+ }
+diff --git a/lib/crates/fabro-store/src/run_state.rs b/lib/crates/fabro-store/src/run_state.rs
+index 0da6e957c..246784d40 100644
+--- a/lib/crates/fabro-store/src/run_state.rs
++++ b/lib/crates/fabro-store/src/run_state.rs
+@@ -1757,12 +1757,16 @@ mod tests {
+                     question_type:   "multiple_choice".to_string(),
+                     options:         vec![
+                         InterviewOption {
+-                            key:   "approve".to_string(),
+-                            label: "Approve".to_string(),
++                            key:         "approve".to_string(),
++                            label:       "Approve".to_string(),
++                            description: Some("Ship it".to_string()),
++                            preview:     Some("deploy --prod".to_string()),
+                         },
+                         InterviewOption {
+-                            key:   "revise".to_string(),
+-                            label: "Revise".to_string(),
++                            key:         "revise".to_string(),
++                            label:       "Revise".to_string(),
++                            description: None,
++                            preview:     None,
+                         },
+                     ],
+                     allow_freeform:  true,
+@@ -1781,6 +1785,14 @@ mod tests {
+         assert_eq!(pending.question.stage, "gate");
+         assert_eq!(pending.question.question_type, QuestionType::MultipleChoice);
+         assert_eq!(pending.question.options.len(), 2);
++        assert_eq!(
++            pending.question.options[0].description.as_deref(),
++            Some("Ship it")
++        );
++        assert_eq!(
++            pending.question.options[0].preview.as_deref(),
++            Some("deploy --prod")
++        );
+         assert!(pending.question.allow_freeform);
+         assert_eq!(pending.question.timeout_seconds, Some(30.0));
+         assert_eq!(
+diff --git a/lib/crates/fabro-types/src/run_event/misc.rs b/lib/crates/fabro-types/src/run_event/misc.rs
+index b2280c243..3f88cdbd3 100644
+--- a/lib/crates/fabro-types/src/run_event/misc.rs
++++ b/lib/crates/fabro-types/src/run_event/misc.rs
+@@ -4,10 +4,14 @@ use serde_json::Value;
+ use super::ExecOutputTail;
+ use crate::{CommandTermination, PullRequestLink};
+ 
+-#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
++#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, Default)]
+ pub struct InterviewOption {
+-    pub key:   String,
+-    pub label: String,
++    pub key:         String,
++    pub label:       String,
++    #[serde(default, skip_serializing_if = "Option::is_none")]
++    pub description: Option<String>,
++    #[serde(default, skip_serializing_if = "Option::is_none")]
++    pub preview:     Option<String>,
+ }
+ 
+ #[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
+diff --git a/lib/crates/fabro-workflow/src/handler/agent.rs b/lib/crates/fabro-workflow/src/handler/agent.rs
+index 24feb5650..9bfb742b5 100644
+--- a/lib/crates/fabro-workflow/src/handler/agent.rs
++++ b/lib/crates/fabro-workflow/src/handler/agent.rs
+@@ -11,6 +11,7 @@ use super::{EngineServices, Handler, NodeTimeoutPolicy};
+ use crate::context::{Context, WorkflowContext, keys};
+ use crate::error::Error;
+ use crate::event::{Emitter, Event, StageScope};
++use crate::interview_runtime::WorkflowAgentQuestionRuntime;
+ use crate::outcome::{
+     BilledModelUsage, FailureCategory, FailureDetail, Outcome, OutcomeExt, StageOutcome,
+ };
+@@ -27,14 +28,15 @@ pub enum CodergenResult {
+ }
+ 
+ pub struct CodergenRunRequest<'a> {
+-    pub node:         &'a Node,
+-    pub prompt:       &'a str,
+-    pub context:      &'a Context,
+-    pub thread_id:    Option<&'a str>,
+-    pub emitter:      &'a Arc<Emitter>,
+-    pub sandbox:      &'a Arc<dyn Sandbox>,
+-    pub tool_hooks:   Option<Arc<dyn fabro_agent::ToolHookCallback>>,
+-    pub cancel_token: CancellationToken,
++    pub node:               &'a Node,
++    pub prompt:             &'a str,
++    pub context:            &'a Context,
++    pub thread_id:          Option<&'a str>,
++    pub emitter:            &'a Arc<Emitter>,
++    pub sandbox:            &'a Arc<dyn Sandbox>,
++    pub tool_hooks:         Option<Arc<dyn fabro_agent::ToolHookCallback>>,
++    pub cancel_token:       CancellationToken,
++    pub agent_tool_runtime: fabro_agent::AgentToolRuntime,
+ }
+ 
+ pub struct OneShotRequest<'a> {
+@@ -258,6 +260,15 @@ impl Handler for AgentHandler {
+             .or_else(|| Some(services.run.provider_id.to_string()));
+         let prompt_model = node.model().map(String::from);
+         let stage_scope = StageScope::for_handler(context, &node.id);
++        let agent_tool_runtime = fabro_agent::AgentToolRuntime::with_question_runtime(Arc::new(
++            WorkflowAgentQuestionRuntime::new(
++                Arc::clone(&services.interviewer),
++                Arc::clone(&services.run.emitter),
++                stage_scope.clone(),
++                node.id.clone(),
++                Arc::clone(&services.run.interview_blocker),
++            ),
++        ));
+         services.run.emitter.emit_scoped(
+             &Event::Prompt {
+                 stage:    node.id.clone(),
+@@ -299,6 +310,7 @@ impl Handler for AgentHandler {
+                         sandbox: &services.run.sandbox,
+                         tool_hooks,
+                         cancel_token: services.run.cancel_token(),
++                        agent_tool_runtime: agent_tool_runtime.clone(),
+                     })
+                     .await;
+                 match result {
+diff --git a/lib/crates/fabro-workflow/src/handler/fan_in.rs b/lib/crates/fabro-workflow/src/handler/fan_in.rs
+index 455b27b26..3a712ec0f 100644
+--- a/lib/crates/fabro-workflow/src/handler/fan_in.rs
++++ b/lib/crates/fabro-workflow/src/handler/fan_in.rs
+@@ -269,6 +269,7 @@ async fn llm_evaluate(
+             sandbox,
+             tool_hooks: None,
+             cancel_token,
++            agent_tool_runtime: fabro_agent::AgentToolRuntime::default(),
+         })
+         .await
+     {
+diff --git a/lib/crates/fabro-workflow/src/handler/human.rs b/lib/crates/fabro-workflow/src/handler/human.rs
+index a46a6e525..77b83454a 100644
+--- a/lib/crates/fabro-workflow/src/handler/human.rs
++++ b/lib/crates/fabro-workflow/src/handler/human.rs
+@@ -1,13 +1,12 @@
+ use std::path::Path;
+ use std::str::FromStr;
+ use std::sync::Arc;
+-use std::sync::atomic::{AtomicUsize, Ordering};
+ use std::time::Instant;
+ 
+ use async_trait::async_trait;
+ use fabro_graphviz::graph::{Graph, Node};
+ use fabro_interview::{Answer, AnswerValue, Interviewer, Question, ask_with_timeout};
+-use fabro_types::{BlockedReason, InterviewOption, Principal, QuestionType, SystemActorKind};
++use fabro_types::{InterviewOption, Principal, QuestionType, SystemActorKind};
+ use ulid::Ulid;
+ 
+ use super::{EngineServices, Handler, NodeTimeoutPolicy};
+@@ -112,8 +111,10 @@ fn build_human_gate_question(
+     question.options = choices
+         .iter()
+         .map(|choice| InterviewOption {
+-            key:   choice.key.clone(),
+-            label: choice.label.clone(),
++            key:         choice.key.clone(),
++            label:       choice.label.clone(),
++            description: None,
++            preview:     None,
+         })
+         .collect();
+     question.allow_freeform = freeform_target.is_some();
+@@ -138,61 +139,10 @@ fn build_human_gate_question(
+     })
+ }
+ 
+-/// Refcount of open interviews for this handler's run. Emits `run.blocked`
+-/// exactly once when the count transitions 0→1, and `run.unblocked` exactly
+-/// once when it transitions back to 0. Internal to `HumanHandler`; shared
+-/// across concurrent `execute` calls fanned out by `ParallelHandler`.
+-struct BlockedStateTracker {
+-    unresolved_interviews: AtomicUsize,
+-}
+-
+-impl BlockedStateTracker {
+-    fn new() -> Self {
+-        Self {
+-            unresolved_interviews: AtomicUsize::new(0),
+-        }
+-    }
+-
+-    fn interview_started(&self, emitter: &Emitter) {
+-        if self.unresolved_interviews.fetch_add(1, Ordering::AcqRel) == 0 {
+-            emitter.emit(&Event::RunBlocked {
+-                blocked_reason: BlockedReason::HumanInputRequired,
+-            });
+-        }
+-    }
+-
+-    fn interview_resolved(&self, emitter: &Emitter) {
+-        // Guard against unmatched resolves (e.g., tests that over-resolve) so
+-        // the counter cannot underflow. `compare_exchange_weak` loops until we
+-        // either observe zero (and bail) or successfully decrement.
