This merge adds three major features: an eval domain (`eval-store.ts`, `eval-types.ts`) with persistence schema for evaluation data; a plugin dashboard view registry with navigation integration for third-party dashboard extensions; and GitHub source metadata traceability that locks and enforces issu
Fusion-Task-Id: FN-3513
Merges three major changesets: ephemeral agent cleanup for FN-3481 (runtime and spawned agent teardown), a fix for planning-mode refine continuation flow (FN-3209) plus a new local startup script, and chat SSE broadcast isolation with QuickChat backend unification. Key components affected include th
Fusion-Task-Id: FN-3481
Auto-archive previously wiped .fusion/tasks/{id}/ for stale done tasks
even while downstream agents (triage/todo/in-progress/in-review) still
needed to read those sibling specs from disk. Now the sweep skips any
done task that has an active dependent. The executor prompt also
instructs the agent to fall back to fn_task_show when sibling spec
files are missing on disk (e.g., manually archived deps).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Updates the restart integration test mock to account for plugin template behavior, keeping the test in sync with recent plugin template changes.
Fusion-Task-Id: FN-3096
Adds comprehensive tests for cleanup reliability in the in-process runtime (131-line test suite) plus related coverage in the executor tests, with a small fix to the runtime implementation itself to address the reliability issue.
Fusion-Task-Id: FN-3352
The merge completes FN-3320 by implementing configuration tools for the agent system (`agent-tools.ts`), wiring them through the executor, and adding test coverage (`agent-tools-config.test.ts`). Documentation was updated in `AGENTS.md` and the engine tools reference, with minor updates to the agent
Fusion-Task-Id: FN-3320
This merge lands FN-3231 across two steps: it preserves a merge-active fix when verification bounces occur (step 1) and ensures the fix is retained during board routing transitions (step 2). Changes span the dashboard Board routing logic and the engine executor, with corresponding test coverage adde
Fusion-Task-Id: FN-3231
Cleans up lint and type errors in the engine by removing unused imports from executor and merger, and replacing `any` types with proper type annotations in verification-utils.
Fusion-Task-Id: FN-3345
Merges FN-3008 to add a "fallback-used" notification system: the engine now emits events when AI model fallbacks are triggered, dispatches notifications via ntfy/webhook providers, surfaces a session banner in the dashboard, and exposes a settings toggle to enable or disable these alerts.
Fusion-Task-Id: FN-3008
This merge implements a "preserve progress" option for task resets across the system. FN-3185 adds a `preserveProgress` flag to `moveTask` that keeps status/history when resetting tasks back to `todo`, with required explicit confirmation dialogs to prevent accidental resets. The feature is wired thr
Fusion-Task-Id: FN-3185
This release (v0.15.0) brings significant plugin system enhancements including a new dependency graph plugin with dashboard view, plugin skills in session selection, and extended plugin UI slot metadata. Database improvements add SQLite WAL tuning, integrity checks, and batch writes for agent logs.
Fusion-Task-Id: FN-3117
Demote the refresh message from console.error to debugMcp so it no
longer appears as an error in normal output.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This merge adds droid CLI path reconciliation as a new pi extension, wires agent task auto-summarization for the agent tools layer, makes the UsageIndicator component resizable with improved styling, and updates related tests and documentation. The core changes include a new `reconcile-droid-cli-pat
Fusion-Task-Id: FN-3060
engine: reviewer subagents previously kept running after the parent task
was moved out of in-progress, paused, or globally paused — they spawn
their own AgentSession outside `activeSessions`/`activeStepExecutors`,
so the existing kill paths never reached them. Track them in a per-task
`activeSubagentSessions` map (mirrored in TriageProcessor) and dispose
on the same triggers as the main session. ReviewOptions gains
`onSessionCreated` / `onSessionEnded` callbacks the executor and triage
processor wire to register/unregister.
droid-cli: probe timeouts (`validateCliPresence`, `validateCliAuth`,
`runDroidProbe`) raised from 5s to 45s — observed cold-start is ~20s,
so 5s reported the binary as missing even when present. Provider gains
a `FIRST_LINE_TIMEOUT_MS` (60s) cold-start guard so a hung droid binary
is reported with an actionable error instead of being indistinguishable
from a slow-thinking turn. Fix the await race in `streamViaCli`: when
SIGKILL destroys stdout mid-buffer, `rl` may never emit "close", so the
promise also resolves on `proc.close` and forces `rl.close()` — prevents
the engine's "executor did not unwind within 60s — hung subprocess".
