- db.ts: restrict migration-105 orphan-step cleanup to JSON arrays
(json_type guard so json_each can't expand objects/strings)
- project-engine.ts: requestInterpreterMerge throws on null task lookup
instead of casting null into MergeResult (seam converts to clean failure)
- executor.ts: truncate dual-observe shadow stage walk at the live terminal
stage so healthy in-review tasks don't record a phantom merge transition
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
maybeObserveWorkflowParity compares the selected workflow graph's routing
against the legacy authoritative run for the same task and records
workflow:parity-observed / -drift audit events (aggregated by
getWorkflowParitySummary). The shadow walks the graph via WorkflowGraphTaskRunner
with no-side-effect seams whose outcomes mirror the legacy task's reality
(execute reached review?, review verdict, merged?), then maps visited seam nodes
to execute/review/merge stages. Hooked in handoffTaskToReview, gated by
workflowInterpreterDualObserve (off by default), wrapped so it never affects the
authoritative handoff. Scope: execute->review->merge routing parity; full
execution-fidelity shadow is future graduation work.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Greptile + CodeRabbit findings across core/engine/dashboard. Stale findings
(written against earlier commits) verified and skipped; valid ones fixed.
Engine:
- await-input: do not clear pausedReason in the /input route (the node's
marker must survive unpause); the node clears it after consuming input.
Embed a colon-free epoch watermark in the marker so only post-pause steering
comments count as the reply (ISO timestamps collided with the colon
separator and the dashboard question parser).
- gate nodes without a registered runner now fail closed (throw) instead of
silently passing.
- a thrown interpreter error in maybeExecuteWorkflowGraph now falls back to the
legacy pipeline instead of stranding the task in-progress.
- approved-CLI path clears the stale awaiting-cli-approval status/marker.
Core:
- persist+cascade workflow selection: purge task_workflow_selection rows and
compiled workflow_steps on physical task deletes; migration 105 cleans
already-orphaned rows; catch-cleanup for materialized steps when the owner
write fails; WF-id allocation now in a BEGIN IMMEDIATE transaction.
- compiler validates the canonical execute->review->merge seam order (rejects
duplicate/misordered seams).
- disk-backed reopen round-trip + tightened updatedAt/list assertions.
Dashboard:
- WorkflowSelector clears stale default/options across project changes and on
fetch failure; InlineCreateCard/NewTaskModal reset the workflow on all
clear/discard paths and include it in dirty-state.
- WorkflowNodeEditor: config-key deletion now persists; removed an invalid
eslint-disable that was itself a hard lint error.
- TaskCard: single status badge for awaiting-input (no duplicate).
- WorkflowResultsTab: reset paused-action UI between pauses; surface
resume/approve failures inline.
- TaskDetailModal: treat awaiting-user-input/awaiting-cli-approval/paused as
not-in-progress for the live-log subscription.
- workflow-flow-mapping: don't write synthetic node names back into IR.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Agent-native parity: users can list workflows and select one for a task in
the dashboard, so agents should be able to as well. Adds two task-session
tools — fn_workflow_list (read: built-ins + user definitions) and
fn_workflow_select (assign a workflow to a task, defaulting to the current
one) — wired into the executor's customTools and exported from the engine
index. Covered by unit tests against mock stores.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The node "Auto-approve requests" toggle was captured but unused. The only
human-approval pause reachable from a custom node is the CLI first-run
trust-on-first-use gate (review-style nodes run as ephemeral readonly agents
with no permission gate), so autoApprove now bypasses that pause — a superset
of the CLI-specific cliSkipApproval flag. The inspector explains the effect
when the toggle is on.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
In-review tasks showed other tasks' files in their "files changed"
list. Task branches fork from local main, but the base capture
measured merge-base(HEAD, origin/main) — when local main carried
merged-but-unpushed task commits, the recorded base rewound past
them, and after the post-merge rebase-and-push rewrote those SHAs,
baseCommitSha..HEAD permanently swept the predecessors' files into
the new task's diff.
Extract the capture into base-commit-capture.ts, measure local main
first (origin/main fallback) to match the contamination-base sites,
and add a real-git regression suite covering local-ahead-of-origin.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Address PR #1363 review findings:
- core: pausedReason was written in-memory and read by SELECT but never
persisted by the task upsert (missing column/value) nor mapped back in
rowToTask — so it was lost on every reload. Add it to both. This is the
root cause behind the workflow CLI-approval / await-input pause cycle and
also fixes token-budget / worktrunk pause reasons silently vanishing.
- dashboard: approve-cli now derives the approved command exclusively from
the task's pausedReason; a caller-supplied body.command is ignored, closing
a trust-on-first-use bypass.