+-        let mut current = self.unresolved_interviews.load(Ordering::Acquire);
+-        loop {
+-            if current == 0 {
+-                return;
+-            }
+-            match self.unresolved_interviews.compare_exchange_weak(
+-                current,
+-                current - 1,
+-                Ordering::AcqRel,
+-                Ordering::Acquire,
+-            ) {
+-                Ok(_) => {
+-                    if current == 1 {
+-                        emitter.emit(&Event::RunUnblocked);
+-                    }
+-                    return;
+-                }
+-                Err(observed) => current = observed,
+-            }
+-        }
+-    }
+-}
+-
+ /// Blocks until a human selects an option derived from outgoing edges.
+ pub struct HumanHandler {
+     interviewer: Arc<dyn Interviewer>,
+     emitter:     Option<Arc<Emitter>>,
+-    tracker:     BlockedStateTracker,
+ }
+ 
+ impl HumanHandler {
+@@ -200,7 +150,6 @@ impl HumanHandler {
+         Self {
+             interviewer,
+             emitter: None,
+-            tracker: BlockedStateTracker::new(),
+         }
+     }
+ 
+@@ -295,8 +244,10 @@ impl Handler for HumanHandler {
+                     .options
+                     .iter()
+                     .map(|option| InterviewOption {
+-                        key:   option.key.clone(),
+-                        label: option.label.clone(),
++                        key:         option.key.clone(),
++                        label:       option.label.clone(),
++                        description: option.description.clone(),
++                        preview:     option.preview.clone(),
+                     })
+                     .collect(),
+                 allow_freeform:  question.allow_freeform,
+@@ -305,8 +256,10 @@ impl Handler for HumanHandler {
+             },
+             &stage_scope,
+         );
+-        self.tracker
+-            .interview_started(services.run.emitter.as_ref());
++        let interview_guard = services
++            .run
++            .interview_blocker
++            .block(Arc::clone(&services.run.emitter));
+         let interview_start = Instant::now();
+         let answer_submission = ask_with_timeout(self.interviewer.as_ref(), question).await;
+         let answer_actor = answer_submission.actor.clone();
+@@ -327,8 +280,7 @@ impl Handler for HumanHandler {
+                 },
+                 &stage_scope,
+             );
+-            self.tracker
+-                .interview_resolved(services.run.emitter.as_ref());
++            interview_guard.resolve();
+             let default_choice = node
+                 .attrs
+                 .get("human.default_choice")
+@@ -371,8 +323,7 @@ impl Handler for HumanHandler {
+                 },
+                 &stage_scope,
+             );
+-            self.tracker
+-                .interview_resolved(services.run.emitter.as_ref());
++            interview_guard.resolve();
+             return Ok(unanswered_human_gate(
+                 "human interaction interrupted before an answer was provided",
+             ));
+@@ -389,8 +340,7 @@ impl Handler for HumanHandler {
+                 },
+                 &stage_scope,
+             );
+-            self.tracker
+-                .interview_resolved(services.run.emitter.as_ref());
++            interview_guard.resolve();
+             return Ok(unanswered_human_gate("human skipped interaction"));
+         }
+ 
+@@ -406,8 +356,7 @@ impl Handler for HumanHandler {
+             },
+             &stage_scope,
+         );
+-        self.tracker
+-            .interview_resolved(services.run.emitter.as_ref());
++        interview_guard.resolve();
+ 
+         // Try fixed-choice match
+         if let Some(selected) = find_choice_match(&answer, &choices) {
+@@ -894,8 +843,10 @@ mod tests {
+     async fn wait_human_emits_blocked_then_unblocked_around_interview() {
+         let interviewer = Arc::new(CallbackInterviewer::new(|_| {
+             Answer::selected("A", InterviewOption {
+-                key:   "A".to_string(),
+-                label: "Approve".to_string(),
++                key:         "A".to_string(),
++                label:       "Approve".to_string(),
++                description: None,
++                preview:     None,
+             })
+         }));
+         let handler = HumanHandler::new(interviewer);
+@@ -1128,9 +1079,10 @@ mod tests {
+ 
+     #[test]
+     fn blocked_state_tracker_emits_once_across_parallel_interview_races() {
+-        let tracker = BlockedStateTracker::new();
++        let blocker = Arc::new(crate::interview_runtime::RunInterviewBlocker::new());
+         let emitter = Arc::new(Emitter::new(fabro_types::fixtures::RUN_1));
+         let event_names = Arc::new(Mutex::new(Vec::new()));
++        let guards = Arc::new(Mutex::new(Vec::new()));
+ 
+         emitter.on_event({
+             let event_names = Arc::clone(&event_names);
+@@ -1148,22 +1100,25 @@ mod tests {
+ 
+         std::thread::scope(|scope| {
+             for _ in 0..8 {
+-                let tracker = &tracker;
++                let blocker = Arc::clone(&blocker);
+                 let emitter = Arc::clone(&emitter);
+-                scope.spawn(move || tracker.interview_started(emitter.as_ref()));
++                let guards = Arc::clone(&guards);
++                scope.spawn(move || {
++                    guards.lock().unwrap().push(blocker.block(emitter));
++                });
+             }
+         });
+ 
+         std::thread::scope(|scope| {
+             for _ in 0..8 {
+-                let tracker = &tracker;
+-                let emitter = Arc::clone(&emitter);
+-                scope.spawn(move || tracker.interview_resolved(emitter.as_ref()));
++                let guards = Arc::clone(&guards);
++                scope.spawn(move || {
++                    let guard = guards.lock().unwrap().pop().unwrap();
++                    guard.resolve();
++                });
+             }
+         });
+ 
+-        tracker.interview_resolved(emitter.as_ref());
+-
+         assert_eq!(event_names.lock().unwrap().as_slice(), [
+             "run.blocked",
+             "run.unblocked"
+diff --git a/lib/crates/fabro-workflow/src/handler/llm/acp.rs b/lib/crates/fabro-workflow/src/handler/llm/acp.rs
+index 40929cc97..316882de6 100644
+--- a/lib/crates/fabro-workflow/src/handler/llm/acp.rs
++++ b/lib/crates/fabro-workflow/src/handler/llm/acp.rs
+@@ -449,14 +449,15 @@ mod tests {
+         let context = Context::new();
+         let result = backend
+             .run(CodergenRunRequest {
+-                node:         &node,
+-                prompt:       "write hello",
+-                context:      &context,
+-                thread_id:    None,
+-                emitter:      &emitter,
+-                sandbox:      &sandbox,
+-                tool_hooks:   None,
+-                cancel_token: CancellationToken::new(),
++                node:               &node,
++                prompt:             "write hello",
++                context:            &context,
++                thread_id:          None,
++                emitter:            &emitter,
++                sandbox:            &sandbox,
++                tool_hooks:         None,
++                cancel_token:       CancellationToken::new(),
++                agent_tool_runtime: fabro_agent::AgentToolRuntime::default(),
+             })
+             .await
+             .unwrap();
+@@ -516,14 +517,15 @@ mod tests {
+         let context = Context::new();
+         let result = backend
+             .run(CodergenRunRequest {
+-                node:         &node,
+-                prompt:       "write hello",
+-                context:      &context,
+-                thread_id:    None,
+-                emitter:      &emitter,
+-                sandbox:      &sandbox,
+-                tool_hooks:   None,
+-                cancel_token: CancellationToken::new(),
++                node:               &node,
++                prompt:             "write hello",
++                context:            &context,
++                thread_id:          None,
++                emitter:            &emitter,
++                sandbox:            &sandbox,
++                tool_hooks:         None,
++                cancel_token:       CancellationToken::new(),
++                agent_tool_runtime: fabro_agent::AgentToolRuntime::default(),
+             })
+             .await
+             .unwrap();
+@@ -563,14 +565,15 @@ mod tests {
+         let context = Context::new();
+         let result = backend
+             .run(CodergenRunRequest {
+-                node:         &node,
+-                prompt:       "write hello",
+-                context:      &context,
+-                thread_id:    None,
+-                emitter:      &emitter,
+-                sandbox:      &sandbox,
+-                tool_hooks:   None,
+-                cancel_token: CancellationToken::new(),
++                node:               &node,
++                prompt:             "write hello",
++                context:            &context,
++                thread_id:          None,
++                emitter:            &emitter,
++                sandbox:            &sandbox,
++                tool_hooks:         None,
++                cancel_token:       CancellationToken::new(),
++                agent_tool_runtime: fabro_agent::AgentToolRuntime::default(),
+             })
+             .await
+             .unwrap();
+@@ -601,14 +604,15 @@ mod tests {
+         let context = Context::new();
+         let result = backend
+             .run(CodergenRunRequest {
+-                node:         &node,
+-                prompt:       "write hello",
+-                context:      &context,
+-                thread_id:    None,
+-                emitter:      &emitter,
+-                sandbox:      &sandbox_dyn,
+-                tool_hooks:   None,
+-                cancel_token: CancellationToken::new(),
++                node:               &node,
++                prompt:             "write hello",
++                context:            &context,
++                thread_id:          None,
