diff --git a/docs-internal/run-directory-keys.md b/docs-internal/run-directory-keys.md
new file mode 100644
index 000000000..41c573c45
--- /dev/null
+++ b/docs-internal/run-directory-keys.md
@@ -0,0 +1,266 @@
+# Run Directory Keys
+
+All keys that may be written to the run store during a workflow execution.
+
+## 1. `_init.json`
+
+Store initialization metadata. Written when the store is created.
+
+- `run_id` — ULID string
+- `created_at` — RFC 3339 timestamp
+- `db_prefix` — SlateDB key prefix
+- `run_dir` — path to run directory (optional)
+
+## 2. `run.json`
+
+Run configuration snapshot. Written at run creation.
+
+- `run_id` — ULID string
+- `created_at` — RFC 3339 timestamp
+- `settings` — full `FabroSettings` object (see `workflow.toml` fields)
+- `graph` — parsed workflow graph
+  - `name` — graph name
+  - `nodes` — map of node id to node object
+    - `id` — node identifier
+    - `attrs` — map of attribute name to typed value (`String`, `Integer`, `Float`, `Boolean`, `Duration`)
+    - `classes` — list of CSS-like class names (optional)
+  - `edges` — list of edge objects
+    - `from` — source node id
+    - `to` — target node id
+    - `attrs` — map of attribute name to typed value
+  - `attrs` — graph-level attributes map
+- `workflow_slug` — workflow slug (optional)
+- `working_directory` — path string
+- `host_repo_path` — original host repo path (optional)
+- `base_branch` — base git branch name (optional)
+- `labels` — string key-value map (optional)
+
+## 3. `start.json`
+
+Start timestamp and git context. Written when execution begins.
+
+- `run_id` — ULID string
+- `start_time` — RFC 3339 timestamp
+- `run_branch` — git branch created for the run (optional)
+- `base_sha` — base commit SHA (optional)
+
+## 4. `checkpoint.json`
+
+Latest execution state. Updated after each node completes.
+
+- `timestamp` — RFC 3339 timestamp
+- `current_node` — id of the node being executed
+- `completed_nodes` — list of completed node ids
+- `node_retries` — map of node id to retry count
+- `context_values` — map of context key to arbitrary JSON value
+- `node_outcomes` — map of node id to outcome object (optional)
+  - `status` — `"success"` | `"fail"` | `"skipped"` | `"partial_success"` | `"retry"`
+  - `preferred_label` — edge label hint for routing (optional)
+  - `suggested_next_ids` — list of suggested successor node ids (optional)
+  - `context_updates` — map of context key to JSON value (optional)
+  - `jump_to_node` — target node for non-edge jump (optional)
+  - `notes` — free-text notes (optional)
+  - `failure` — failure detail object (optional)
+    - `message` — error description
+    - `failure_class` — `"transient_infra"` | `"deterministic"` | `"budget_exhausted"` | `"compilation_loop"` | `"canceled"` | `"structural"`
+    - `failure_signature` — dedup key for repeated failures (optional)
+  - `usage` — token usage object (optional, null when absent)
+    - `model` — model identifier
+    - `input_tokens` — input token count
+    - `output_tokens` — output token count
+    - `cache_read_tokens` — cache read tokens (optional)
+    - `cache_write_tokens` — cache write tokens (optional)
+    - `reasoning_tokens` — reasoning/thinking tokens (optional)
+    - `speed` — speed tier (optional)
+    - `cost` — estimated cost in USD (optional)
+  - `files_touched` — list of file paths modified (optional)
+  - `duration_ms` — stage duration in milliseconds (optional)
+- `next_node_id` — pre-selected next node (optional)
+- `git_commit_sha` — current HEAD SHA (optional)
+- `loop_failure_signatures` — map of failure signature to count (optional)
+- `restart_failure_signatures` — map of failure signature to count (optional)
+- `node_visits` — map of node id to visit count (optional)
+
+## 5. `conclusion.json`
+
+Final run summary. Written when the run finishes.
+
+- `timestamp` — RFC 3339 timestamp
+- `status` — `"success"` | `"fail"` | `"skipped"` | `"partial_success"` | `"retry"`
+- `duration_ms` — total run duration in milliseconds
+- `failure_reason` — error message (optional)
+- `final_git_commit_sha` — final HEAD SHA (optional)
+- `stages` — list of stage summary objects (optional)
+  - `stage_id` — node id
+  - `stage_label` — display label
+  - `duration_ms` — stage duration
+  - `cost` — estimated cost in USD (optional)
+  - `retries` — retry count
+- `total_cost` — aggregate cost in USD (optional)
+- `total_retries` — aggregate retry count
+- `total_input_tokens` — aggregate input tokens
+- `total_output_tokens` — aggregate output tokens
+- `total_cache_read_tokens` — aggregate cache read tokens
+- `total_cache_write_tokens` — aggregate cache write tokens
+- `total_reasoning_tokens` — aggregate reasoning tokens
+- `has_pricing` — whether cost data is available
+
+## 6. `retro.json`
+
+Retrospective analysis. Written after the retro agent completes.