- engine: await-input nodes resume only when THIS node paused the task (its
marker on pausedReason), not on any pre-existing steering comment.
- engine: write-capable custom nodes (coding/script/CLI) are refused until a
task worktree exists, so they never mutate the shared repo root before the
execute seam.
- engine: document cliSkipApproval as an intentional workflow-author-only
escape hatch; scriptName is now const (ESLint).
- tests: pausedReason round-trip coverage in store-persistence; approve-cli
body-command-ignored + no-pending-command coverage; built-in-aware list
assertion in workflow-routes.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A CLI node can set cliSkipApproval to bypass the trust-on-first-use pause and
run its command immediately. Exposed as a checkbox in the node inspector.
CLI prompt nodes now accept a raw cliCommand (any command + args), not just
named scripts. A raw command must be explicitly approved by the user before it
runs: an unapproved command pauses the task (status awaiting-cli-approval) with
the command shown; the user approves via POST /tasks/:id/workflow/approve-cli,
which records the exact command string in settings.approvedWorkflowCliCommands
and resumes. Named scripts (settings.scripts) still never require approval.
Adds POST /tasks/:id/workflow/input to answer await-input nodes (records a
steering comment + resumes).
Prompt nodes now support an executor kind: model (provider/model override),
agent (adopts a named agent's model + custom instructions), skill (prompt
becomes a skill invocation), and cli (named project script with the prompt in
FUSION_NODE_PROMPT — raw commands still never accepted). Per-node maxRetries
overrides the executor-wide default (capped at 10). Await-input nodes pause the
task with status 'awaiting-user-input' and the question as pausedReason; on
unpause the newest steering comment is consumed as the answer and exposed in
graph context.
- liveness: graph-routed tasks count as executing in getExecutingTaskIds/
isTaskActive and are skipped by recoverCompletedTask + the completed-task
resume fast-path — recovery can no longer drive a parallel lifecycle (P0)
- handleGraphFailure: sets status 'failed' (self-healing revival exemption,
prevents FN-5704-style re-run loop) and leaves paused tasks untouched
- execute seam distinguishes pause/abort from implementation failure
- merge seam: 30-minute timeout so a wedged queue cannot strand the run
- awaitAbortInFlightTaskWork: defensive interceptor/routing cleanup
- deleteWorkflowDefinition: cascades to selections, materialized steps,
affected tasks' enabledWorkflowSteps, and the project default
Real engine seams: execute delegates to the legacy implementation phase via a
completion interceptor that stops execute() at the implementation-complete
boundary (no double review/merge); review performs the in-review handoff; merge
resolves through ProjectEngine.onMerge over the same serialized merge queue
(wired via a late-bound setMergeRequester, mirroring setMergeEnqueuer). Custom
graph nodes run on the proven WorkflowStep machinery (readonly tool policy,
verdict parsing). Adds a 'planning' seam to the vocabulary (no-op for
pre-specified tasks; custom planning is a prompt node today).
Entry point: execute() routes graph-selected tasks through the runner when
experimentalFeatures.workflowGraphExecutor is on, with process-wide routing
claims (FN-4811 posture), duplicate-dispatch dropping, pre-run errors falling
back to legacy, and mid-run errors parking the task in review (never re-running
the implementation, never stranding the task).
Flag off by default: all 587 executor tests pass unchanged.
Retry incomplete executor work with a fresh session after non-continuable session errors.
- add executor recovery handling that clears session state and requeues incomplete tasks to todo while retry budget remains
- preserve terminal failure behavior once the fresh-session retry budget is exhausted
- extend reliability coverage and AGENTS.md backstop notes for the new retry path
Files changed:
AGENTS.md | 1 +
.../post-done-continuation-no-wedge.test.ts | 48 +++++++++++++++++++++-
packages/engine/src/executor.ts | 40 ++++++++++++++++++
3 files changed, 87 insertions(+), 2 deletions(-)
Fusion-Task-Id: FN-5888
Fusion-Task-Lineage: 4ebce260-8923-4599-aced-541533b94543
Implements graceful suppression of transient `task.json` ENOENT errors in the executor, logging a suppression signal and surfacing a banner in the UI, with test coverage for both the executor behavior and notification service. Documentation in `docs/architecture.md` and a changeset for `@runfusion/f
Fusion-Task-Id: FN-5624
Co-authored-by: Fusion (runfusion.ai) <noreply@runfusion.ai>
Fusion-Task-Id: FN-5624
Cold-start dashboard responsiveness went from ~99s to ~6-11s. CPU profiling
identified two synchronous-spawn hotspots blocking the event loop:
- `GitHubTrackingReconciler` scanned up to 200 done tasks per startup,
each call into `getIssue` invoking `isGhAvailable()` + `isGhAuthenticated()`
via `execFileSync`. `gh auth status` makes a network roundtrip, so 400
sync spawns ≈ 71s of pure event-loop blocking (69% of cold-start CPU).