++                emitter:            &emitter,
++                sandbox:            &sandbox_dyn,
++                tool_hooks:         None,
++                cancel_token:       CancellationToken::new(),
++                agent_tool_runtime: fabro_agent::AgentToolRuntime::default(),
+             })
+             .await;
+         assert!(result.is_err());
+@@ -650,14 +654,15 @@ mod tests {
+         let context = Context::new();
+         let result = backend
+             .run(CodergenRunRequest {
+-                node:         &node,
+-                prompt:       "cancel",
+-                context:      &context,
+-                thread_id:    None,
+-                emitter:      &emitter,
+-                sandbox:      &sandbox,
+-                tool_hooks:   None,
+-                cancel_token: CancellationToken::new(),
++                node:               &node,
++                prompt:             "cancel",
++                context:            &context,
++                thread_id:          None,
++                emitter:            &emitter,
++                sandbox:            &sandbox,
++                tool_hooks:         None,
++                cancel_token:       CancellationToken::new(),
++                agent_tool_runtime: fabro_agent::AgentToolRuntime::default(),
+             })
+             .await;
+         let Err(err) = result else {
+@@ -703,14 +708,15 @@ mod tests {
+         let context = Context::new();
+         backend
+             .run(CodergenRunRequest {
+-                node:         &node,
+-                prompt:       "write hello",
+-                context:      &context,
+-                thread_id:    None,
+-                emitter:      &emitter,
+-                sandbox:      &sandbox,
+-                tool_hooks:   None,
+-                cancel_token: CancellationToken::new(),
++                node:               &node,
++                prompt:             "write hello",
++                context:            &context,
++                thread_id:          None,
++                emitter:            &emitter,
++                sandbox:            &sandbox,
++                tool_hooks:         None,
++                cancel_token:       CancellationToken::new(),
++                agent_tool_runtime: fabro_agent::AgentToolRuntime::default(),
+             })
+             .await
+             .unwrap();
+@@ -744,14 +750,15 @@ mod tests {
+         let context = Context::new();
+         let result = backend
+             .run(CodergenRunRequest {
+-                node:         &node,
+-                prompt:       "write hello",
+-                context:      &context,
+-                thread_id:    None,
+-                emitter:      &emitter,
+-                sandbox:      &sandbox_dyn,
+-                tool_hooks:   None,
+-                cancel_token: CancellationToken::new(),
++                node:               &node,
++                prompt:             "write hello",
++                context:            &context,
++                thread_id:          None,
++                emitter:            &emitter,
++                sandbox:            &sandbox_dyn,
++                tool_hooks:         None,
++                cancel_token:       CancellationToken::new(),
++                agent_tool_runtime: fabro_agent::AgentToolRuntime::default(),
+             })
+             .await;
+         let Err(err) = result else {
+@@ -796,14 +803,15 @@ mod tests {
+         let context = Context::new();
+         let result = backend
+             .run(CodergenRunRequest {
+-                node:         &node,
+-                prompt:       "write hello",
+-                context:      &context,
+-                thread_id:    None,
+-                emitter:      &emitter,
+-                sandbox:      &sandbox_dyn,
+-                tool_hooks:   None,
+-                cancel_token: CancellationToken::new(),
++                node:               &node,
++                prompt:             "write hello",
++                context:            &context,
++                thread_id:          None,
++                emitter:            &emitter,
++                sandbox:            &sandbox_dyn,
++                tool_hooks:         None,
++                cancel_token:       CancellationToken::new(),
++                agent_tool_runtime: fabro_agent::AgentToolRuntime::default(),
+             })
+             .await;
+         let Err(err) = result else {
+diff --git a/lib/crates/fabro-workflow/src/handler/llm/api.rs b/lib/crates/fabro-workflow/src/handler/llm/api.rs
+index 38f20a5d9..41c6ca1aa 100644
+--- a/lib/crates/fabro-workflow/src/handler/llm/api.rs
++++ b/lib/crates/fabro-workflow/src/handler/llm/api.rs
+@@ -7,7 +7,7 @@ use fabro_agent::tool_registry::{RegisteredTool, ToolContext, ToolRegistry};
+ use fabro_agent::{
+     AgentEvent, AgentProfile, AnthropicProfile, CompletionCoordinator, GeminiProfile,
+     Message as AgentMessage, OpenAiProfile, Sandbox, Session, SessionOptions, StaticEnvProvider,
+-    ToolEnvProvider,
++    ToolEnvProvider, register_question_tools,
+ };
+ use fabro_auth::{CredentialSource, EnvCredentialSource};
+ use fabro_graphviz::graph::{AttrValue, Node};
+@@ -733,6 +733,7 @@ impl AgentApiBackend {
+         });
+ 
+         profile.register_subagent_tools(manager, factory, 0);
++        register_question_tools(provider.profile_kind, profile.tool_registry_mut());
+         if let Some(services) = fabro_run_tools {
+             register_fabro_run_tools(profile.tool_registry_mut(), &services);
+         }
+@@ -971,6 +972,7 @@ impl CodergenBackend for AgentApiBackend {
+         let sandbox = request.sandbox;
+         let tool_hooks = request.tool_hooks;
+         let cancel_token = request.cancel_token;
++        let agent_tool_runtime = request.agent_tool_runtime;
+ 
+         let fidelity = context.fidelity();
+         let reuse_key = if fidelity == Fidelity::Full {
+@@ -1081,7 +1083,9 @@ impl CodergenBackend for AgentApiBackend {
+                         return Err(err);
+                     }
+                 }
+-                session.process_input(prompt).await
++                session
++                    .process_input_with_runtime(prompt, agent_tool_runtime.clone())
++                    .await
+             }
+             Err(err) => Err(err),
+         };
+@@ -1205,7 +1209,10 @@ impl CodergenBackend for AgentApiBackend {
+                                 return Err(err);
+                             }
+                         }
+-                        match session.process_input(prompt).await {
++                        match session
++                            .process_input_with_runtime(prompt, agent_tool_runtime.clone())
++                            .await
++                        {
+                             Ok(()) => {
+                                 succeeded = true;
+                                 break;
+diff --git a/lib/crates/fabro-workflow/src/handler/manager_loop.rs b/lib/crates/fabro-workflow/src/handler/manager_loop.rs
+index cf6fcaf8e..d7e1dff7b 100644
+--- a/lib/crates/fabro-workflow/src/handler/manager_loop.rs
++++ b/lib/crates/fabro-workflow/src/handler/manager_loop.rs
+@@ -231,6 +231,7 @@ impl Handler for SubWorkflowHandler {
+ 
+         let parent_run = Arc::clone(&services.run);
+         let registry = Arc::clone(&services.registry);
++        let interviewer = Arc::clone(&services.interviewer);
+         let base_env = services.base_env.clone();
+         let github_token = services.github_token.clone();
+         let inputs = services.inputs.clone();
+@@ -269,6 +270,7 @@ impl Handler for SubWorkflowHandler {
+                 engine:        Arc::new(EngineServices {
+                     run: child_run,
+                     registry,
++                    interviewer,
+                     git_state: std::sync::RwLock::new(None),
+                     base_env,
+                     github_token,
+diff --git a/lib/crates/fabro-workflow/src/handler/parallel.rs b/lib/crates/fabro-workflow/src/handler/parallel.rs
+index 4ec7fbcd8..58cbe93f5 100644
+--- a/lib/crates/fabro-workflow/src/handler/parallel.rs
++++ b/lib/crates/fabro-workflow/src/handler/parallel.rs
+@@ -304,6 +304,7 @@ impl Handler for ParallelHandler {
+         for setup in branch_setups {
+             let parent_run = Arc::clone(&services.run);
+             let registry = Arc::clone(&services.registry);
++            let interviewer = Arc::clone(&services.interviewer);
+             let base_env = services.base_env.clone();
+             let github_token = services.github_token.clone();
+             let inputs = services.inputs.clone();
+@@ -375,6 +376,7 @@ impl Handler for ParallelHandler {
+                 let branch_services = EngineServices {
+                     run: parent_run.with_sandbox(Arc::clone(&setup.sandbox)),
+                     registry: Arc::clone(&registry),
++                    interviewer,
+                     git_state: std::sync::RwLock::new(None),
+                     base_env: base_env.clone(),
+                     github_token: github_token.clone(),
+diff --git a/lib/crates/fabro-workflow/src/interview_runtime.rs b/lib/crates/fabro-workflow/src/interview_runtime.rs
+new file mode 100644
+index 000000000..e844de8b3
+--- /dev/null
++++ b/lib/crates/fabro-workflow/src/interview_runtime.rs
+@@ -0,0 +1,632 @@
++use std::sync::Arc;
++use std::sync::atomic::{AtomicUsize, Ordering};
++use std::time::Instant;
++
++use async_trait::async_trait;
++use fabro_agent::{
++    AgentQuestion, AgentQuestionAnswer, AgentQuestionAnswerStatus, AgentQuestionRuntime,
++};
++use fabro_interview::{Answer, AnswerSubmission, AnswerValue, Interviewer, Question};
++use fabro_types::{BlockedReason, InterviewOption, Principal, SystemActorKind};
++use futures::future;
++use tokio_util::sync::CancellationToken;
++use ulid::Ulid;
++
++use crate::event::{Emitter, Event, StageScope};
++use crate::millis_u64;
++
++/// Run-scoped refcount for unresolved human input. Emits `run.blocked` on the
++/// first unresolved human/agent interview and `run.unblocked` after the last
++/// one resolves.