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This merge brings multiple substantial features: new research extension tools wired through the AI engine (with full test coverage and documentation), legacy routines agentId backward compatibility with migration paths, migration of experimental remote settings to the global scope, Nerd Font glyph a
Fusion-Task-Id: FN-2996
This merge brings several enhancements across the dashboard and engine. Research settings are now fully integrated into the settings modal and exported for alias builds, including a new schema and dedicated resolver. The AgentDetailView receives a heartbeat markdown viewer modal and expanded styling
Fusion-Task-Id: FN-3029
Adds Nerd Font as the terminal font stack default with a migration to update existing user settings, updates the terminal and settings modals to reflect the new font choice, and includes corresponding tests plus a patch changeset for `@runfusion/fusion`.
Fusion-Task-Id: FN-3026
Workflow-step REVISE retries, pause→todo handoffs, and the
context-overflow fresh-session requeue were all routing tasks back to
`todo` before returning to `in-progress`. The default reopen-to-todo
path reset every step to pending and rewrote PROMPT.md checkboxes, so
each retry restarted from step 0 even when earlier steps had already
been done — the symptom seen on FN-2978, where every workflow REVISE
or pause cycle wiped the task's progress.
- Add `preserveResumeState` to `TaskStore.moveTask`. When set, skip
`resetAllStepsToPending` + `resetPromptCheckboxes` and keep
`worktree` and `executionStartedAt` so the resumed run reattaches to
the same checkout. `status`, `error`, and `blockedBy` still clear.
- Use it on the workflow-rerun bounce, the three pause-graceful
handoffs, and the context-overflow requeue. The agent-terminated
pause path still discards (it nukes worktree+branch by design).
- Context-overflow requeue clears `sessionFile` synchronously in the
awaited `updateTask` immediately before `moveTask`, so the next
dispatch cannot reopen the saturated session via a stale pointer.
- `fn_task_update` no longer silently regresses `done`/`skipped` steps
to `in-progress`, no longer captures a stale rewind checkpoint when
it does, and tells the agent honestly when a regression is ignored.
- Mobile chat keyboard: ChatView/QuickChatFAB gate layout on the new
`keyboardOpen` flag so focused-input + viewport-shrink iOS cases
still adjust when the computed overlap is zero.
Tests:
- New `preserveResumeState` coverage in store.test.ts; updated
workflow-rerun + pause-graceful assertions in executor.test.ts.
- Restructured the previously-flaky "routes exhausted prompt-mode
workflow hard failures" test to drive the bounce inline; passes in
isolation and in the wider workflow/pause/context sweep (59/59).
- Added regression tests in ChatView.test.tsx and QuickChatFAB.test.tsx
for the iOS last-resort `keyboardOpen=true, keyboardOverlap=0` case.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three coordinated fixes for the FN-2978 incident class — auto-requeues
that orphaned committed work and watchdog kills on long verification runs.
**Auto-requeue branch reuse** (executor.ts, worktree-pool.ts)
- executor.ts:1782 now uses `task.branch || fusion/<id>` so persisted
branches are honored on requeue. Previously the hardcoded fallback
always tried to re-create the original branch, hit a conflict with
the prior run's ref, and got suffix -2/-3. Other call sites already
honor task.branch — this aligns the worktree-acquisition path.
- worktree-pool.ts:181 prepareForTask now probes existing branches with
`git rev-parse --verify` and checks them out as-is. Falls through to
suffixed creation only when the branch is genuinely in use by another
live worktree. Previously force-reset with `checkout -B`, destroying
prior commits.