+
+- `run_id` — ULID string
+- `workflow_name` — workflow name
+- `goal` — workflow goal text
+- `timestamp` — RFC 3339 timestamp
+- `smoothness` — `"effortless"` | `"smooth"` | `"bumpy"` | `"struggled"` | `"failed"` (optional)
+- `stages` — list of stage retro objects
+  - `stage_id` — node id
+  - `stage_label` — display label
+  - `status` — status string
+  - `duration_ms` — stage duration
+  - `retries` — retry count
+  - `cost` — estimated cost in USD (optional)
+  - `notes` — free-text notes (optional)
+  - `failure_reason` — error message (optional)
+  - `files_touched` — list of file paths (optional)
+- `stats` — aggregate stats object
+  - `total_duration_ms` — total duration
+  - `total_cost` — aggregate cost (optional)
+  - `total_retries` — aggregate retries
+  - `files_touched` — deduplicated file list (optional)
+  - `stages_completed` — count of completed stages
+  - `stages_failed` — count of failed stages
+- `intent` — what the run intended to do (optional)
+- `outcome` — what actually happened (optional)
+- `learnings` — list of learning objects (optional)
+  - `category` — `"repo"` | `"code"` | `"workflow"` | `"tool"`
+  - `text` — learning description
+- `friction_points` — list of friction point objects (optional)
+  - `kind` — `"retry"` | `"timeout"` | `"wrong_approach"` | `"tool_failure"` | `"ambiguity"`
+  - `description` — friction description
+  - `stage_id` — related node id (optional)
+- `open_items` — list of open item objects (optional)
+  - `kind` — `"tech_debt"` | `"follow_up"` | `"investigation"` | `"test_gap"`
+  - `description` — item description
+
+## 7. `sandbox.json`
+
+Sandbox environment details. Written when the sandbox is ready.
+
+- `provider` — sandbox provider name (e.g. `"local"`, `"docker"`, `"daytona"`)
+- `working_directory` — working directory inside the sandbox
+- `identifier` — sandbox instance identifier (optional)
+- `host_working_directory` — host-side working directory (optional)
+- `container_mount_point` — mount point inside the container (optional)
+
+## 8. `workflow.fabro`
+
+Raw Graphviz dot source for the workflow graph. Plain text, not JSON.
+
+## 9. `workflow.toml`
+
+Workflow configuration in TOML format. Same schema as `settings` in `run.json`.
+
+## 10. `checkpoints/{seq:04}-{epoch_ms}.json`
+
+Checkpoint history snapshots. Same schema as `checkpoint.json` (#4). Filename is a zero-padded sequence number followed by epoch milliseconds (e.g. `0042-1706234567890.json`).
+
+## 11. `nodes/{node_id}/prompt.md`
+
+Prompt sent to the LLM for agent or prompt nodes. Plain text/markdown, not JSON.
+
+## 12. `nodes/{node_id}/response.md`
+
+Response received from the LLM. Plain text/markdown, not JSON.
+
+## 13. `nodes/{node_id}/status.json`
+
+Node execution status. Written when a node completes.
+
+- `status` — `"success"` | `"fail"` | `"skipped"` | `"partial_success"` | `"retry"`
+- `notes` — free-text notes (optional)
+- `failure_reason` — error message (optional)
+- `timestamp` — RFC 3339 timestamp
+
+## 14. `nodes/{node_id}/stdout.log`
+
+Standard output from command nodes. Plain text, not JSON.
+
+## 15. `nodes/{node_id}/stderr.log`
+
+Standard error from command nodes. Plain text, not JSON.
+
+## 16. `nodes/{node_id}/cli_stdout.log`
+
+Standard output from CLI-backend LLM invocations. Plain text, not JSON.
+
+## 17. `nodes/{node_id}/cli_stderr.log`
+
+Standard error from CLI-backend LLM invocations. Plain text, not JSON.
+
+## 18. `nodes/{node_id}/diff.patch`
+
+Git diff of sandbox changes made by the node. Plain text unified diff, not JSON.
+
+## 19. `nodes/{node_id}/provider_used.json`
+
+LLM provider metadata. Written for agent, prompt, and CLI-backend nodes.
+
+- `mode` — `"agent"` | `"prompt"` | `"cli"`
+- `provider` — provider name (e.g. `"anthropic"`, `"openai"`, `"gemini"`)
+- `model` — model identifier
+- `command` — CLI command string (only present when `mode` is `"cli"`)
+
+## 20. `nodes/{node_id}/script_invocation.json`
+
+Command node invocation details. Written before the command runs.
+
+- `command` — shell command or script body
+- `language` — `"shell"` | `"python"`
+- `timeout_ms` — timeout in milliseconds (null if no timeout)
+
+## 21. `nodes/{node_id}/script_timing.json`
+
+Command node execution timing. Written after the command completes.
+
+- `duration_ms` — execution duration in milliseconds
+- `exit_code` — process exit code (null if timed out)
+- `timed_out` — whether the command was killed by timeout
+
+## 22. `nodes/{node_id}/parallel_results.json`
+
+Results from parallel branch execution. Written by the parallel handler.
+
+Array of objects, each with:
+- `id` — branch node id
+- `status` — status string (e.g. `"success"`, `"fail"`)
+- `head_sha` — git HEAD SHA from the branch worktree (optional, present only if set)
+
+## 23. `retro/prompt.md`
+
+Prompt sent to the retro agent. Plain text/markdown, not JSON.
+
+## 24. `retro/response.md`
+
+Response received from the retro agent. Plain text/markdown, not JSON.
+
+## 25. `retro/status.json`
+
+Retro agent execution status.
+
+- `outcome` — `"success"` | `"failure"`
+- `failure_reason` — error message (null on success)
+- `timestamp` — RFC 3339 timestamp
+
+## 26. `retro/provider_used.json`
+
+Retro agent LLM provider metadata.
+
+- `mode` — always `"agent"`
+- `provider` — provider name (e.g. `"anthropic"`, `"openai"`)
+- `model` — model identifier
+
+---
+
+**Node visit directories:** The first visit writes to `nodes/{node_id}/`. Subsequent visits write to `nodes/{node_id}-visit_{N}/` where N is the visit number.