++pub(crate) struct RunInterviewBlocker {
++    unresolved_interviews: AtomicUsize,
++}
++
++impl RunInterviewBlocker {
++    #[must_use]
++    pub(crate) fn new() -> Self {
++        Self {
++            unresolved_interviews: AtomicUsize::new(0),
++        }
++    }
++
++    pub(crate) fn block(self: &Arc<Self>, emitter: Arc<Emitter>) -> RunInterviewGuard {
++        if self.unresolved_interviews.fetch_add(1, Ordering::AcqRel) == 0 {
++            emitter.emit(&Event::RunBlocked {
++                blocked_reason: BlockedReason::HumanInputRequired,
++            });
++        }
++        RunInterviewGuard {
++            blocker: Arc::clone(self),
++            emitter,
++            resolved: false,
++        }
++    }
++
++    fn resolved(&self, emitter: &Emitter) {
++        let mut current = self.unresolved_interviews.load(Ordering::Acquire);
++        loop {
++            if current == 0 {
++                return;
++            }
++            match self.unresolved_interviews.compare_exchange_weak(
++                current,
++                current - 1,
++                Ordering::AcqRel,
++                Ordering::Acquire,
++            ) {
++                Ok(_) => {
++                    if current == 1 {
++                        emitter.emit(&Event::RunUnblocked);
++                    }
++                    return;
++                }
++                Err(observed) => current = observed,
++            }
++        }
++    }
++}
++
++impl Default for RunInterviewBlocker {
++    fn default() -> Self {
++        Self::new()
++    }
++}
++
++pub(crate) struct RunInterviewGuard {
++    blocker:  Arc<RunInterviewBlocker>,
++    emitter:  Arc<Emitter>,
++    resolved: bool,
++}
++
++impl RunInterviewGuard {
++    pub(crate) fn resolve(mut self) {
++        if !self.resolved {
++            self.blocker.resolved(self.emitter.as_ref());
++            self.resolved = true;
++        }
++    }
++}
++
++impl Drop for RunInterviewGuard {
++    fn drop(&mut self) {
++        if !self.resolved {
++            self.blocker.resolved(self.emitter.as_ref());
++            self.resolved = true;
++        }
++    }
++}
++
++pub(crate) struct WorkflowAgentQuestionRuntime {
++    interviewer: Arc<dyn Interviewer>,
++    emitter:     Arc<Emitter>,
++    stage_scope: StageScope,
++    stage_id:    String,
++    blocker:     Arc<RunInterviewBlocker>,
++}
++
++impl WorkflowAgentQuestionRuntime {
++    #[must_use]
++    pub(crate) fn new(
++        interviewer: Arc<dyn Interviewer>,
++        emitter: Arc<Emitter>,
++        stage_scope: StageScope,
++        stage_id: impl Into<String>,
++        blocker: Arc<RunInterviewBlocker>,
++    ) -> Self {
++        Self {
++            interviewer,
++            emitter,
++            stage_scope,
++            stage_id: stage_id.into(),
++            blocker,
++        }
++    }
++}
++
++struct PreparedQuestion {
++    agent_question: AgentQuestion,
++    question:       Question,
++}
++
++struct PendingAgentQuestionBatch {
++    emitter:     Arc<Emitter>,
++    stage_scope: StageScope,
++    stage_id:    String,
++    questions:   Vec<(String, String)>,
++    started_at:  Instant,
++    guard:       Option<RunInterviewGuard>,
++}
++
++impl PendingAgentQuestionBatch {
++    fn new(
++        emitter: Arc<Emitter>,
++        stage_scope: StageScope,
++        stage_id: String,
++        prepared: &[PreparedQuestion],
++        guard: RunInterviewGuard,
++        started_at: Instant,
++    ) -> Self {
++        Self {
++            emitter,
++            stage_scope,
++            stage_id,
++            questions: prepared
++                .iter()
++                .map(|prepared_question| {
++                    (
++                        prepared_question.question.id.clone(),
++                        prepared_question.question.text.clone(),
++                    )
++                })
++                .collect(),
++            started_at,
++            guard: Some(guard),
++        }
++    }
++
++    fn resolve(mut self) {
++        if let Some(guard) = self.guard.take() {
++            guard.resolve();
++        }
++    }
++}
++
++impl Drop for PendingAgentQuestionBatch {
++    fn drop(&mut self) {
++        if self.guard.is_none() {
++            return;
++        }
++        let duration_ms = millis_u64(self.started_at.elapsed());
++        for (question_id, question) in &self.questions {
++            self.emitter.emit_scoped(
++                &Event::InterviewInterrupted {
++                    actor: Some(Principal::System {
++                        system_kind: SystemActorKind::Engine,
++                    }),
++                    question_id: question_id.clone(),
++                    question: question.clone(),
++                    stage: self.stage_id.clone(),
++                    reason: "interrupted".to_string(),
++                    duration_ms,
++                },
++                &self.stage_scope,
++            );
++        }
++        if let Some(guard) = self.guard.take() {
++            guard.resolve();
++        }
++    }
++}
++
++#[async_trait]
++impl AgentQuestionRuntime for WorkflowAgentQuestionRuntime {
++    async fn ask_questions(
++        &self,
++        tool_call_id: &str,
++        questions: Vec<AgentQuestion>,
++        cancel_token: CancellationToken,
++    ) -> Result<Vec<AgentQuestionAnswer>, String> {
++        if questions.is_empty() {
++            return Ok(Vec::new());
++        }
++
++        let prepared = questions
++            .into_iter()
++            .enumerate()
++            .map(|(index, question)| self.prepare_question(tool_call_id, index, question))
++            .collect::<Vec<_>>();
++
++        for prepared_question in &prepared {
++            let question = &prepared_question.question;
++            self.emitter.emit_scoped(
++                &Event::InterviewStarted {
++                    question_id:     question.id.clone(),
++                    question:        question.text.clone(),
++                    stage:           self.stage_id.clone(),
++                    question_type:   question.question_type.to_string(),
++                    options:         question.options.clone(),
++                    allow_freeform:  question.allow_freeform,
++                    timeout_seconds: None,
++                    context_display: question.context_display.clone(),
++                },
++                &self.stage_scope,
++            );
++        }
++
++        let interview_start = Instant::now();
++        let cleanup = PendingAgentQuestionBatch::new(
++            Arc::clone(&self.emitter),
++            self.stage_scope.clone(),
++            self.stage_id.clone(),
++            &prepared,
++            self.blocker.block(Arc::clone(&self.emitter)),
++            interview_start,
++        );
++        let ask_all = future::join_all(
++            prepared
++                .iter()
++                .map(|prepared_question| self.interviewer.ask(prepared_question.question.clone())),
++        );
++        tokio::pin!(ask_all);
++
++        let answers = tokio::select! {
++            submissions = &mut ask_all => Some(submissions),
++            () = cancel_token.cancelled() => None,
++        };
++
++        let results = match answers {
++            Some(submissions) => prepared
++                .iter()
++                .zip(submissions)
++                .map(|(prepared_question, submission)| {
++                    self.emit_submission_event(
++                        prepared_question,
++                        &submission,
++                        millis_u64(interview_start.elapsed()),
++                    );
++                    answer_from_submission(&prepared_question.agent_question, &submission)
++                })
++                .collect::<Vec<_>>(),
++            None => prepared
++                .iter()
++                .map(|prepared_question| {
++                    self.emit_interrupted(
++                        prepared_question,
++                        Some(Principal::System {
++                            system_kind: SystemActorKind::Engine,
++                        }),
++                        "interrupted",
++                        millis_u64(interview_start.elapsed()),
++                    );
++                    AgentQuestionAnswer {
++                        original_id:       prepared_question.agent_question.original_id.clone(),
++                        original_question: prepared_question
++                            .agent_question
++                            .original_question
++                            .clone(),
++                        answers:           Vec::new(),
++                        status:            AgentQuestionAnswerStatus::Interrupted,
++                    }
++                })
++                .collect::<Vec<_>>(),
++        };
++
++        cleanup.resolve();
++        Ok(results)
++    }
++}
++
++impl WorkflowAgentQuestionRuntime {
++    fn prepare_question(
++        &self,
++        tool_call_id: &str,
++        index: usize,
++        agent_question: AgentQuestion,
++    ) -> PreparedQuestion {
++        let mut question = Question::new(agent_question.text.clone(), agent_question.question_type);
++        question.id = internal_question_id(&self.stage_scope, tool_call_id, index);
++        question.options.clone_from(&agent_question.options);
++        question.allow_freeform = agent_question.allow_freeform;
++        question.stage.clone_from(&self.stage_id);
++        question.metadata.insert(
++            "agent.tool_call_id".to_string(),
++            serde_json::json!(tool_call_id),
++        );
++        question.metadata.insert(
++            "agent.original_question".to_string(),
++            serde_json::json!(agent_question.original_question),
++        );
++        if let Some(original_id) = &agent_question.original_id {
++            question.metadata.insert(
++                "agent.original_id".to_string(),
++                serde_json::json!(original_id),
++            );
++        }
++        if let Some(header) = &agent_question.header {
++            question
++                .metadata
++                .insert("agent.header".to_string(), serde_json::json!(header));
++        }
++        PreparedQuestion {
++            agent_question,
++            question,
++        }
++    }
++
++    fn emit_submission_event(
++        &self,
++        prepared: &PreparedQuestion,
++        submission: &AnswerSubmission,
++        duration_ms: u64,
++    ) {
++        match submission.answer.value {
++            AnswerValue::Timeout => self.emitter.emit_scoped(
++                &Event::InterviewTimeout {
++                    actor: Some(Principal::System {
++                        system_kind: SystemActorKind::Timeout,
++                    }),
