## Summary
Removes the standalone `agent.context_window.snapshot` event and instead
attaches the context-window projection directly to `agent.message`. This
eliminates the async provider token-count API calls that the old
approach required, and simplifies the event log to a single event type
carrying all post-response agent data.
## What Changed and Why
**Before:** After each LLM turn, the agent emitted a separate
`agent.context_window.snapshot` event — first a local estimate, then
potentially a second one after an async `count_input_tokens` call
resolved (or after response usage arrived). This required fingerprint
deduplication state, a `close_token` to cancel in-flight counts, and
frontend handling for the extra event type.
**After:** The `AgentEvent::AssistantMessage` variant carries an
`Option<StageContextWindowProjection>`. The projection is computed
locally at request-build time and then refined using response token
usage when available (`ResponseUsageScaledBreakdown`), or kept as a
`LocalEstimate` when response usage is absent. No provider API calls are
made.
### Plan Summary
- **Task 1:** Added `context_window:
Option<StageContextWindowProjection>` to `AgentMessageProps` (Rust types
+ OpenAPI), removed `AgentContextWindowSnapshotProps` and
`EventBody::AgentContextWindowSnapshot`.
- **Task 2:** Removed the spawned `count_input_tokens` task,
`close_token`, fingerprint sets, and both snapshot-emit methods from
`Session`. Added `context_window_from_response_usage` to
`context_window.rs`; `BuiltRequest` now holds the local projection
instead of the tool list.
- **Task 3:** Workflow conversion copies `context_window` from
`AgentEvent::AssistantMessage` into `AgentMessageProps`; store reducer
reads it from `AgentMessage` instead of the removed snapshot variant and
stamps `event_seq`.
- **Task 4:** GET endpoint tests updated to seed data via
`agent.message` with embedded context-window; endpoint behavior
unchanged.
- **Task 5:** Frontend constant and tests for
`agent.context_window.snapshot` removed; `agent.message` already
invalidates `stageContextWindow` through existing stage-activity
handling. TypeScript client regenerated with the new `AgentMessageProps`
model.
### Key Design Decisions
- **No provider token-count API calls** during normal execution —
context-window accuracy relies on local estimates scaled by response
usage, which is always available for successful turns.
- **Failed-before-response turns** emit no context-window data
(`context_window: None`), matching the old behavior where a snapshot
would have been emitted but response-usage scaling would never arrive.
- `BuiltRequest` drops the `tools` field (only needed for the
now-removed snapshot emission path); the local projection is computed at
build time and stored directly.
### Fabro Details
<details>
<summary>Ran 8 stages in 60m 3s for $55.78</summary>
| Stage | Duration | Cost | Retries |
|---|---|---|---|
| start | 0s | – | 0 |
| toolchain | 1s | – | 0 |
| preflight_compile | 2m 1s | – | 0 |
| preflight_lint | 2m 16s | – | 0 |
| implement | 30m 39s | $44.82 | 0 |
| simplify_opus | 10m 48s | $4.03 | 0 |
| simplify_gpt | 5m 1s | $6.93 | 0 |
| verify | 8m 47s | – | 0 |
| **Total** | **60m 3s** | **$55.78** | **0** |
</details>
<details>
<summary>Ran <code>ImplementPlan.fabro</code> (11 nodes and 14
edges)</summary>
```dot
digraph ImplementPlan {
graph [
goal="Implement and simplify",
model_stylesheet="
* { model: claude-opus-4-7; }
"
]
rankdir=LR
start [shape=Mdiamond, label="Start"]
exit [shape=Msquare, label="Exit"]
toolchain [label="Toolchain", shape=parallelogram, 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", max_retries=0]
preflight_compile [label="Preflight Compile", shape=parallelogram, script="cargo check -q --workspace 2>&1", max_retries=0]
preflight_lint [label="Preflight Lint", shape=parallelogram, script="cargo +nightly-2026-04-14 clippy -q --workspace --all-targets -- -D warnings 2>&1", max_retries=0]
