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
- merger-ai: resolve+persist concrete landedSha when a sub-repo is recognized
already-landed via the Fusion-Task-Id trailer fallback, so finalize no longer
drops it and mis-finalizes a fully-landed workspace task as a no-op
- project-engine: manual-merge land-lease busy errors reject the resolver without
burning mergeRetries; clear stale busy-reenqueue counter on real partial land;
persist retry count before arming the backoff timer (fail closed on write error)
- cli/dashboard + task: use shared isWorkspaceTask predicate instead of inlining
- base-commit-capture: POSIX single-quote shell escaping for integration ref
- git-repository: validate workspace.json repos elements are strings
- merger-ai: drop dead store param from landOneRepo
- tests: assert the 60s backoff cap across cycles; exercise the real runAiMerge
merge door; fix non-git-root assertion; re-export real workspace error classes
in the merger-ai mock (fixes 24 pre-existing instanceof-undefined failures);
remove generic fake-timer smoke test now covered by the live engine assertion
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Wrap the fatal-path acquisition observability writes (logEntry + audit.git)
in safeObserve so a store/audit throw can't replace the original
acquisition error, keeping WorkspaceRepoAcquireBusyError instanceof checks
reliable upstream.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
5-persona review of the Phase-C per-repo merge loop. No P0; the no-push invariant
and retry/park accounting verified clean. Fixed:
Land mechanics (merger-ai.ts / active-session-registry.ts):
- persistRepoLandedSha no longer swallows the DB write: a failed landedSha write
after the ref advanced now escalates to WorkspacePartialLandError so the engine
parks/retries instead of silently re-landing (duplicate squash). isRepoLanded
gains a landedSha-independent fallback — it scans the integration ref for this
task's Fusion-Task-Id trailer (a squash commit is NOT a branch descendant, so a
branch-ancestor check is provably wrong), so an actually-landed repo is skipped
on retry.
- The land lease is now taskId-aware across kinds: any foreign-task holder on a
sub-repo path is contention (a merging task can't run over an executing task's
acquire lease), and registerPath throws ActiveSessionPathHeldByForeignTaskError
instead of silently clobbering a different task's entry.
- The per-repo loop is wrapped in try/finally(setStatus(null)) so the busy/partial
throws can't leave the task stuck 'merging'. WorkspacePartialLandError is a real
exported class (not a .name-mutated Error). finalizeWorkspaceTask re-reads fresh
and no longer swallows the mergeDetails write (TOCTOU). isRepoLanded exported for
Phase D.
Dispatch + doors (project-engine.ts / dashboard.ts / task.ts / @fusion/core):
- getTask-null in the partial-land catch fails closed (park) instead of defaulting
retries to 0 and scheduling an indefinite retry storm.
- The merge-confirmed reachability fast-path skips workspace tasks (its
representative commitSha is a sub-repo squash sha, unreachable in the root cwd —
it was demoting fully-merged tasks); they're verified by per-repo landedSha.
- The CLI/dashboard merge doors now return merged:true on full land (were hardcoded
merged:false). WorkspaceRepoLandBusyError re-enqueues with backoff WITHOUT burning
the mergeRetries quota (bounded busy counter) so contention can't park a healthy
task. Backoff capped at 60s. shouldRetryWorkspacePartialLand folded into
shouldRetryAutoMergeConflict. Catch switched to instanceof. New canonical
isWorkspaceTask predicate in @fusion/core.
Gate green: build, typecheck, lint, test:gate (649+58); workspace-merger + oracle
+ project-engine 174.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Addresses the follow-up review on the foundation fixes:
- Re-read the task via store.getTask immediately before merging the per-repo
entry, so a concurrent sibling-repo acquisition that landed since the initial
read isn't clobbered by updateTask's wholesale map replace (narrows the
read-modify-write window to the store lock; a fully atomic per-repo store-level
merge remains a follow-up).
- Normalize the inline FNXC comment to the FNXC:Area yyyy-MM-dd-hh:mm: convention.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- base-commit-capture.ts: shell-quote integration branch with a real
single-quoted POSIX literal instead of JSON.stringify (not shell-safe).