++                    question_id: prepared.question.id.clone(),
++                    question: prepared.question.text.clone(),
++                    stage: self.stage_id.clone(),
++                    duration_ms,
++                },
++                &self.stage_scope,
++            ),
++            AnswerValue::Interrupted => self.emit_interrupted(
++                prepared,
++                Some(submission.actor.clone()),
++                "interrupted",
++                duration_ms,
++            ),
++            AnswerValue::Cancelled => self.emit_interrupted(
++                prepared,
++                Some(submission.actor.clone()),
++                "cancelled",
++                duration_ms,
++            ),
++            _ => self.emitter.emit_scoped(
++                &Event::InterviewCompleted {
++                    actor: Some(submission.actor.clone()),
++                    question_id: prepared.question.id.clone(),
++                    question: prepared.question.text.clone(),
++                    answer: answer_labels(&prepared.question.options, &submission.answer)
++                        .join(", "),
++                    duration_ms,
++                },
++                &self.stage_scope,
++            ),
++        }
++    }
++
++    fn emit_interrupted(
++        &self,
++        prepared: &PreparedQuestion,
++        actor: Option<Principal>,
++        reason: &str,
++        duration_ms: u64,
++    ) {
++        self.emitter.emit_scoped(
++            &Event::InterviewInterrupted {
++                actor,
++                question_id: prepared.question.id.clone(),
++                question: prepared.question.text.clone(),
++                stage: self.stage_id.clone(),
++                reason: reason.to_string(),
++                duration_ms,
++            },
++            &self.stage_scope,
++        );
++    }
++}
++
++fn answer_from_submission(
++    agent_question: &AgentQuestion,
++    submission: &AnswerSubmission,
++) -> AgentQuestionAnswer {
++    let status = match &submission.answer.value {
++        AnswerValue::Cancelled => AgentQuestionAnswerStatus::Cancelled,
++        AnswerValue::Interrupted => AgentQuestionAnswerStatus::Interrupted,
++        AnswerValue::Skipped => AgentQuestionAnswerStatus::Skipped,
++        AnswerValue::Timeout => AgentQuestionAnswerStatus::Timeout,
++        _ => AgentQuestionAnswerStatus::Answered,
++    };
++    let answers = if status == AgentQuestionAnswerStatus::Answered {
++        answer_labels(&agent_question.options, &submission.answer)
++    } else {
++        Vec::new()
++    };
++    AgentQuestionAnswer {
++        original_id: agent_question.original_id.clone(),
++        original_question: agent_question.original_question.clone(),
++        answers,
++        status,
++    }
++}
++
++fn answer_labels(options: &[InterviewOption], answer: &Answer) -> Vec<String> {
++    match &answer.value {
++        AnswerValue::Selected(key) => {
++            vec![option_label(options, key, answer.selected_option.as_ref())]
++        }
++        AnswerValue::MultiSelected(keys) => keys
++            .iter()
++            .map(|key| option_label(options, key, None))
++            .collect(),
++        AnswerValue::Text(text) => vec![text.clone()],
++        AnswerValue::Yes => vec!["yes".to_string()],
++        AnswerValue::No => vec!["no".to_string()],
++        AnswerValue::Cancelled => vec!["cancelled".to_string()],
++        AnswerValue::Interrupted => vec!["interrupted".to_string()],
++        AnswerValue::Skipped => vec!["skipped".to_string()],
++        AnswerValue::Timeout => vec!["timeout".to_string()],
++    }
++}
++
++fn option_label(
++    options: &[InterviewOption],
++    key: &str,
++    selected_option: Option<&InterviewOption>,
++) -> String {
++    selected_option
++        .filter(|option| option.key == key)
++        .or_else(|| options.iter().find(|option| option.key == key))
++        .map_or_else(|| key.to_string(), |option| option.label.clone())
++}
++
++fn internal_question_id(scope: &StageScope, tool_call_id: &str, index: usize) -> String {
++    format!(
++        "agentq-{}-v{}-{}-{}-{}",
++        slug(&scope.node_id),
++        scope.visit,
++        slug(tool_call_id),
++        index + 1,
++        Ulid::new()
++    )
++}
++
++fn slug(value: &str) -> String {
++    let mut out = value
++        .chars()
++        .filter_map(|ch| {
++            if ch.is_ascii_alphanumeric() {
++                Some(ch.to_ascii_lowercase())
++            } else if matches!(ch, '-' | '_') {
++                Some(ch)
++            } else {
++                None
++            }
++        })
++        .take(48)
++        .collect::<String>();
++    if out.is_empty() {
++        out.push('x');
++    }
++    out
++}
++
++#[cfg(test)]
++mod tests {
++    use fabro_interview::ControlInterviewer;
++    use fabro_types::{EventBody, RunId};
++
++    use super::*;
++
++    #[test]
++    fn answer_labels_return_user_facing_labels_in_submission_order() {
++        let options = vec![
++            InterviewOption {
++                key: "a".to_string(),
++                label: "Alpha".to_string(),
++                ..InterviewOption::default()
++            },
++            InterviewOption {
++                key: "b".to_string(),
++                label: "Beta".to_string(),
++                ..InterviewOption::default()
++            },
++        ];
++        let answer = Answer::multi_selected(vec!["b".to_string(), "a".to_string()]);
++
++        assert_eq!(answer_labels(&options, &answer), vec!["Beta", "Alpha"]);
++    }
++
++    #[test]
++    fn internal_question_id_includes_stage_visit_and_tool_call_context() {
++        let scope = StageScope {
++            node_id:            "Review Changes".to_string(),
++            visit:              3,
++            parallel_group_id:  None,
++            parallel_branch_id: None,
++        };
++
++        let id = internal_question_id(&scope, "call_123", 1);
++
++        assert!(id.starts_with("agentq-reviewchanges-v3-call_123-2-"));
++    }
++
++    #[tokio::test]
++    async fn batch_questions_are_all_started_before_run_is_blocked_and_return_labels() {
++        let interviewer = Arc::new(ControlInterviewer::new());
++        let emitter = Arc::new(Emitter::new(RunId::new()));
++        let events = Arc::new(std::sync::Mutex::new(Vec::new()));
++        emitter.on_event({
++            let events = Arc::clone(&events);
++            move |event| events.lock().unwrap().push(event.clone())
++        });
++        let runtime = WorkflowAgentQuestionRuntime::new(
++            interviewer.clone(),
++            Arc::clone(&emitter),
++            StageScope {
++                node_id:            "ask".to_string(),
++                visit:              1,
++                parallel_group_id:  None,
++                parallel_branch_id: None,
++            },
++            "ask",
++            Arc::new(RunInterviewBlocker::new()),
++        );
++        let option = InterviewOption {
++            key:         "ship".to_string(),
++            label:       "Ship it".to_string(),
++            description: Some("Deploy".to_string()),
++            preview:     Some("preview".to_string()),
++        };
++
++        let ask = tokio::spawn(async move {
++            runtime
++                .ask_questions(
++                    "call_1",
++                    vec![
++                        AgentQuestion {
++                            original_id:       Some("q1".to_string()),
++                            original_question: "First?".to_string(),
++                            header:            None,
++                            text:              "First?".to_string(),
++                            question_type:     fabro_types::QuestionType::MultipleChoice,
++                            options:           vec![option.clone()],
++                            allow_freeform:    true,
++                        },
++                        AgentQuestion {
++                            original_id:       Some("q2".to_string()),
++                            original_question: "Second?".to_string(),
++                            header:            None,
++                            text:              "Second?".to_string(),
++                            question_type:     fabro_types::QuestionType::MultipleChoice,
++                            options:           vec![option.clone()],
++                            allow_freeform:    true,
++                        },
++                    ],
++                    CancellationToken::new(),
++                )
++                .await
++                .unwrap()
++        });
++
++        tokio::task::yield_now().await;
++        let question_ids = {
++            let events = events.lock().unwrap();
++            assert!(matches!(events[0].body, EventBody::InterviewStarted(_)));
++            assert!(matches!(events[1].body, EventBody::InterviewStarted(_)));
++            assert!(matches!(events[2].body, EventBody::RunBlocked(_)));
++            events
++                .iter()
++                .filter_map(|event| match &event.body {
++                    EventBody::InterviewStarted(props) => Some(props.question_id.clone()),
++                    _ => None,
++                })
++                .collect::<Vec<_>>()
++        };
++
++        for question_id in question_ids {
++            let option = InterviewOption {
++                key: "ship".to_string(),
++                label: "Ship it".to_string(),
++                ..InterviewOption::default()
++            };
++            interviewer
++                .submit(
++                    &question_id,
++                    AnswerSubmission::system(
++                        Answer::selected("ship", option),
++                        SystemActorKind::Engine,
++                    ),
++                )
++                .await
++                .unwrap();
++        }
++
++        let answers = ask.await.unwrap();
++
++        assert_eq!(answers.len(), 2);
++        assert_eq!(answers[0].answers, vec!["Ship it"]);
++        assert_eq!(answers[1].answers, vec!["Ship it"]);
++        assert!(
++            events
++                .lock()
++                .unwrap()
++                .iter()
++                .any(|event| matches!(event.body, EventBody::RunUnblocked(_)))
++        );
++    }
++}
+diff --git a/lib/crates/fabro-workflow/src/lib.rs b/lib/crates/fabro-workflow/src/lib.rs
+index 194a0ca2c..42f5c7f43 100644
+--- a/lib/crates/fabro-workflow/src/lib.rs
++++ b/lib/crates/fabro-workflow/src/lib.rs
+@@ -298,6 +298,7 @@ pub mod github_token_source;
+ pub(crate) mod graph;
+ pub mod handler;
+ mod hook_context;
++mod interview_runtime;
+ #[allow(
+     dead_code,
+     reason = "The lifecycle module remains crate-visible for tests and pending integrations."