fix_lints [label="Fix Lints", prompt="The preflight lint step failed. Read the build output from context and fix all clippy lint warnings.", max_visits=3]
implement [label="Implement", prompt="Read the plan file referenced in the goal and implement every step. Make all the code changes described in the plan. Use red/green TDD.", model="gpt-55", reasoning_effort="xhigh"]
simplify_opus [label="Simplify (Opus)", prompt="@prompts/simplify.md"]
simplify_gpt [label="Simplify (GPT-55)", prompt="@prompts/simplify.md", model="gpt-55"]
verify [label="Verify", shape=parallelogram, script="git fetch origin main 2>&1 && git merge --no-edit --no-stat origin/main 2>&1 && cargo +nightly-2026-04-14 fmt --all 2>&1 && cargo dev docs refresh 2>&1 && cargo +nightly-2026-04-14 fmt --check --all 2>&1 && { command -v rg >/dev/null 2>&1 || { echo 'rg is required for verify'; exit 127; }; } && ! rg -n 'AuthMode::Disabled|RunAuthMethod|RunSubjectProvenance|\bActorRef\b|\bActorKind\b|AuthenticatedSubject|AuthenticatedService|AuthorizeRunScoped|AuthorizeRunBlob|AuthorizeStageArtifact|AuthorizeCommandLog|auth_method\s*==\s*\"disabled\"' lib/crates apps lib/packages docs/public/api-reference/fabro-api.yaml 2>&1 && cargo +nightly-2026-04-14 clippy --workspace --all-targets -- -D warnings 2>&1 && cargo nextest run --workspace --status-level slow --profile ci 2>&1 && cargo dev docs check 2>&1 && bun install --frozen-lockfile 2>&1 && (cd apps/fabro-web && bun run typecheck) 2>&1 && (cd apps/fabro-web && bun run test) 2>&1 && (cd lib/packages/fabro-api-client && bun run typecheck) 2>&1 && cargo dev build -- -p fabro-cli --release 2>&1", goal_gate=true, retry_target="fixup"]
fixup [label="Fixup", prompt="The verify step failed. Read the build output from context and fix all format, clippy, Rust test, docs, TypeScript typecheck/test, and build failures.", max_visits=3]
start -> toolchain
toolchain -> preflight_compile [condition="outcome=succeeded"]
toolchain -> exit
preflight_compile -> preflight_lint [condition="outcome=succeeded"]
preflight_compile -> exit
preflight_lint -> implement [condition="outcome=succeeded"]
preflight_lint -> fix_lints
fix_lints -> preflight_lint
implement -> simplify_opus -> simplify_gpt -> verify
verify -> exit [condition="outcome=succeeded"]
verify -> fixup
fixup -> verify
}
```
</details>
⚒️ Generated with [Fabro](https://fabro.sh)
---------
Co-authored-by: Fabro <noreply@fabro.sh>
Co-authored-by: Bryan Helmkamp <bryan@brynary.com>
## Summary
Adds support for MCP servers that use the SSE-based HTTP transport,
including Playwright MCP.
Fabro already supports stdio and Streamable HTTP MCP servers. Some MCP
servers still expose the SSE transport shape where the client opens an
SSE stream, receives an `endpoint` event, and sends JSON-RPC requests
back to that endpoint. This PR adds an explicit `protocol = "sse"`
option while keeping Streamable HTTP as the default.
## What Changed
- Added `McpHttpProtocol` with `streamable_http` as the default and
`sse` as an opt-in protocol.
- Added an SSE MCP client transport implementation.
- Wired HTTP MCP setup to choose Streamable HTTP or SSE based on config.
- Added `protocol = "sse"` support for both `http` and `sandbox` MCP
entries.
- Updated sandbox MCP resolution so SSE sandbox servers connect through
the preview `/sse` path.
- Documented `protocol = "sse"` for Playwright MCP.
- Added an integration test covering SSE initialize, tool listing, and
tool calls.
## Example
```toml
[run.agent.mcps.playwright]
type = "sandbox"
protocol = "sse"
command = ["npx", "@playwright/mcp@latest", "--port", "3100", "--headless", "--browser", "chromium"]
port = 3100
startup_timeout = "60s"
tool_timeout = "2m"
```
## Compatibility
Existing MCP configs are unchanged because `protocol` defaults to
`streamable_http`.