- TaskCard.tsx: memo compares full workspaceWorktrees values, not just key
sets, so a same-key worktreePath/branch change re-renders.
- TaskDetailModal.tsx: gate/render workspace summary off hydrated workingTask.
- worktree-acquisition.ts: null the singular worktree/branch columns in the
workspaceWorktrees write so isWorkspaceTask stays true; wrap non-fatal
post-acquire observability so logEntry/audit can't re-escalate to fatal.
- agent-tools.ts: register sub-repo worktree via onAcquired unconditionally
(idempotent) so a resumed/already-acquired path is tracked after restart.
- executor.ts: DB liveness fallback also checks task.workspaceWorktrees paths.
- executor-workspace.test.ts: root non-git assertion runs in fx.rootDir (".").
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Resolves the actionable CodeRabbit threads on the workspace-mode foundation:
- project-resolver: defer saveWorkspaceConfig until after the user confirms init
and store.init() succeeds (no partial .fusion/ on a declined/non-interactive run).
- git-repository: validate each candidate with a real `git rev-parse` work-tree
probe before counting it (no false-positive repos from stray .git markers);
loadWorkspaceConfig now rejects absolute paths, `..` escapes, and non-string
entries so a corrupt/malicious config can't resolve outside the workspace root.
- executor: gate workspace mode on repos.length > 0 at all three sites so an
empty { repos: [] } can't bypass the git-repo guard or enable an empty workspace.
- worktree-acquisition: thread the configured-command runner through the workspace
acquire path (sub-repos run their init setup); validate repoRelPath as an in-root
relative path before joining; liveness-check a remembered worktree before
reporting it ready (pruned paths fall through to re-acquire); clear the singular
task.worktree/branch after persisting per-repo state (per-repo state lives only
in workspaceWorktrees).
- agent-tools: forward runContext into acquireWorkspaceRepoWorktree for log attribution.
The executor-workspace test's mock-the-subject pattern is left for the
session-scoping follow-up that rewrites it with a real two-repo fixture (FN-5048).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
landWorkspaceTask now holds a per-repo land lease around each landOneRepo call:
a new activeSessionRegistry kind "workspace-repo-land" keyed on the sub-repo
absolute path, registered synchronously before the per-repo try and released in
a finally (on success and failure, only yanking our own taskId+ownerKey entry —
never a foreign/different-kind entry). Two workspace tasks landing the same
sub-repo serialize; the loser throws the retryable WorkspaceRepoLandBusyError,
which reuses the U2 partial-land retry/park machinery (consume a mergeRetry,
backoff re-enqueue up to MAX skipping landed repos, then operator-park). Disjoint
sub-repos never falsely serialize.
The lease is for serialization / clean-room-collision avoidance, not ref
correctness — advanceIntegrationBranchRef's CAS already makes interleaved
update-ref safe. Distinct from the execution-phase "workspace-repo-acquire" lease
(different kind, different lifecycle phase, each ignores the other's entry).
3 new tests (serialize, independence, release-on-failure); oracle (56) + U1/U2
(idempotency) stay green. Gate: build, typecheck, lint, test:gate (649+58).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
landWorkspaceTask now tracks per-repo landing and finalizes the task exactly once.
After a repo lands, its advanced integration tip is persisted as
workspaceWorktrees[repo].landedSha (fresh-read merge, siblings untouched). Before
landing, isRepoLanded skips a repo iff its landedSha is present AND an ancestor of
(or equal to) its local integration ref — so a retry after a partial land never
re-advances an already-landed ref. finalizeWorkspaceTask runs only when every
acquired repo is landed: it builds an aggregate MergeResult (representative
commitSha + a workspaceLandedShas map in MergeDetails) and calls the existing
task-global finalizeTask once, satisfying the task:merged consumer. No premature
done on the first repo.
