Makes graph-native post-merge the default and cuts the merger's legacy post-merge
path over so post-merge runs exactly once (via the graph), with migration data prep.
- experimentalFeatures.graphNativePostMerge → default ON; merger
hasEnabledPostMergeWorkflowSteps/runPostMergeWorkflowSteps become inert when on
(no double-run; proven by a new no-double-run test). Merger code kept (U7c removes it).
- Migration 130 (SCHEMA_VERSION 129→130) rewrites each task's enabledWorkflowSteps
entries that are legacy built-in pre-merge workflow_steps row ids → the optional-group
node id (browser-verification/code-review); dedupes; idempotent; identity-stable;
leaves node-ids/compiled/custom entries untouched. Table KEPT (U7c drops it).
- Investigation (real DBs): NO custom/plugin post-merge steps exist; the only post-merge
step is compound-engineering's 'document' graph node — so the merger no-op strands
nothing. Custom-step re-pointing was verified unnecessary and skipped.
Safe-to-drop in U7c still blocked by live readers: merger post-merge fns, store CRUD,
migrateLegacyWorkflowSteps/readConfig materialization (executor recovery reader is
already null-safe→advisory).
Plan U7b.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Adds graph-native post-merge step execution behind experimentalFeatures.graphNativePostMerge
(default OFF — byte-identical behavior until enabled). After a successful merge-attempt
(the merge seam awaits the merge Promise), the graph runs post-merge optional-group nodes
and records phase:"post-merge" results, non-blocking. The merge-region traversal hop is
inert when the flag is off, and empty for builtin:coding even flag-on (its merge exits only
reach merge-region nodes or end), so the parity oracle holds. Optional-group recording now
derives phase + log prefix from config.phase (defaults pre-merge). Adds postMergeOptionalGroupNode
factory for migration/custom workflows. Legacy merger post-merge path untouched (U7b cutover).
Plan U7a.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
From the U1-U6 code review (correctness/adversarial/reliability/maintainability):
- Delete dead code left by the runWorkflowSteps removal: parkTaskAfterWorkflowStepPause,
handleWorkflowRevisionRequest (+ createWorkflowRevisionFollowUpTask,
injectWorkflowRevisionInstructions), handleWorkflowStepFailure, the dead
partitionWorkflowRevisionFeedback export + its test, and 2 orphaned jsdocs;
reword 2 stale comments (executor.ts FN-6722, self-healing.ts jsdoc).
- Record step output/notes on graph workflow-step results: runGraphCustomNode now
emits contextPatch:{output,notes} so the Workflow tab shows real review feedback
and [pre-merge] revision logs carry detail (was always the fallback before).
- Sync the workflow-step seam contract: core (workflow-compiler SEAM_NAMES/order,
workflow-ir column map, builtin-workflow-prompts) now rejects the workflow-step
seam to match the engine's resolveSeamName, preventing a latent run-time crash on
a persisted/cloned def that core would otherwise parse.
Residual (tracked in the PR): re-introduce the FN-4343 per-step scope gate on the
graph path; the parked-failed recovery log wording; malformed-advisory->passed edge
case; recording for non-optional-group/split-branch step realizations.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
U6: delete the built-in WORKFLOW_STEP_TEMPLATES catalog + its materializer
(getBuiltInWorkflowTemplate/ensureWorkflowStepForTemplate/toBuiltInWorkflowStep);
inline the browser-verification + code-review name/prompt/toolMode/gateMode into
their optional-group IR builders (node bytes unchanged); simplify
resolveEnabledWorkflowSteps to an identity-stable pass-through (no materialization,
so the optionalGroupIdSet collision guard is no longer needed). Plugin-contributed
step templates are kept as the editor palette.
U5: remove the legacy /api/workflow-steps REST surface (GET/POST/PATCH/DELETE +
/refine + /workflow-step-templates/:id/create), the dead client fns, and the
Settings management UI; GET /api/workflow-step-templates now serves plugin
templates only. The create-time optional-step toggles remain.
Scope: the workflow_steps store CRUD + table are intentionally KEPT — still consumed
by the engine (merger/recovery) and needed by U7's migration; their removal + the
table drop land in U7.
Plan U5 + U6.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Removes the legacy workflow-step EXECUTION path now that the graph records results
(U2): delete runWorkflowSteps(), the workflow-step seam + runWorkflowStep primitive
(runtime-primitives, workflow-node-handlers, authoritative-driver), and the legacy
execute() step blocks. Keeps task.workflowStepResults + its store write path (the
graph's sink) and executeWorkflowStep/executeScriptWorkflowStep (reused by the graph).
- Watchdog recoverCompletedTask now re-enters via maybeExecuteWorkflowGraph so the
graph re-runs pending gates, records results, and owns the in-review/back-for-fix
transition (KTD-2).
- maybeExecuteWorkflowGraph fails CLOSED (parks) when a store lacks
getTaskWorkflowSelection AND the task has enabled pre-merge steps — closing the
FN-7039 silent-skip class without changing minimal-store implementation runs (KTD-5).
KNOWN GAP (follow-up): the FN-4343 per-step workflowStepScopeEnforcement leak check
lived only in runWorkflowSteps and is NOT yet replicated on the graph path. Merge-time
File Scope enforcement (FileScopeViolationError, squash overlap) is unaffected.
Plan U4.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Prior WIP added a default-on code-review optional-group to builtin:coding on the
success path (execute -> browser-verification -> code-review -> review). 9 exact
visitedNodeIds expectations across 3 suites predated that node and were failing.
Updated expectations to match the real traversal (insert code-review in position).
No product code touched; no assertions weakened.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Enabled optional-group nodes now upsert their outcome into the existing
task.workflowStepResults field keyed by node id, and emit [pre-merge] logs at
parity with the legacy runWorkflowSteps path. Disabled groups stay byte-inert.