## Validation
- `cargo +nightly-2026-04-14 fmt --check --all`
- `cargo nextest run -p fabro-mcp`
- `cargo check -p fabro-agent -p fabro-workflow -p fabro-config`
---------
Co-authored-by: Bryan Helmkamp <bryan@brynary.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Clicking a column header on /runs?view=list silently did nothing.
handleSortClick called updateParam three times in a row; each call
cloned a fresh URLSearchParams from the same closure-captured
searchParams and invoked setSearchParams independently. React Router's
setSearchParams calls don't merge in a single tick, so only the last
one's params landed in the URL -- the sort change was overwritten by
the trailing page-reset. setPageSize had the same shape and was also
quietly losing the size change.
Replace the per-key URL mutator with a reducer-shaped
updatePreferences((prev) => next) that operates on the typed
RunsWorkspacePreferences model. URL and localStorage are derived from
the same next object via the existing converters, and each handler
makes exactly one call -- so concurrent-update races are
structurally impossible. Use setSearchParams((prev) => ...) so the
updater reads the latest committed URL params instead of a closure.
Add page to RunsWorkspacePreferences so the model describes the full
URL view state; strip it before persisting to localStorage since page
is ephemeral. Rename persistRunsWorkspaceSearchParams to
persistRunsWorkspacePreferences to match what it now consumes. Guard
the hydration useEffect with a useRef so it runs only on mount.
Add a regression test that clicks a SortHeader and asserts the URL
gains sort=status while preserving view=list&archived=1, and that a
second click toggles direction=asc.
The local server wasn't setting git_root, and wasn't defaulting to the
workspace working directory, so it didn't find project-specific skills.
---------
Co-authored-by: Bryan Helmkamp <bryan@brynary.com>
Fixes the flaky `attach_json_errors_without_prompting_for_human_input`
snapshot by moving elapsed JSON duration redaction into the shared
`fabro-test` snapshot filters. `duration_ms`, `wall_time_ms`,
`inference_time_ms`, `tool_time_ms`, and `active_time_ms` now use one
common normalization path, and the attach, wait, and events integration
snapshots use that shared helper instead of hand-rolled per-test
regexes.
This keeps snapshots focused on event shape and command behavior rather
than exact runtime timing, while leaving exact timing relationships to
direct assertions in lower-level tests.
Verified with `cargo nextest run -p fabro-test`, `cargo nextest run -p
fabro-cli --test it cmd::attach`, `cargo nextest run -p fabro-cli --test
it cmd::wait`, `cargo nextest run -p fabro-cli --test it cmd::events`,
`cargo +nightly-2026-04-14 fmt --check --all`, and `cargo
+nightly-2026-04-14 clippy -p fabro-test -p fabro-cli --test it -- -D
warnings`.
---
[](https://github.com/EveryInc/compound-engineering-plugin)
🤖 Generated with GPT-5 (context unknown, reasoning effort unknown) via
[Codex](https://openai.com/codex)
- Add /automations/new with Basics/Source/Goal/Triggers panels and a
kebab-case Slug auto-derived from Name until the user edits it
- Wire the "Create Automation" button on /automations to the new page
- Move individual automation URL from /automations/<slug> to
/automation/<slug>; /automations and /automations/new are unchanged
- Refresh /settings/secrets/new to use the Panel + Row layout pattern:
breadcrumb header, one field per row, plain footer, no inside-Panel
stacked form
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The lifecycle status pill next to the run title now hides when the run
is in the initializing column, mirroring the board view behavior. This
removes the duplicate "Initializing" / "STARTING" indicators that
appeared together on the row during startup.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The Connect menu (Preview / SSH) on the run detail header was wired up
only in demo mode and the items were never connected to real actions.
Drop the dead UI.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The Pending, Skipped, and Cancelled status pills used `text-fg-muted`
(#4B5768) on `bg-overlay-strong`, producing ~1.7:1 contrast against the
panel — well below WCAG AA. Switch to `text-fg-3` (#A8B5C5) for ~6:1
while keeping the subdued look that distinguishes these states from
active/result tones.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Reuses the StagePopover from the stages sidebar so graph nodes reveal
the same handler, timing, model, and status-specific detail on hover.