+diff --git a/lib/crates/fabro-workflow/src/pipeline/initialize.rs b/lib/crates/fabro-workflow/src/pipeline/initialize.rs
+index 953f39d2d..3165f7855 100644
+--- a/lib/crates/fabro-workflow/src/pipeline/initialize.rs
++++ b/lib/crates/fabro-workflow/src/pipeline/initialize.rs
+@@ -675,6 +675,7 @@ pub async fn initialize(
+     let engine = Arc::new(EngineServices {
+         run: Arc::clone(&run_services),
+         registry,
++        interviewer: Arc::clone(&options.interviewer),
+         git_state: std::sync::RwLock::new(None),
+         base_env,
+         github_token,
+diff --git a/lib/crates/fabro-workflow/src/services.rs b/lib/crates/fabro-workflow/src/services.rs
+index 72c5499f9..9210998d3 100644
+--- a/lib/crates/fabro-workflow/src/services.rs
++++ b/lib/crates/fabro-workflow/src/services.rs
+@@ -9,6 +9,7 @@ use fabro_auth::CredentialSource;
+ #[cfg(test)]
+ use fabro_auth::ResolvedCredentials;
+ use fabro_hooks::{HookContext, HookDecision, HookExecutionContext, HookRunner};
++use fabro_interview::Interviewer;
+ use fabro_model::{Catalog, ProviderId};
+ use fabro_types::{ManifestPath, RunId};
+ use tokio_util::sync::CancellationToken;
+@@ -16,6 +17,7 @@ use tokio_util::sync::CancellationToken;
+ use crate::event::Emitter;
+ use crate::github_token_source::GitHubTokenSource;
+ use crate::handler::HandlerRegistry;
++use crate::interview_runtime::RunInterviewBlocker;
+ use crate::run_metadata::{RunMetadataRuntime, RunMetadataWriterHandle};
+ use crate::runtime_store::RunStoreHandle;
+ use crate::sandbox_git::GitState;
+@@ -91,19 +93,20 @@ pub struct FabroRunToolServices {
+ /// does NOT count as cancellation.
+ #[derive(Clone)]
+ pub struct RunServices {
+-    pub run_store:               RunStoreHandle,
+-    pub emitter:                 Arc<Emitter>,
+-    pub sandbox:                 Arc<dyn Sandbox>,
+-    pub hook_runner:             Option<Arc<HookRunner>>,
+-    pub locations:               RunLocations,
+-    pub(crate) cancel_token:     CancellationToken,
+-    pub provider_id:             ProviderId,
+-    pub model:                   String,
+-    pub llm_source:              Arc<dyn CredentialSource>,
+-    pub catalog:                 Arc<Catalog>,
+-    pub(crate) sandbox_git:      Arc<SandboxGitRuntime>,
+-    pub(crate) metadata_runtime: Arc<RunMetadataRuntime>,
+-    pub(crate) metadata_writer:  Option<RunMetadataWriterHandle>,
++    pub run_store:                RunStoreHandle,
++    pub emitter:                  Arc<Emitter>,
++    pub sandbox:                  Arc<dyn Sandbox>,
++    pub hook_runner:              Option<Arc<HookRunner>>,
++    pub locations:                RunLocations,
++    pub(crate) cancel_token:      CancellationToken,
++    pub provider_id:              ProviderId,
++    pub model:                    String,
++    pub llm_source:               Arc<dyn CredentialSource>,
++    pub catalog:                  Arc<Catalog>,
++    pub(crate) sandbox_git:       Arc<SandboxGitRuntime>,
++    pub(crate) metadata_runtime:  Arc<RunMetadataRuntime>,
++    pub(crate) metadata_writer:   Option<RunMetadataWriterHandle>,
++    pub(crate) interview_blocker: Arc<RunInterviewBlocker>,
+ }
+ 
+ impl RunServices {
+@@ -137,6 +140,7 @@ impl RunServices {
+             sandbox_git,
+             metadata_runtime,
+             metadata_writer,
++            interview_blocker: Arc::new(RunInterviewBlocker::new()),
+         })
+     }
+ 
+@@ -224,6 +228,7 @@ impl RunServices {
+ pub struct EngineServices {
+     pub run:              Arc<RunServices>,
+     pub registry:         Arc<HandlerRegistry>,
++    pub interviewer:      Arc<dyn Interviewer>,
+     /// Git state for the current run. Set via `set_git_state` at the start of
+     /// `execute` and read by parallel/fan-in handlers.
+     pub(crate) git_state: std::sync::RwLock<Option<Arc<GitState>>>,
+@@ -330,6 +335,7 @@ impl EngineServices {
+                 None,
+             ),
+             registry:        Arc::new(HandlerRegistry::new(Box::new(start::StartHandler))),
++            interviewer:     Arc::new(fabro_interview::AutoApproveInterviewer::engine()),
+             git_state:       std::sync::RwLock::new(None),
+             base_env:        HashMap::new(),
+             github_token:    None,
+diff --git a/lib/crates/fabro-workflow/src/test_support.rs b/lib/crates/fabro-workflow/src/test_support.rs
+index aa9c8117a..43b0c27cf 100644
+--- a/lib/crates/fabro-workflow/src/test_support.rs
++++ b/lib/crates/fabro-workflow/src/test_support.rs
+@@ -7,6 +7,7 @@ use std::time::Duration;
+ use fabro_agent::Sandbox;
+ use fabro_auth::{CredentialSource, EnvCredentialSource};
+ use fabro_graphviz::graph::Graph as GvGraph;
++use fabro_interview::AutoApproveInterviewer;
+ use fabro_model::Catalog;
+ use fabro_store::{ArtifactStore, Database, RunProjection};
+ use object_store::local::LocalFileSystem;
+@@ -175,6 +176,7 @@ async fn initialized(
+                     None,
+                 ),
+                 registry:        Arc::new(registry),
++                interviewer:     Arc::new(AutoApproveInterviewer::engine()),
+                 git_state:       std::sync::RwLock::new(None),
+                 base_env:        options.env,
+                 github_token:    None,
+diff --git a/lib/packages/fabro-api-client/src/.openapi-generator/FILES b/lib/packages/fabro-api-client/src/.openapi-generator/FILES
+index 653eea117..278e89ec9 100644
+--- a/lib/packages/fabro-api-client/src/.openapi-generator/FILES
++++ b/lib/packages/fabro-api-client/src/.openapi-generator/FILES
+@@ -24,7 +24,6 @@ index.ts
+ models/agent-permissions.ts
+ models/aggregate-billing-totals.ts
+ models/aggregate-billing.ts
+-models/api-question-option.ts
+ models/api-question.ts
+ models/append-event-response.ts
+ models/approval-mode.ts
+@@ -410,6 +409,10 @@ models/system-resources-response.ts
+ models/system-run-counts.ts
+ models/timeline-entry-response.ts
+ models/tls-mode.ts
++models/todo-list-kind.ts
++models/todo-list-projection.ts
++models/todo-projection.ts
++models/todo-status.ts
+ models/update-run-parent-request.ts
+ models/update-run-request.ts
+ models/user-response.ts
+@@ -425,4 +428,4 @@ models/workflow-ref.ts
+ models/workflow-reference.ts
+ models/workflow-schedule-summary.ts
+ models/workflow-settings.ts
+-models/write-blob-response.ts
+\ No newline at end of file
++models/write-blob-response.ts
+diff --git a/lib/packages/fabro-api-client/src/models/api-question-option.ts b/lib/packages/fabro-api-client/src/models/api-question-option.ts
+deleted file mode 100644
+index ddbe5577d..000000000
+--- a/lib/packages/fabro-api-client/src/models/api-question-option.ts
++++ /dev/null
+@@ -1,29 +0,0 @@
+-/* tslint:disable */
+-/* eslint-disable */
+-/**
+- * Fabro Run API
+- * HTTP API for managing Fabro workflow run executions.
+- *
+- * The version of the OpenAPI document: 0.1.0
+- *
+- *
+- * NOTE: This class is auto generated by OpenAPI Generator (https://openapi-generator.tech).
+- * https://openapi-generator.tech
+- * Do not edit the class manually.
+- */
+-
+-
+-
+-/**
+- * A selectable option for a multiple-choice or multi-select question.
+- */
+-export interface ApiQuestionOption {
+-    /**
+-     * Machine-readable option key used when submitting an answer.
+-     */
+-    'key': string;
+-    /**
+-     * Human-readable label displayed to the user.