- New private reconcileStepsFromGitHistory walks `git log
baseCommitSha..HEAD` for `feat(FN-X): complete Step N` commits and
marks matching steps[] as done so resumes don't redo committed work.
**Manual reset endpoint + UI** (dashboard)
- POST /api/tasks/:id/reset (requires `confirm: true`) — clears worktree,
branch, all retry counters, resets steps[] to pending, moves to todo.
Distinct from /retry which is the soft-resume path.
- Reset button alongside Retry in TaskDetailModal with confirm dialog,
wired through useTasks → AppModals → API.
**fn_run_verification tool** (run-verification-tool.ts, executor.ts)
- New custom tool wrapping test/lint/build commands with a heartbeat
callback (per-line + 60s synthetic), 200KB head+tail output cap, hard
timeout with SIGTERM→SIGKILL escalation, and auto-bootstrap detection
for missing node_modules. Prevents the inactivity watchdog from
killing sessions during long compiles.
- Cross-platform via `shell: true` (Node picks /bin/sh on POSIX,
cmd.exe on Windows). Prompt section in EXECUTOR_SYSTEM_PROMPT and
EXECUTOR_PROMPT_TEXT instructs agents to prefer package-scoped
verification first and reserve workspace-scoped runs for final
integration.
**Tests** (64 passing)
- detect-pseudo-pause.test.ts (27 tests) — covers all 7 regex patterns,
structural fallback, FN-2978 regression text.
- reconcile-step-regex.test.ts (25 tests) — pins the commit-message
regex against a wide variant set.
- run-verification-command.test.ts (12 tests) — basic execution, output
capture, heartbeat callbacks, timeout, error handling. POSIX-specific
cases (multi-cmd `;`, `>&2`, `\$USER`) gated behind itPosix.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Agents sometimes end turns by asking permission instead of calling tools.
Add turn-ending rules to both executor prompts and runtime detection via
detectPseudoPause() to identify these stalls and issue targeted retry
prompts that correct the behavior.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
FN-2910 surfaced concurrent reviewer + merger activity on the same task.
Root cause: asymmetric in-flight guards let an unpause-resume kick off a
fresh executor session while a recovery path was already running, and the
auto-merge handoff fired before the executor's finally block finished
cleanup. This sweeps the surrounding lifecycle paths for similar races and
tightens the reviewer pause gate against TOCTOU through runtime setup.
- Symmetric in-flight tracking across `executing`, `recoveringCompleted`,
and `resumingUnpaused`; `recoverCompletedTask` bails when any are set.
- Atomic claim of the recovery slot in the completed-task watchdog before
any awaited work.
- Workflow-rerun bounce returns "bounced" | "skipped-pending" so the
watchdog can no longer log a false-success retry when the original
bounce is still mid-flight.
- Self-healing's completed-task scan re-checks executing IDs inside the
loop instead of trusting a pre-await snapshot.
- 300ms grace period before auto-merge enqueue, giving the executor's
finally block (session disposal, child cleanup) time to drain and
eliminating the residual log-overlap symptom from FN-2910. Test uses
fake timers, no real sleep added.
- New AgentSemaphore.runNested for synchronously nested helper agents
(reviewers): bumps activeCount for honest observability while bypassing
the wait queue, preserving forward-progress fairness for the parent at
low maxConcurrent. Both createReviewStepTool and triage's
createReviewSpecTool now use it.
- New beforeSpawnSession hook on AgentRuntimeOptions/AgentOptions fired
inside createFnAgent immediately before createAgentSession, past every
awaited setup step. Reviewer wires a pause re-check that throws a
sentinel error converted to UNAVAILABLE, closing the TOCTOU window
where pause flipped during runtime resolution or resource loading.
All 2887 engine tests pass; engine + core + cli + dashboard + plugin-sdk
+ pi-claude-cli + desktop typecheck clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Audit of merge commit sites surfaced four paths that could produce an
empty commit body when the branch had no unique commits relative to its
base (so `git log <base>..<branch>` returns empty), or when the `git log`
collection itself failed. Subject was always non-empty so git accepted
the commit, but the body was missing — breaking downstream consumers
(release notes, dashboard summaries, mergeDetails) that read it.