The graph is server-rendered SVG injected via innerHTML, so listeners
are attached imperatively alongside the existing click handlers; the
popover is portal-positioned via the shared hoverCardStyle helper.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Hovering a stage row reveals handler, timing, model, and status-specific
detail — failure reason for failed/retrying, notes for skipped/partial,
tokens and files-touched for succeeded. Lazy-fetches per-stage events
on first hover via the existing useRunStageEvents hook; HoverCard gains
an openDelay so a cursor sweep doesn't trigger fetches for every row.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two icon buttons on the stage events toolbar, to the right of the
model usage label: copy all loaded events as pretty-printed JSON, or
download them as JSONL. Both read from the existing SWR cache, so no
new API surface — for in-flight stages they're a snapshot of what the
client has fetched so far.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replace the native browser title tooltip on the model chip with a
structured HoverCard listing provider, model, reasoning, and speed.
Shorten the chip label to `model[effort]` (e.g. `gpt-5.5[xhigh]`).
Search expands on focus with a width transition and collapses on blur
when empty. Ghost icon style when collapsed; full input styling slides
in on expand.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
On /settings/models, unconfigured providers now offer "Add secret →"
alongside "Get API key →", deep-linking to /settings/secrets/new with
the expected vault secret name prefilled. Driven by a new
`expected_secret_name` field on the Provider API, derived from the
first vault credential in the catalog so the suggestion stays in sync
with the catalog instead of being hardcoded on the frontend.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Surfaces the existing summary.size (XS/S/M/L/XL) as a compact bold
badge to the right of the last-event chip, with a tooltip showing
the underlying billed cost. Hidden when no billing data is available
so we don't show a misleading "XS" for runs with zero cost.
Wires the existing POST /api/v1/runs/delete endpoint into the runs list
selection toolbar. Surfaces a confirmation dialog before calling the
fail-soft batch delete, since deletion is irreversible. Only archived
runs are eligible, matching the single-run delete semantics.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
## Summary
Three small UX tweaks to the agent stage insights sidebar so it reads as
informational rather than alarming.
- **Permission badge stays neutral.** Removed `text-coral` (red) from
Full and `text-amber` (orange) from Read/write — every level now sits in
the foreground palette (`fg-2` / `fg-3`). Icon shape (lock / pencil /
bolt) carries the level distinction and the badge label spells it out.
- **Collapsed footer always uses the muted lock icon.** The footer is a
static affordance, not a danger signal, so a Full-access stage no longer
splashes a colored icon in the corner of the page.
- **Hide the Todos section when there are zero todos.** No header row,
no `0/0` count, no "No todos." line — saves vertical space on stages
where the agent never used TodoWrite.
## Test plan
- [x] `bun run typecheck` (apps/fabro-web)
- [x] `bun test app/components/stage-insights-sidebar.test.tsx` (8/8
pass)
- [ ] Visually confirm in a browser: Full-access agent stage shows a
neutral bolt + "Full access" label (no red); collapsed sidebar footer
shows a single muted lock regardless of level; stage with zero todos has
no Todos section.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Track which MCP servers the agent invoked. New `invoked: bool` on
`McpServerProjection` (OpenAPI + Rust type + generated TS client),
set by the projector when an `AgentToolStarted` event has an
`mcp__<server>__*` tool_name. UI shows `used/total` in the section
header, replaces the tool count with `used` on invoked rows, and dims
rows that weren't invoked. Sticky across status re-reads.
- Quiet noisy context-window warnings. When the snapshot's total is
provider-authoritative (ProviderApiScaledBreakdown or
ResponseUsageScaledBreakdown), drop local-estimator warning codes
from the snapshot — they imply the user-facing total is wrong when
it isn't. Also dedupe by code so a 35-turn conversation with opaque
reasoning blocks no longer surfaces 35 copies of the same warning.
- Reword the legitimately-local warnings. "opaque provider context
estimated from JSON" → "Some content couldn't be precisely
tokenized; total is approximate." Same treatment for the media,
provider-options, and whole-request local-estimate messages.
- Rename the sidebar header from "INSIGHTS" to "AGENT" to better
describe what it shows.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Renders a second left sidebar on /runs/<id>/stages/<agent-stage> with
todos, color-coded context-window usage and breakdown, skills, MCP
servers, and permission level. Data comes from the existing
StageProjection and the context-window endpoint added in #378; no API
changes.