+-     */
+-    'label': string;
+-}
+diff --git a/lib/packages/fabro-api-client/src/models/api-question.ts b/lib/packages/fabro-api-client/src/models/api-question.ts
+index 0cf9e785b..74908ff63 100644
+--- a/lib/packages/fabro-api-client/src/models/api-question.ts
++++ b/lib/packages/fabro-api-client/src/models/api-question.ts
+@@ -15,7 +15,7 @@
+ 
+ // May contain unused imports in some cases
+ // @ts-ignore
+-import type { ApiQuestionOption } from './api-question-option';
++import type { InterviewOption } from './interview-option';
+ // May contain unused imports in some cases
+ // @ts-ignore
+ import type { QuestionType } from './question-type';
+@@ -40,7 +40,7 @@ export interface ApiQuestion {
+     /**
+      * Available options for selection-based questions. Empty for freeform questions.
+      */
+-    'options': Array<ApiQuestionOption>;
++    'options': Array<InterviewOption>;
+     /**
+      * Whether the user may provide freeform text in addition to selecting options.
+      */
+diff --git a/lib/packages/fabro-api-client/src/models/index.ts b/lib/packages/fabro-api-client/src/models/index.ts
+index 23a07a598..c2b6db816 100644
+--- a/lib/packages/fabro-api-client/src/models/index.ts
++++ b/lib/packages/fabro-api-client/src/models/index.ts
+@@ -2,7 +2,6 @@ export * from './agent-permissions';
+ export * from './aggregate-billing';
+ export * from './aggregate-billing-totals';
+ export * from './api-question';
+-export * from './api-question-option';
+ export * from './append-event-response';
+ export * from './approval-mode';
+ export * from './artifact-batch-upload-entry';
+@@ -405,4 +404,4 @@ export * from './workflow-ref';
+ export * from './workflow-reference';
+ export * from './workflow-schedule-summary';
+ export * from './workflow-settings';
+-export * from './write-blob-response';
+\ No newline at end of file
++export * from './write-blob-response';
+diff --git a/lib/packages/fabro-api-client/src/models/interview-option.ts b/lib/packages/fabro-api-client/src/models/interview-option.ts
+index 4215b382e..16c536f80 100644
+--- a/lib/packages/fabro-api-client/src/models/interview-option.ts
++++ b/lib/packages/fabro-api-client/src/models/interview-option.ts
+@@ -18,6 +18,20 @@
+  * Option stored with an interview question in the event log.
+  */
+ export interface InterviewOption {
++    /**
++     * Machine-readable option key used when submitting an answer.
++     */
+     'key': string;
++    /**
++     * Human-readable label displayed to the user.
++     */
+     'label': string;
++    /**
++     * Optional untrusted model-authored option description for display.
++     */
++    'description'?: string | null;
++    /**
++     * Optional untrusted model-authored option preview captured for clients.
++     */
++    'preview'?: string | null;
+ }
+diff --git a/lib/packages/fabro-api-client/src/models/run-projection.ts b/lib/packages/fabro-api-client/src/models/run-projection.ts
+index 0505c0cd0..0ab7b799f 100644
+--- a/lib/packages/fabro-api-client/src/models/run-projection.ts
++++ b/lib/packages/fabro-api-client/src/models/run-projection.ts
+@@ -87,4 +87,4 @@ export interface RunProjection {
+      * Map from StageId (`node_id@visit`) to stage projection data.
+      */
+     'stages': { [key: string]: StageProjection; };
+-}
+\ No newline at end of file
++}
diff --git a/stages/005-implement@1/status.json b/stages/005-implement@1/status.json
new file mode 100644
index 000000000..e178f5f76
--- /dev/null
+++ b/stages/005-implement@1/status.json
@@ -0,0 +1,6 @@
+{
+  "outcome": "succeeded",
+  "notes": "Stage completed: implement",
+  "failure_reason": null,
+  "timestamp": "2026-05-23T17:14:03.543746Z"
+}
\ No newline at end of file
diff --git a/stages/006-simplify_opus@1/prompt.md b/stages/006-simplify_opus@1/prompt.md
new file mode 100644
index 000000000..04e80bb39
--- /dev/null
+++ b/stages/006-simplify_opus@1/prompt.md
@@ -0,0 +1,155 @@
+Goal: # Mid-Stage Agent Interview Tools
+
+## Summary
+
+Add model-native question tools that let agents pause mid-stage and ask the human for input through Fabro's existing interview system.
+
+OpenAI-profile agents get `request_user_input`; Anthropic-profile agents get `AskUserQuestion`. When either tool is called, Fabro creates pending interview questions, surfaces them through the existing web/API/Slack paths, waits for answers, then returns provider-shaped tool results so the model can continue the same stage.
+
+## Key Changes
+
+- Extend the existing interview contract without type sprawl:
+  - Add optional `description` and `preview` fields to the canonical `fabro_types::InterviewOption`; reuse that type through `fabro-api` replacements instead of introducing `AgentQuestionOption`, API-only aliases, or adapter-only duplicate types.
+  - Update OpenAPI, generated Rust/TypeScript clients, event conversion, projection, Slack/web mappers, and the existing `with_replacement("InterviewOption", "fabro_types::InterviewOption", ...)` parity tests.
+  - Treat both fields as untrusted model-authored display data. Store and expose them after enforcing bounded lengths; truncate or reject oversized values consistently before persistence.
+  - Initial UI behavior: display `description` under option labels where practical. Capture and expose `preview`, but do not render preview content specially in web or Slack v1.
+
+- Add a shared run-level interview runtime:
+  - Move the private human-node blocked-state refcount into a reusable run-level guard used by both `HumanHandler` and agent question tools, so `RunUnblocked` is emitted only when all human and agent interviews for the run are resolved.
+  - Runtime accepts the interviewer, workflow emitter, stage scope, stage id, tool call id, and normalized questions.
+  - Support batch asks as a first-class operation: emit/register all questions first, mark the run blocked once, await all answers concurrently, then emit completion/timeout/interrupted events per question and unblock when the batch resolves.
+  - Batch support applies only to multiple `questions[]` inside one question-tool call. Do not aggregate multiple separate question-tool calls from the same model round.
+  - Generate safe internal question IDs with a ULID/UUID plus stage visit/tool-call context; store original model question IDs/text in question metadata for provider result mapping.
+
+- Add provider-specific agent tools:
+  - `request_user_input` for `AgentProfileKind::OpenAi`.
+    - Accept Codex-compatible schema: `questions[]` with `id`, `header`, `question`, and `options[] { label, description }`.
+    - Normalize each question to `QuestionType::MultipleChoice` with `allow_freeform: true`.
+    - Return JSON text matching Codex shape, keyed by the original model question ID: `{"answers":{"id":{"answers":["..."]}}}`.
+  - `AskUserQuestion` for `AgentProfileKind::Anthropic`.
+    - Accept Claude-compatible schema: `questions[]` with `question`, `header`, `options[] { label, description, preview? }`, and `multiSelect`.
+    - Normalize single-select to `MultipleChoice`, multi-select to `MultiSelect`, always with `allow_freeform: true`.
+    - Return Claude-style tool result text keyed by the original question text: `User has answered your questions: "...question..."="answer". You can now continue...`.
+  - Answer formatting for both tools returns user-facing option labels to the model. Preserve internal option keys for validation and event storage. For multi-select, preserve the submission order supplied by the answer path.
+
+- Thread workflow interview context into agent tool execution:
+  - Add an explicit per-turn agent tool runtime context passed into `process_input` or an adjacent `process_input_with_runtime` API. It carries the interviewer, workflow emitter, stage scope/id, shared block guard, and provider answer formatter.
+  - Do not capture stage-specific interview handles in the profile registry or cached session construction; cached full-fidelity sessions must receive the current turn's stage context dynamically.
+  - Child/subagent sessions must not expose these question tools. If somehow called outside the root session, return a model-visible error.
+  - Question tools must execute alone in a model tool round. If a round contains one question tool plus any other tool call, execute the question tool and return model-visible error results for the non-question peers, preserving tool-call/tool-result ordering. If a round contains multiple separate question-tool calls, execute only the first and return model-visible error results for the later question-tool calls instructing the model to combine questions into one `questions[]` batch.
+  - Agent-originated questions have no per-question timeout in v1 because the provider schemas do not include timeout. They rely on existing stage timeout, wall-clock timeout, cancellation, and interruption behavior.
+
+- Preserve existing answer paths:
+  - Do not add a new answer endpoint.
+  - Continue using `GET /runs/{id}/questions` and `POST /runs/{id}/questions/{qid}/answer`.
+  - Keep `ControlInterviewer`, web `InterviewDock`, Slack blocks, and run projection as the delivery mechanism.
+
+## Test Plan
+
+- Unit tests for schema parsing and normalization:
+  - Codex request with descriptions maps to Fabro multiple-choice questions and returns answers by model question ID.
+  - Claude request with `multiSelect: true` maps to `MultiSelect` and returns comma-separated answer text.
+  - Batched Codex and Claude requests surface all questions as pending before awaiting answers, then return one result with every answer mapped to the original model ID/text.
+  - Optional `preview` and `description` survive event, projection, API conversion, OpenAPI replacement tests, and TypeScript client generation.
+  - Oversized `description`/`preview` values are bounded before persistence and never rendered as trusted HTML.
+
+- Workflow and agent tests:
+  - OpenAI-profile session advertises `request_user_input`; Anthropic-profile session advertises `AskUserQuestion`; Gemini advertises neither.