The AI merge agent is the primary author of merge commit messages; these
fallback paths only run when the agent didn't commit and the merger has
to commit on its behalf. Previously they used `-m "${commitLog}"` which
silently produced `-m ""` on empty input.
Now uses a 4-tier resolveSafeCommitBody cascade — most informative
first, with a deterministic floor so the function never returns empty
and never throws:
1. The branch's commit log if non-empty.
2. AI-generated body via aiGenerateCommitBody — a fresh readonly
session that summarizes the diff stat into 2–6 bullet points.
Bounded by a 30s timeout (forwards the caller's abort signal too)
so engine pause / shutdown tears it down promptly. Any failure
falls through.
3. The diff stat itself, formatted as a "Files changed" listing.
4. A synthetic `- merge <branch>` placeholder.
Wired into all three merger fallback commit sites:
- Auto-resolved-conflicts commit (Attempt 2 success path)
- -X ours / -X theirs side-strategy commit (Attempt 3)
- Agent-didn't-commit fallback commit (post-AI verification)
Also defensive: removed `--allow-empty-message` from the executor's
squash-import commit. The message is hardcoded non-empty (subject +
body), but the flag was a footgun — git would silently accept an empty
message if the construction ever broke. Switched to two `-m` args
(subject and body separately) so empty would now correctly fail at
git's level rather than silently land a blank-message commit.
Tests + checks: engine 2887/2887 pass, typecheck clean, workspace lint
clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a layered recovery cascade to the merger's pre-rebase stage so tasks
no longer get stuck in in-review when their declared dependency was
squash-merged to main and left orphan raw commits in the dependent's
history. Also prevents the orphan situation at the source for new tasks.
Why:
- 13 tasks were stuck in in-review for hours, all hitting the same
pre-merge rebase abort because they shared 6 raw commits inherited
from FN-2729's branch (declared baseBranch). FN-2729 was then
squash-merged to main, turning those raw commits into orphans whose
content is in main but in a different commit shape, conflicting with
later-merged tasks. The merger's `smart-prefer-main` strategy
correctly refused -X ours (which would silently re-introduce main's
deletions), but the only escape hatch was a 30-min cooldown loop
that retried the same impossible rebase forever.
Recovery cascade (merger.ts pre-rebase stage):
- Layer 1: surgical `git rebase --onto <main> <dep-tip> <branch>` when
task.baseBranch is set. Resolves the dep tip from the live branch ref
or recorded baseCommitSha; peels off the dep's inherited commits
cleanly. Captures the squash-merge-of-dep case end-to-end.
- Layer 2: generic patch-id duplicate-content stripping. Walks the last
500 main commits, computes patch-ids, then drops branch commits whose
patch-id matches and cherry-picks the remainder onto main. Captures
manual cherry-picks, double-merges, and any other duplicate-content
variant Layer 1 doesn't see. Restores the branch's pre-mutation SHA
on partial-failure so worst case leaves the worktree no worse than
before the recovery attempt.
- Layer 3: AI arbitration fall-through. If Layers 1+2 fail, log the
situation and proceed to the existing 3-attempt AI merge cascade
instead of throwing. The deterministic post-merge verification
(test + build) gates whatever the AI produces — that gate is what
enforces prefer-main's safety contract under fall-through (no silent
re-introduction of main's deletions).
- Critical: the unsafe `-X ours` Attempt 3 is suppressed under
fall-through. AI Attempts 1+2 are the only paths that can complete
the merge; if both fail and verification rejects them, the task
bounces back to in-progress via the existing engine path rather than
silently merging.
Prevention (executor.ts worktree creation):
- When a task declares a non-main `baseBranch`, branch the worktree
off main (origin/<defaultBranch> when worktreeRebaseBeforeMerge is
enabled and a remote is resolvable; otherwise local rootDir HEAD)
and `git merge --squash` the dep's content as a single import commit.