Partial lands surface as WorkspacePartialLandError; the engine consumes a
mergeRetry and re-enqueues landWorkspaceTask (skipping landed repos) with the
existing conflict-retry backoff up to MAX, then operator-parks (status:failed) —
mirroring shouldRetryAutoMergeConflict (new exported shouldRetryWorkspacePartialLand
seam). The defense-in-depth WorkspaceTaskMergeError still hard-fails without
burning retries; manual merges fall through to rejectMergeResolvers.
types: workspaceWorktrees entry gains landedSha?; MergeDetails gains
workspaceLandedShas?. 6 new idempotency/predicate/finalize-once/retry-park tests;
oracle (52) + U1 (5) stay green. Gate: build, typecheck, lint, test:gate (649+58).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The foundation imports acquireWorkspaceRepoWorktree in executor.ts but
deliberately stops before wiring it into the executor lifecycle, so the import
is unused and fails @typescript-eslint/no-unused-vars (the sole Lint failure on
this PR). Remove the dead import; it is reintroduced with real usage in the
session-scoping follow-up.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Extracts the per-repo land mechanics out of runAiMerge's inline clean-room
closure into an exported landOneRepo(store, repoRootDir, branch, integrationBranch,
ctx): pre-merge prune (rooted at the sub-repo), the clean-room temp worktree,
mergeAndReview, landSquash, and the CAS concurrent-advance retry that advances ONE
local integration ref — no remote push. runAiMerge is rewired as the single-repo
caller (its task-global finalization unchanged); the merger-ai suite (56 tests)
stays green as the byte-for-byte oracle.
landWorkspaceTask loops a workspace task's acquired sub-repos (sorted keys),
re-resolving each repo's integration branch with the shared override stripped
({...settings, integrationBranch: undefined, baseBranch: undefined}) so each
sub-repo lands on its own origin/HEAD, calls landOneRepo per repo, and aggregates
repo-tagged results — land-as-you-go on each repo's LOCAL ref (D2/D5). It does NOT
finalize/move the task (finalize-once + landed-tracking + idempotent retry are U2).
Door routing (KTD2): the engine dispatch and the user-facing CLI `fn task merge`
+ dashboard merge doors route workspace tasks to landWorkspaceTask so manual merge
works; store.mergeTask, aiMergeTask, and the runAiMerge chokepoint guard keep
throwing WorkspaceTaskMergeError as defense-in-depth.
New two-repo fixture tests: both repos land + no-push assertion, per-repo
override-stripped resolution onto distinct branches, repo-B conflict partial land
(task not moved), defense-in-depth throws. Gate green: typecheck, lint, build,
test:gate (649+58).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
ce-code-review (4 personas) on Phase B. No P0; the review conjunction was
confirmed safe (no false-done — empty map and per-repo throws both route to
UNAVAILABLE, which blocks). Applied:
P1: the fn_task_done scope-leak guard now fails CLOSED in workspace mode — a
per-repo capture throw blocks completion ("refusing as a precaution") instead of
the outer .catch returning {blocked:false} and letting an incomplete check pass.
A scoped task that acquired ZERO sub-repo worktrees is now blocked rather than
silently passing scope enforcement.
P2: reviewWorkspacePerRepo breaks on the first non-APPROVE repo so a later repo's
throw can't discard an already-determined REVISE (callers were seeing UNAVAILABLE
instead). captureWorkspaceModifiedFiles isolates each per-repo capture in
try/catch so one repo's throw can't skip the modifiedFiles write. The .changeset
always-allowed carve-out is honored in workspace mode: the scope-leak branch now
filters repo-LOCAL paths via the (previously dead) workspace-paths.ts
deriveRepoScopeSubset helper through the same filter as the singular path, so a
sub-repo .changeset/* no longer falsely blocks fn_task_done. All four per-repo
loops iterate sorted keys for deterministic offending-repo reporting; the dead
repoRel callback param and the duplicate path-normalizer are removed.
Verified safe (no change): the reviewer semaphore releases on throw (try/finally),
and per-repo reviewers inherit the task abort via session disposal.