+  - Subagent profiles do not advertise the question tools.
+  - Root agent can ask a question and resume after the answer.
+  - Subagent or missing interview context returns a clear tool error.
+  - Cached full-fidelity session emits interview events against the current stage, not the original cached stage.
+  - A mixed tool round containing a human-question tool plus another tool preserves all required tool results and rejects the peer calls with model-visible errors.
+  - A round with multiple separate question-tool calls executes only the first and rejects later question-tool calls with model-visible errors.
+
+- Server, projection, and UI tests:
+  - `InterviewStarted` with option metadata appears in pending questions.
+  - Submitting valid selected, multi-selected, and freeform answers unblocks the waiting tool.
+  - Duplicate answer submission remains rejected through existing accepted-question logic.
+  - Parallel human gate plus agent question keeps the run blocked until both are answered.
+  - Pause, cancel, and interrupt while an agent question is waiting resolve pending questions consistently and do not leave the run blocked.
+  - Stage timeout or wall-clock timeout while an agent question is waiting interrupts the batch; no per-question timeout event is expected unless a future schema adds timeout.
+  - Slack answer submissions work for agent-originated questions using the same pending interview transport.
+
+- Run checks:
+  - `cargo nextest run -p fabro-interview -p fabro-workflow -p fabro-server -p fabro-agent`
+  - `cd apps/fabro-web && bun test && bun run typecheck`
+  - Regenerate and verify OpenAPI-derived Rust and TypeScript clients after schema changes.
+
+## Assumptions
+
+- This feature is only for in-process/API-backed agent sessions, not ACP external agents in v1.
+- `preview` is stored and exposed but not rendered specially in the first implementation.
+- Human-question tools are available only during root agent execution inside a workflow run with an active interviewer.
+- Existing interview events remain the source of truth for pending questions; no separate agent-question event family is added.
+- The implementation should prefer extending existing interview structs and replacement mappings over adding parallel API DTOs or conversion-only aliases.
+
+
+## Completed stages
+- **toolchain**: succeeded
+  - Script: `command -v cargo >/dev/null || { curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y && sudo ln -sf $HOME/.cargo/bin/* /usr/local/bin/; }; cargo --version 2>&1`
+  - Output:
+    ```
+    cargo 1.95.0 (f2d3ce0bd 2026-03-21)
+    ```
+- **preflight_compile**: succeeded
+  - Script: `cargo check -q --workspace 2>&1`
+  - Output: (empty)
+- **preflight_lint**: succeeded
+  - Script: `cargo +nightly-2026-04-14 clippy -q --workspace --all-targets -- -D warnings 2>&1`
+  - Output: (empty)
+- **implement**: succeeded
+  - Model: gpt-5.5, 483.3k tokens in / 63.1k out
+  - Files: /home/daytona/workspace/fabro/lib/crates/fabro-agent/src/question_tools.rs, /home/daytona/workspace/fabro/lib/crates/fabro-workflow/src/interview_runtime.rs
+
+
+# Simplify: Code Review and Cleanup
+
+Review changes vs. origin for reuse, quality, and efficiency. Fix any issues found.
+
+## Phase 1: Identify Changes
+
+Run git diff (or git diff HEAD if there are staged changes) to see what changed. If there are no git changes, review the most recently modified files that the user mentioned or that you edited earlier in this conversation.
+
+## Phase 2: Launch Three Review Agents in Parallel
+
+Use the Agent tool to launch all three agents concurrently in a single message. Pass each agent the full diff so it has the complete context.
+
+### Agent 1: Code Reuse Review
+
+For each change:
+
+1. Search for existing utilities and helpers that could replace newly written code. Use Grep to find similar patterns elsewhere in the codebase — common locations are utility directories, shared modules, and files adjacent to the changed ones.
+2. Flag any new function that duplicates existing functionality. Suggest the existing function to use instead.
+3. Flag any inline logic that could use an existing utility — hand-rolled string manipulation, manual path handling, custom environment checks, ad-hoc type guards, and similar patterns are common candidates.
+
+Note: This is a greenfield app, so focus on maximizing simplicity and don't worry about changing things to achieve it.
+
+### Agent 2: Code Quality Review
+
+Review the same changes for hacky patterns:
+
+1. Redundant state: state that duplicates existing state, cached values that could be derived, observers/effects that could be direct calls
+2. Parameter sprawl: adding new parameters to a function instead of generalizing or restructuring existing ones
+3. Copy-paste with slight variation: near-duplicate code blocks that should be unified with a shared abstraction
+4. Leaky abstractions: exposing internal details that should be encapsulated, or breaking existing abstraction boundaries
+5. Stringly-typed code: using raw strings where constants, enums (string unions), or branded types already exist in the codebase
+
+Note: This is a greenfield app, so be aggressive in optimizing quality.
+
+### Agent 3: Efficiency Review
+
+Review the same changes for efficiency:
+
+1. Unnecessary work: redundant computations, repeated file reads, duplicate network/API calls, N+1 patterns
+2. Missed concurrency: independent operations run sequentially when they could run in parallel
+3. Hot-path bloat: new blocking work added to startup or per-request/per-render hot paths
+4. Unnecessary existence checks: pre-checking file/resource existence before operating (TOCTOU anti-pattern) — operate directly and handle the error
+5. Memory: unbounded data structures, missing cleanup, event listener leaks
+6. Overly broad operations: reading entire files when only a portion is needed, loading all items when filtering for one
+
+## Phase 3: Fix Issues
+
+Wait for all three agents to complete. Aggregate their findings and fix each issue directly. If a finding is a false positive or not worth addressing, note it and move on — do not argue with the finding, just skip it.
+
+When done, briefly summarize what was fixed (or confirm the code was already clean).
\ No newline at end of file
diff --git a/stages/006-simplify_opus@1/provider_used.json b/stages/006-simplify_opus@1/provider_used.json
new file mode 100644
index 000000000..672cc50e5
--- /dev/null
+++ b/stages/006-simplify_opus@1/provider_used.json
@@ -0,0 +1,5 @@
+{
+  "mode": "agent",
+  "provider": "anthropic",
+  "model": "claude-opus-4-7"
+}
\ No newline at end of file
diff --git a/stages/006-simplify_opus@1/response.md b/stages/006-simplify_opus@1/response.md
new file mode 100644
index 000000000..f229c5535
--- /dev/null
+++ b/stages/006-simplify_opus@1/response.md
@@ -0,0 +1,31 @@
+All clean. Final summary:
+
+## Cleanup Applied
+
+Aggregated findings from three reviews and applied 7 focused fixes (3 files, net −31 lines):
+
+**Efficiency**
+- `bounded_display_field` now truncates in a single `char_indices().nth(...)` pass instead of `chars().count()` + `chars().take()`.
+- Dropped redundant `Ulid::new()` from `internal_question_id` — slug + stage visit + tool-call-id + index is already unique.
+
+**Code quality**
+- De-duped the access-denial / hook-block / normal-execution event triple via shared `emit_tool_call_started` and `emit_tool_call_result` helpers in `tool_execution.rs`, dropping ~25 lines of copy-pasted emit code.
+- Cleaned up dead `unwrap_or_else` fallbacks in `format_anthropic_answers` (infallible string serialization).
+- Extracted repeated "root agent session" error string into `ROOT_SESSION_REQUIRED_ERROR` constant.
+- Renamed `option_label` to `label_for_key`, dropping the redundant `selected_option.filter(key)` defensive check. Preserved the `selected_option`-first lookup for `Selected` (control-protocol path supplies `Selected(key)` with `selected_option: None`).
+- Removed the unused `Default for RunInterviewBlocker` impl.
+- De-duped `RunInterviewGuard::resolve` and `Drop` via a shared `resolve_in_place` method.
+
+**Findings noted but not applied** (with rationale):
+- Lifting `agent_tool_runtime` from a parameter into a task-local set once at the top: large cross-crate refactor; skipped to keep scope bounded.
+- Unifying OpenAI/Anthropic question tools into one schema: schemas intentionally differ (`id`/`header` required for OpenAI, `multiSelect` and `preview` for Anthropic); separation matches provider contracts.
+- `AgentToolRuntime` one-field wrapper: keeping a struct allows future expansion at no extra friction.
+- `register_question_tools` moving into `AgentProfile` trait: cross-crate refactor; current explicit dispatch is fine for two profiles.
+- Cleaning `ControlInterviewer.pending` on agent-batch cancel: needs a new `Interviewer` API; flagged as the most substantial follow-up.
+- `..InterviewOption::default()` adoption in tests: high-churn, low-value compared to other items in this pass.
+- The `mem::forget` suggestion for `RunInterviewGuard::resolve` is buggy — it would leak the `Arc<RunInterviewBlocker>` and `Arc<Emitter>` refs. Used a shared method instead.
+
+**Verification**
+- `cargo nextest run -p fabro-agent -p fabro-workflow -p fabro-interview -p fabro-server`: all tests pass except 3 pre-existing flaky SVG-render tests (confirmed by stashing changes and reproducing the same failures on the pre-cleanup tree).
+- `cargo +nightly-2026-04-14 clippy --workspace --all-targets -- -D warnings`: clean.
+- `cargo +nightly-2026-04-14 fmt --check --all`: clean.
\ No newline at end of file