Also set permission_level to Full on workflow agent SessionOptions —
workflow agents run with no tool_access_policy and expose the full
tool registry, so Full is the honest report and avoids "Unknown"
rendering in the new sidebar.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The Runs nav link goes to /runs (no query string), so the route briefly
rendered with default params (view=columns, archived=false) before a
post-commit useEffect restored the URL from stored preferences. On
repeat clicks the useAllRuns SWR cache for {includeArchived:false}
returned zero rows immediately, flashing the Quick Start landing for
users whose only runs are archived.
Resolve workspace search params synchronously during render via
resolveRunsWorkspaceSearchParams(), so the first frame already reflects
stored prefs. The effect now just writes the URL back to match.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
## Summary
Adds a black-box server API regression test for the per-run event append
race observed as projection cache sequence gaps and non-contiguous event
streams.
The test creates a run through the public API, simulates a server
restart with a fresh AppState over the same backing object store, then
concurrently appends events through `POST /api/v1/runs/{id}/events`. The
expected contract is that all appends succeed and the resulting event
sequence is durable and contiguous.
## Test Plan
- `cargo nextest run -p fabro-server --features test-support --test it
concurrent_event_appends_after_restart_keep_projection_cache_contiguous`
Expected current result: fails RED, demonstrating the existing race.
Example observed failure showed duplicate/missing seqs such as `[1, 2,
3, 3, 4, 4, ...]` instead of a contiguous sequence through 66.
Fixes implement-plan runs failing at verify time when cloned sandboxes
lack Git identity, and prevents the verify forbidden-pattern scan from
being silently skipped when `rg` is unavailable.
## Changes
- Install `ripgrep` in the Fabro Daytona image and bump the snapshot ref
to `fabro-v12` so Daytona rebuilds it.
- Configure repository-local Git `user.name` and `user.email` from
`run.git.author` during workflow initialization before lifecycle setup
commands or workflow stages run.
- Make the implement-plan verify stage fail explicitly if `rg` is
missing.
## Validation
- `cargo test -p fabro-workflow
configure_sandbox_git_identity_uses_run_author --quiet`
- `cargo +nightly-2026-04-14 fmt --check --all`
- `cargo +nightly-2026-04-14 clippy -p fabro-workflow --all-targets --
-D warnings`
- `cargo run -p fabro-cli -- validate
.fabro/workflows/implement-plan/workflow.fabro`
---
[](https://github.com/EveryInc/compound-engineering-plugin)
🤖 Generated with GPT-5 via [Codex](https://openai.com/codex)
## Summary
- Render repeated stage visits as `verify@2` in the sidebar (and
waterfall/header/artifacts) to match Fabro's stage-reference syntax
instead of the parenthesized `verify (2)` form.
- Persist runs workspace preferences across sessions.
- Add multi-select with bulk archive/unarchive on the runs list view.
- Show provider logos on settings/models and integration logos on
settings/integrations, with a slightly wider logo-to-text gap.
## Test plan
- [ ] `cd apps/fabro-web && bun test` passes.
- [ ] Sidebar shows `verify`, `verify@2`, `verify@3` for a looped node
on a run with multiple visits.
- [ ] Runs list: select multiple runs and bulk archive/unarchive.
- [ ] Workspace preference on the runs page persists after reload.
- [ ] Settings → Models and Settings → Integrations render
provider/integration logos with the new spacing.
Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Bumps the flex gap from 12px to 16px so the logo chip and the name/help block
breathe a little more.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Vendor brand icons for GitHub, Slack, Microsoft Teams, Discord, Linear, and
Jira under apps/fabro-web/public/images/integrations/ and render each one in
the same light chip used on the providers page. Refactors the existing rows
into IntegrationRow so name + help sit alongside the logo.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Vendor the 9 SVGs from models.dev under apps/fabro-web/public/images/providers/
and render each one in a light chip alongside the provider name and model count.
LiteLLM has no logo on models.dev; the onError handler falls back to an initial
in the same chip style.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a leading checkbox column with a tri-state select-all header and
a fixed bottom action toolbar that surfaces selection count plus
Archive and Unarchive buttons. Selection clears when pagination,
sort, or filters change. Bulk actions fan out to the existing single-
run endpoints via Promise.allSettled and report per-run outcomes.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