Memoized both checks with a 60s TTL; `resetGhAvailabilityCache()` is
exported for login/logout flows that need immediate invalidation.
- `PRAGMA integrity_check(100)` walks every page of the SQLite file (~7s
per database, multiple DBs × projects). The deferred check was scheduled
3s after init — right in the responsiveness-critical window. Pushed to
60s so the user is already interacting before it runs; check itself is
unchanged.
Also yields the event loop between major InProcessRuntime init phases and
between self-healing recovery steps (34 per project), defers orphan-task
AI agent resumption by 30s (env-overridable, auto-zero under Vitest), and
ships an opt-in `FUSION_TRACE_EL_LAG=/path/to/file` event-loop lag tracer
that diagnosed all of the above.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a static pretest check and a runtime vitest-setup wrapper that block
shell/process calls matching `kill|pkill|killall|fuser|lsof ... <port>` or
`.listen(<port>)` against reserved Fusion ports. Reserved set is dynamic:
default 4040 plus $PORT, $FUSION_SERVER_PORT, $FUSION_RESERVED_PORTS, and any
port responding to /api/health on 4040..4045 at worker startup.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
`notice` is `events.find(...)` which returns `undefined` (not `null`)
when no match. `waitFor(() => expect(...).not.toBeNull())` exited
immediately because `undefined !== null` — the test never actually
waited for the api mock to resolve. Sometimes the followup assertions
happened to land after the events fetched (test passed by luck);
sometimes they ran while notice was still undefined and the assertions
failed.
Switched all five waitFor sites to `.toBeDefined()` so they actually
block on the events-fetch resolution.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a "session runtime resolved" audit event that flows through the engine's main execution lanes — triage, executor, reviewer, merger, heartbeat, step-session-executor, and mission-execution-loop — with runtime mutation support and test coverage, plus a compile-fix for the merger auditor wiring.
Fusion-Task-Id: FN-5544
Co-authored-by: Fusion (runfusion.ai) <noreply@runfusion.ai>
Fusion-Task-Id: FN-5544
Follow-up to bf4428c00 (FF-only ref advance). After the prevention fix
new orphans can't form, but pre-fix orphans like f6358ce4 on
fusion/fn-5419 still need a path back onto the integration branch.
Adds an `orphan-our-advance` classification to contamination recovery:
a "unique" foreign commit whose Fusion-Task-Id trailer points at a
`done` task AND that is unreachable from refs/heads/<integrationBranch>
is treated as a stranded merger output.
For these, the executor attempts a fast-forward rehome onto the
integration branch via advanceIntegrationBranchRef (which still enforces
the FF-only invariant). When successful, the orphan sha is added to the
existing `shasToDrop` set so the same recovery pass that drops
already-upstream/misrouted commits also drops the now-upstream orphan.
Non-FF orphans (diverged from current integration tip) are refused.
Doing a cherry-pick onto the integration branch from inside automated
recovery would introduce conflict-resolution surface that's too high
blast radius for a never-event recovery path. The refusal log line
includes the exact `git cherry-pick <sha>` command an operator can run
manually.
Two new GitMutationType audit events:
- merger:orphan-rehome-ff (successful FF rehome)
- merger:orphan-rehome-refused (non-FF, manual cherry-pick required)
Tests in merger-orphan-rehome.test.ts cover classification (orphan,
not-done, already-reachable, no-trailer) and the rehome operation
(FF success advances the ref + emits the audit event; non-FF refusal
emits the hint and leaves the ref untouched).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Drops the "workflow-step" MockSessionPurpose enum value and the
workflowStepId / workflowStepTemplateId plumbing through
agent-runtime, agent-session-helpers, mock-provider, executor, and
merger. The seeded-workflow-prompts script loses its FN-5205
rationale comment + test (no longer applicable now that workflow
steps run through the regular session purposes).
Also strips the stale FN-5482 architecture-invariant bullet from
AGENTS.md and the corresponding audit-event line from
docs/architecture.md (the self-healing reclaim invariant they
described no longer holds).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Address the code-review finding on the prior commit: a natural premise-stale
summary like "PREMISE STALE: the task has no remaining work — implementation
is already done on HEAD" matches /\b(incomplete|not implemented|not done|
not finished)\b/i with 'the task' inside the 40-char first-person window,
refusing fn_task_done with summary-claims-incomplete and deadlocking the
escape hatch.