Reuses the existing WorkflowStepResult type + store.updateTask path (no new
table/type/store method). Unblocks the unified progress bar for graph-run steps.
Plan U2.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Strengthen the fn_run_verification allowFullSuite parameter description, add an
AGENTS.md standing rule, and update docs/testing.md so agents default to a
file-scoped verification command and reserve allowFullSuite for genuinely full
runs with no targetable test set. allowFullSuite is the main way verification
balloons past its budget; the thin merge gate is the cross-cutting safety net.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Diff-proportional verification (deriveFileScopedPnpmTestCommand) + scope-aware
verification timeout, so merge/step checks finish in seconds. Propagated to this
worktree directly.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
echo ok is workspace-scoped, so the default verification budget is now 900s
(VERIFICATION_TIMEOUT_WORKSPACE_MS) rather than the retired flat 600s. Assert via
defaultVerificationTimeoutMs so the expectation tracks the scope-aware default.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The merger and executor verification gate (shared runVerificationCommand) used
a flat 10-min default (VERIFICATION_COMMAND_TIMEOUT_MS) for any configured
test/build command, while the fn_run_verification tool already derived its
default from command scope. A workspace-scoped command (a full suite, ~10+ min)
hit the flat 10-min wall and was killed as an infra timeout; a package-scoped
command got a too-generous bound.
Derive the default from command scope to match the tool: package-scoped
(pnpm --filter/-F ...) → 300s, workspace-scoped (root command like pnpm test)
→ 900s. An explicit project verificationCommandTimeoutMs still overrides, and
the 30-min hard cap still clamps the result. Covers both the merger and the
executor per-step auto-gate, which share runVerificationCommand.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A workspace land squash-merges each sub-repo via a clean room that first runs the
configured/inferred install (npm/pnpm/yarn). The install hard-fails by design so
merge verification never runs against an uninstalled checkout — but that let ONE
sub-repo with a manifest npm refuses to install (e.g. a corrupt `-@0.0.1`
lockfile entry rejected by npm 11) block landing every other sub-repo.
landWorkspaceTask now passes nonFatalDependencySync to landOneRepo: a clean-room
install failure is caught, logged + audited as a non-fatal degradation, and the
land proceeds (the git squash needs no installed deps; only dep-dependent
verification degrades for that repo). A real abort signal still propagates. The
single-repo land path keeps the documented hard-fail (flag defaults off).
Tests: new workspace-merger-deps-resilient asserts both the resilient workspace
land (all repos land despite install throwing) and the preserved single-repo
hard-fail. Also fix a pre-existing getTask mock gap in workspace-merger.test
(mergeAndReview reads getTask().comments) that broke 3 tests at the land step.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- dashboard: remove the "Branch needs reattachment" banner. It fired for any
in-review task with a null singular task.branch — the NORMAL state for a
workspace task (attachment is per-sub-repo worktrees in workspaceWorktrees), so
it was a permanent false positive. Genuine lost bindings are already reattached
automatically by self-healing's reconcileInReviewBranchRebind (event-driven on
move-to-in-review + sweep), so no manual user action is needed. Delete the
now-obsolete rebind-banner test + its registry entry.
- engine/self-healing: reconcileInReviewBranchRebind now explicitly skips
workspace tasks (never rebind candidates — their fusion/<id> branches live in
the sub-repos, not the non-git browse root; null root branch is healthy).
- engine/merger-ai: pre-merge prune treats an absent ai-merge search root (ENOENT)
as "nothing to prune" instead of warning on every workspace merge.
- test: add ToggleRight to the TaskDetailModal lucide mock (pre-existing gap from
FN-6880 that broke the whole suite at import).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Multiworkspace tasks could not complete due to two independent bugs:
1. task.workspaceWorktrees had no SQLite column / rowToTask mapping, so
fn_acquire_repo_worktree's updateTask write was dropped on every persist
(applyTaskPatch writes the DB-round-tripped task back to task.json). Every
later getTask returned undefined, so fn_task_done's scope verifier read {}
and blocked with "acquired no sub-repo worktrees", and isWorkspaceTask()
consumers misfired. Persist it mirroring mergeDetails (schema column + v129
migration + db-migrate + defineTaskColumn + TaskRow + rowToTask).
2. In workspace mode every task ran rooted at the shared browse-only root, and
setActiveSession registered that path keyed only by path — so a second
concurrent workspace task was rejected by the foreign-task guard
("active-session path ... is held by ..."). Give each task a task-scoped
synthetic session key (sessionRegistryPath), applied at all register and
unregister sites; the in-memory worktree Set still holds the real root.
Regression tests assert the persistence invariant across getTask/listTasks/
store-reopen and concurrent session registration across all three session
surfaces; both verified to fail without the fix.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add X-Session-Id and X-Session-Affinity headers to all outbound LLM chat
completion requests so LLM gateways can sticky-route consecutive requests
from the same conversation and observability tools (Langfuse, Arize) can
group stateless API calls into a single multi-turn trace.
The headers carry a stable identifier: the task id when available (stable
across pause/resume), otherwise the pi session id. The implementation wraps
modelRegistry.getApiKeyAndHeaders -- the single chokepoint pi-coding-agent
uses for both the main stream and compaction -- merging routing headers into
the resolved output. This covers all HTTP-based providers (built-in, custom,
and HTTP-streaming extensions) without disturbing auth resolution.
Also propagates taskId to four secondary executor sessions (retry,
verification-fix, workflow-step, child-agent) that previously fell back to
a per-instance pi id, fragmenting per-task observability grouping.
Closes#1675
- Await async spawned child session disposal
- Use own-key iteration for structured tool result previews
- Add FNXC requirement comments for new regression assertions