Deferred to Phase C: extracting a workspace-executor.ts module (before the merge
loop lands). Gate green: typecheck, lint, build, test:gate (649+58).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
In workspace mode both review entry points and the completion guards now iterate
every acquired sub-repo. A shared reviewWorkspacePerRepo loops task.workspaceWorktrees
and invokes the existing single-cwd reviewStep once per repo (cwd = the sub-repo —
the reviewer agent runs its own git diff there), aggregating repo-tagged verdicts
as a conjunction: the task is reviewed only if every repo APPROVEs; the first
non-APPROVE repo's verdict becomes the aggregate. Both call sites loop — the
in-session fn_review_step tool AND the step-inversion seam (createReviewStepTool
and the stepReview workflow seam) — so no review surface silently scopes to the
non-git root (FN-5893). reviewStep itself stays single-cwd; the callers loop.
fn_task_done completion verification iterates per repo: verifyWorktreeInvariants
(from U1) already covers all worktrees, and evaluateTaskDoneScopeLeak now loops
each sub-repo (cwd + repo.baseCommitSha, repo-prefixed touched files vs the
repo-prefixed declared File Scope), blocking on the first repo with off-scope
files and naming it. Both return shapes preserved (ReviewResult; {blocked,message}).
New workspace-paths.ts repo-prefix helper (deriveRepoForPath/splitRepoScopedPath/
deriveRepoScopeSubset; segment-wise longest-prefix match, unscoped fallback) —
master U5 reuses it. Singular non-workspace path unchanged. 16 new fixture tests.
Gate green: typecheck, lint, build, test:gate (649+58).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
In workspace mode the executor now captures changes and verifies worktree
invariants per acquired sub-repo instead of degrading to empty against the
non-git root. Post-session capture (:7898) gains a workspace branch that loops
task.workspaceWorktrees and reuses captureModifiedFiles(repo.worktreePath,
repo.baseCommitSha, …) per repo — inheriting resolveDiffBaseRef's merge-base
fallback (repo baseCommitSha may be undefined) and the filterFilesToOwnTaskCommits
contamination/divergence audit — then prefixes each repo's files with the repo
path into task.modifiedFiles. Branch attribution runs per sub-repo (cwd), never
against the root. The no-op assertCleanBranchAtBase is not iterated.
verifyWorktreeInvariants is un-stubbed for workspace mode: it iterates every
workspaceWorktrees entry asserting toplevel match + HEAD on fusion/<id>, and
returns the FIRST failing repo while preserving the exact discriminated union
{ok:true} | {ok:false; reason:'wrong_toplevel'|'wrong_branch'|'no_commits';
observed; expected} (the :10889 consumer switches on reason for requeue/handoff)
— the new repo field is additive. Singular non-workspace path unchanged.
Real two-repo fixture tests (capture A+B repo-prefixed vs own base, undefined-base
fallback, foreign-commit contamination audit, wrong_branch verify failure,
single-repo regression). Gate green: typecheck, lint, test:gate (649+58).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
ce-code-review (5 personas) on Phase A. No P0; the workspace-root-removal path
was ruled out and the contract changes verified additive. Applied:
P1: fn_acquire_repo_worktree now catches WorkspaceRepoAcquireBusyError (and
generic failures) and returns a sanitized retryable isError instead of an
uncaught throw into the agent loop; runContext is forwarded so audit/log keep
attribution. Per-repo acquired worktree paths are now registered into the
executor's activeWorktrees Set (via an onAcquired callback) — previously the Set
only held the browse-only root, making the U1 per-repo liveness invariant hollow.
Post-acquire identity-guard install and base-SHA capture are now non-fatal
(log-and-continue): a hook/branch failure no longer strands the on-disk worktree
(the worktree is usable without the guard; an undefined baseCommitSha is already
an accepted state).
P2: the KTD3 settings-strip also strips settings.baseBranch (resolveFromSettings
falls back integrationBranch → baseBranch, so a shared baseBranch leaked); the
workspaceWorktrees write re-reads the task fresh before merging to avoid a
sibling-repo clobber on sequential acquires (store-level atomic merge deferred to
Phase B); the busy-path logging is wrapped so it can't mask the busy error; the
TaskCard memo compares key-sets not counts; the stuck-kill no-op for workspace
tasks is now logged; the exclusivity check-then-act synchrony is documented.
Residuals (Phase B): per-repo worktree teardown, orphan-scan coverage, reaper
dedup, store-level atomic merge.
Gate green: typecheck, lint, build, test:gate (649+58), affected (25 + TaskCard 251).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