When summary starts (case-insensitive) with PREMISE STALE:, skip the
dissent-pattern and scoped-incomplete summary checks. Pending-code-review
and bulk-step-completion guards still apply unchanged.
Add executor-task-done-premise-stale.test.ts covering: the deadlock case
now passes; dissent phrasing in a sentinel summary is allowed; case-
insensitive sentinel; sentinel must be at the start (mid-summary doesn't
bypass); REVISE verdict still blocks even with the sentinel.
Root-cause prevention for the FN-5521 failure class where PROMPT.md describes
work that is already done on HEAD and the executor still marches through
plan/review/test/doc, then merger burns hours retrying against unrelated
flakes in @fusion/engine reliability-interactions.
- Executor system prompt: add a Preflight escape hatch. When Step 0
reproduces and finds HEAD already matches the desired state, the agent
marks Step 0 done, marks remaining steps skipped, and calls fn_task_done
with a `PREMISE STALE:` summary. Reuses the existing
evaluateTaskDoneRefusal/skipped semantics and the merger's
empty-own-diff fast-path — no new tools or refusal classes needed.
- packages/engine/vitest.config.ts: split into two projects. engine-default
retains full parallelism; engine-reliability scopes
src/__tests__/reliability-interactions/** to
poolOptions.threads.singleThread so event-ordering assertions stop
flaking under workspace-concurrent merge-gate load (the
`expected 24 to be less than 19` rowid interleaving in
merge-reuse-task-worktree).
Switch commits Fusion produces (both executor step commits and merger squash
commits) from setting `--author="Fusion <noreply@runfusion.ai>"` to appending
`-m "Co-authored-by: Fusion <noreply@runfusion.ai>"`. The user's configured
git identity now stays as the primary author/committer, and Fusion is recorded
as a co-author (recognized by GitHub for shared attribution). The
`commitAuthorEnabled` toggle and `commitAuthorName`/`commitAuthorEmail`
settings keep their existing keys; the dashboard settings UI relabels them
from "Author" to "Co-author" to match.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Contamination on fusion/<id> branches (FN-5233 was the recent example:
two untrailered feat(FN-5353): commits sitting on fusion/fn-5233) used
to be invisible until merge time, days after it happened. The executor
already runs assertCleanBranchAtBase at worktree acquisition and reclaim
— the gap was the active session window itself.
Add reportBranchAttribution(repoDir, branch, baseSha, taskId) which
walks base..branch and bins every commit into ownTrailed (healthy),
ownUntrailed (subject tag but commit-msg hook didn't fire), foreign
(different FN-id), or unattributed (no subject pattern, no trailer —
typically a hand-merge or plumbing commit). Wire it into the executor
right after captureModifiedFiles in the post-session path: when any
anomaly bucket is non-empty, emit a structured branch:attribution-
anomaly audit event and a task log entry. The audit itself is wrapped
in a try/catch so a probe failure never destabilizes a completing
session. New branch:attribution-anomaly and branch:auto-reattach-
authoritative GitMutationType variants accept the structured metadata
(the latter for the handoff re-attach added earlier this session).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Added executor logic to skip retries when a review is pending for a task, introducing a `pendingReviewBlockHelper` in the task-done path and updating the retry-gate to consult it; two new reliability-interaction test suites cover the feature behavior and composition with existing retry/backstop laye
Fusion-Task-Id: FN-5436
Close the last fire-and-forget gap from the previous fixes: the task:moved
(away from in-progress) and task:deleted listeners no longer call the
synchronous fire-and-forget `abortInFlightTaskWork`. Instead they track an
awaited disposal promise per task in `pendingTaskDisposals`. The task:moved
(to in-progress) dispatch path awaits any in-flight disposal for the same
task before calling `execute()`, so a fast bounce (in-progress → todo →
in-progress) no longer races the conflict-cleanup path against a still-live
shell.
`awaitAbortInFlightTaskWork` now claims each session surface (activeSessions,
activeStepExecutors, activeWorkflowStepSessions, activeSubagentSessions)
synchronously before awaiting any async abort. This lets concurrent disposal
calls for the same task dedupe naturally — the second call finds the maps
empty and no-ops, preserving the existing single-abort/single-dispose
contract that the soft-delete and user-cancel tests assert.
Adds a regression test in executor-user-cancel covering the re-dispatch
ordering: an immediate task:moved-to-in-progress that follows a still-running
task:moved-away must wait for abort to complete before execute() runs.
The legacy `abortInFlightTaskWork` is removed (no callers).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