The dependent branch then carries main's history + 1 commit instead
of inheriting the dep's raw commits, so a future squash-merge of the
dep produces patch-id-matching content that rebases cleanly.
- Honors settings: respects `worktreeRebaseBeforeMerge`,
`worktreeRebaseRemote`, and falls back to local HEAD when no remote
is resolvable. Fully fail-soft: any squash-import error falls back to
the legacy fork-from-dep behavior so worktree creation still works
for setups where the squash flow can't run.
Engine-side last-retry fix (project-engine.ts):
- Changed conflict-retry condition from `currentRetries < MAX` to
`currentRetries + 1 < MAX` so the bounce-to-in-progress code fires
in the same engine tick as the failing attempt, rather than relying
on a setTimeout-scheduled Nth attempt that dies on engine restart.
Without this, a dev-time engine restart between the 3rd and 4th
retry left the task with mergeRetries=MAX and only the 30-min
cooldown sweep could try again.
Tests:
- New "Layer 1 recovery" test asserts the surgical --onto rebase fires
when baseBranch is set and primary rebase aborts, and that Layer 3
fall-through is NOT triggered when Layer 1 succeeds.
- Updated the "no silent fall-through to -X ours" test to cover the
new fall-through path: even after Layers 1+2 fail and the merge
cascade proceeds, -X ours must not run, and the task log must record
both the Layer 3 fall-through entry and the Attempt 3 suppression.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
On a fresh install useAuthOnboarding's effect ran at mount before the
setup wizard's 500ms auto-open timer fired. The one-shot ref locked,
and the resolved fetch could either stack model onboarding on top of
the wizard or never re-trigger after the wizard closed.
- Gate the trigger on projectId being set so the wizard owns the
bootstrap phase; the auth check only fires once a project exists.
- Re-check setupWizardOpen via a ref when the auth fetch resolves to
avoid stacking onboarding on top of a wizard opened mid-fetch.
- Release the one-shot in that suppressed branch so the effect retries
when the wizard closes.
Adds two regression tests: fresh-install handoff and mid-fetch wizard
suppression.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two related executor fixes:
1. Spawned child agents previously bypassed the executor model lane hierarchy
and used settings.defaultProvider/defaultModelId directly, ignoring
project-level executionProvider/executionModelId from .fusion/config.json.
Resolve via resolveExecutorModelPair() so children honor the same
precedence as the parent executor.
2. Pre-merge workflow step AI calls now have a wall-clock timeout
(settings.workflowStepTimeoutMs, default 6 min) and fall back to the
configured validatorFallback / fallback model on timeout. The 20-min
stuck-detector kill loop was the only escape hatch when a provider's
streaming API hung mid-response, and the kill triggered a same-provider
retry — guaranteeing repeat hangs. The runner now races the prompt against
a timeout; on timeout it disposes the session, logs a clear entry, and
re-runs the step once with a distinct fallback provider/model. If neither
completes (or no fallback is configured), the step returns a normal
failure that flows into the existing handleWorkflowStepFailure retry path.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Stale-merge recovery now calls back into ProjectEngine's auto-merge queue
directly instead of waiting on the 15s polling sweep — wired via a new
InProcessRuntime.setMergeEnqueuer hook so SelfHealingManager can re-enqueue
without leaking engine internals.
createWorktree mirrors the merge-time rebase: when worktreeRebaseBeforeMerge
is enabled, the new task branch is rebased onto <remote>/<defaultBranch>
right after creation, so executors start from origin's tip with local main
replayed on top. Best-effort — fetch/rebase failures abort cleanly and
leave the merge-time rebase as the backstop.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Reset merge metadata, verification counters, and workflow results when tasks move from in-review/done back to in-progress
- Reopen verification-related steps (or the last step fallback) so re-verification runs from a pending state
- Add execute-time guard to clear stale mergeDetails on in-progress tasks before continuing
- Prevent resumeOrphaned fast-path recovery when completed in-progress tasks still carry merge metadata
- Add targeted executor and TaskForm tests covering FN-2883 regression paths
Reviewer subprocesses were spawned via fn_review_spec / fn_review_step
even with globalPause on, because reviewer.ts had no pause awareness.
Stuck detector also kept running, treating pause-disposed sessions as
inactivity and re-queuing tasks. Pause-transition listeners only called
session.dispose(), which doesn't always interrupt an in-flight LLM
stream — letting reviewer spawns leak through after pause flipped.
- reviewer.ts: re-read settings, return UNAVAILABLE without spawning
when globalPause/enginePaused is on.
- stuck-task-detector.ts: skip checkStuckTasks() while paused.
- triage.ts / executor.ts: call session.abort() before dispose() in the
pause-transition listener to interrupt in-flight work.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Capture prompt/completion/total token usage from step-scoped sessions in the step session executor
- Persist per-step token usage in executor run context so stats survive across task execution
- Record single-session token usage totals alongside run context stats logging for consistent aggregation
- Expand executor and step-session executor tests to validate token usage persistence and fixture behavior
Capture per-session token usage from pi-coding-agent's getSessionStats()
after each promptWithFallback in the executor and merger paths, so
task.tokenUsage populates live during runs and reflects final totals on
done tasks. Previously the executor never read session usage and only
the heartbeat path bumped agent token totals, leaving task.tokenUsage
undefined even after completion.
Stats panel and done-card timing also now reflect live state: the modal
overlays the SSE-updated task prop on top of the one-shot fullDetail
snapshot, in-progress workflow steps contribute live elapsed to the
Workflow runtime metric, and the done card uses Timed duration (matching
the stats tab) with workflow runtime as fallback. Time indicator labels
coarsened to <1m / Nm / Nh / Nd.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Fix triage planning model resolution to fall back through project/global planning settings and default overrides
- Fix reviewer model selection to honor validator-specific settings before default provider/model overrides
- Update merger model resolution to apply default override fallback and align shared task setting types/executor flow
- Add regression coverage for triage, reviewer, and merger fallback behavior and update settings hierarchy documentation
- Add isGitRepository utility in worktree-pool using git rev-parse checks
- Fail fast in TaskExecutor with actionable non-git errors before worktree creation starts
- Classify not-a-git-repository worktree add failures as non-retryable in recovery flows
- Warn from in-process runtime startup when the working directory is not a Git repository
- Expand executor and worktree-pool tests to cover non-git, missing-dir, and conflict-classification paths
TUI quit ('q'/Ctrl+C) bypassed signal handlers via process.exit(0), and
neither shutdown path closed the HTTP server, so server.close()'s
stopAllDevServers() listener never ran. In-flight agent bash commands
(spawned detached for their own pgroup) were also never aborted, so
their subprocess trees — including vitest workers — survived as orphans.
Route the TUI quit through SIGINT so the registered shutdown handler
runs, await stopAllDevServers() in both shutdown paths, and abort
in-flight bash on every active agent session at the start of the
runtime drain so killProcessTree reaches every grandchild.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Agents working on a task that depends on other tasks (e.g. documentation
alignment tasks needing the sibling tasks' specs) were repeatedly
rejected by the worktree boundary when reading .fusion/tasks/FN-NNNN/PROMPT.md,
which also contributed to the malformed-tool-result crash we just fixed.
Add a read-only exception to isWorktreeAllowedPath: the read/glob/grep
tools may access .fusion/tasks/*/PROMPT.md and .fusion/tasks/*/task.json
at the project root. Writes and bash cwd remain restricted.
Update the system-prompt boundary docs (executor.ts) so agents know the
exception exists and stop burning turns re-trying rejected reads.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add formatError() helper that extracts both message and full stack from
unknown caught values, and use it at every status:"failed" catch site in
executor, agent-heartbeat, and triage. Stack traces now land in
store.logEntry outcome (persisted to task.log/activityLog) and in stderr
logger output, so failures like "Cannot read properties of undefined
(reading 'filter')" can be diagnosed without re-running.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
