Follow-up to 13a2b2a9d, from a multi-agent review of that commit. Three of its
claims did not hold.
1. fn_delegate_task bypassed the gate entirely (P0). It reaches the same
createAgentTask primitive, was registered unconditionally in both session
lanes, and validated only that the TARGET agent is non-ephemeral — never the
caller. Under Deny an ephemeral worker could enumerate agents and delegate
unlimited tasks. It is now withheld under Deny, and also under
upon_validation: delegation has no proposal channel, so leaving it available
would launder a create past the operator review that policy requires.
2. The widened dedupe window was capped at 5 minutes. The store query in
branch-and-pr-entities.ts carried its own independent `?? 60_000` /
`min(300_000, …)` pair, so widening only duplicate-guard.ts under-delivered
and made the new ceiling unreachable. Both sites now share
FINGERPRINT_WINDOW_DEFAULT_MS / FINGERPRINT_WINDOW_MAX_MS.
3. The pi-extension gate does not fire at all. pi's ExtensionContext carries no
agentId — the read is a speculative cast and only tests supply one, so every
real call short-circuits as a human caller. The fail-closed direction is kept
for the day an identity signal exists, but the limitation is now documented
instead of implied to be enforcement.
Also: the session prompt now states when creation is disabled and names
fn_task_log as the fallback (the base prompt still taught fn_task_create, which
is the same instruction/capability mismatch that fed the retry storm);
suppression emits an `agent:task-create-withheld` run-audit event; and the two
source-text ratchet tests are replaced with behavioral assertions on the tool
list the executor actually hands the model — verified to fail when the guard is
broken, which the string assertions did not.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Operator report: with project policy "Ephemeral agent follow-up tasks = Deny",
an executing agent filed ten follow-up tasks — five parallel fn_task_create
calls it reported as timed out, then five sequential retries.
Two defects:
1. Deny was advisory. fn_task_create was registered for every session and only
refused inside execute(), so the model still saw the tool, planned around it,
and retried it. The pi extension's isEphemeralCallerAgent also failed OPEN
whenever the caller id did not resolve to an agent row — which is the normal
shape of an ephemeral task-worker — so on that lane Deny was a no-op.
2. The deterministic content-fingerprint duplicate window was 60s, which only
covered concurrent in-flight creates. A retry two minutes later saw nothing
and filed a second task.
Fixes: isAgentTaskCreateToolAvailable() withholds the tool from ephemeral
sessions under Deny in both engine lanes (outer execution session, per-step
workflow session); isEphemeralCallerAgent fails closed on an unresolvable
caller id; the fingerprint window goes 60s -> 10m (clamp ceiling 5m -> 1h).
upon_validation keeps the tool, and permanent-agent and human/chat callers are
unaffected.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
## Summary
- Fix a false-positive in `StuckTaskDetector` where legitimate long
single-step work (E2E debugging, iterative fix/test cycles) was
classified as a loop and kill/requeued.
- Root cause: loop meant “no step status transition for
`taskStuckTimeoutMs` + high activity volume,” conflating **step
progress** with **actual activity**. Agents can stay productively busy
on one step for 10+ minutes with zero repetition.
- Loop now requires thrash evidence on top of volume + no step progress:
- **repetitive tool fingerprints** (`toolName` + primary-arg detail in a
sliding window), or
- **elevated ignored step-update rebuffs** (≥ 10)
- Wire tool name/detail from `AgentLogger` → executor / step-session
into `recordActivity(...)` so novelty is measurable.
- Document the thrash-evidence rule in `docs/architecture.md`.
## Test plan
- [x] `pnpm --filter @fusion/engine exec vitest run
src/__tests__/stuck-task-detector.test.ts
src/__tests__/reliability-interactions/non-progress-churn.test.ts`
- [x] Regression: high-volume **diverse** iterative activity (174
events) does **not** classify as loop
- [x] High bare text/heartbeat volume without tools does **not**
classify as loop
- [x] Repetitive identical tool fingerprint + timeout **does** classify
as loop
- [x] Ignored step-update thrash (≥10) with volume **does** classify as
loop
- [x] Existing FN-5168 no-progress-churn + FN-6598 verification
suppression paths still pass
- [ ] CI gate green
<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit
* **Bug Fixes**
* Improved stuck/loop classification by requiring explicit “thrash
evidence” (repetitive tool fingerprints and/or elevated ignored progress
rebuffs), reducing false positives for busy but diverse work.
* Updated loop evidence tracking to incorporate tool name plus
summarized tool-argument detail.
* Cleared loop evidence appropriately after verification, progress
updates, and task resumption.
* Extended tool-start telemetry/callbacks to include optional tool
detail.
* **Documentation**
* Refined loop-classification criteria to match the new evidence gates.
* **Tests**
* Updated/expanded stuck/loop and churn scenarios to validate the
evidence-based behavior and callback ordering.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
## Summary
Ordered task steps can no longer appear active ahead of unfinished
predecessors. Step starts now use the same dependency-aware ordering
guard as completions, while steps explicitly declared independent remain
parallelizable. Rejected executor updates explain that the lifecycle
transition was suppressed instead of implying completed work was
overwritten.
## Validation
- Reproduced the FN-8490 concurrent update sequence and verified later
steps remain pending.
- Passed 15 PostgreSQL step-order tests, the focused executor response
test, core and engine typechecks, changeset validation, and `pnpm
verify:fast` including boot smoke.
- The full `executor-prompt.test.ts` run retains five pause-behavior
expectation failures that reproduce unchanged on `origin/main`.
<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit
* **New Features**
* Enhanced the step start hook to support an awaited “pre-start
projection” that can reject startup via `false` (sync or async),
preventing step-session creation/completion.
* Added a step-start “verdict” so steps can be started or blocked
deterministically (including “resumed” behavior).
* **Bug Fixes**
* Prevented ordered/dependency steps from transitioning out-of-order by
enforcing guards for both in-progress and done transitions, including
concurrent update attempts.
* Improved integrity/out-of-order warning behavior and suppression
details when persisted status doesn’t match expectations.
* **Tests**
* Added/updated PostgreSQL and engine regression coverage for
blocked/resumed start and start-rejection control flow.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
A rate-limited reviewer filled a task's Chat tab with 14 identical
"Reviewer using model: ..." markers and no review text, hammering an
already-limited provider.
Root cause: the reviewer was the only AI lane that never classified
provider errors, so a 429 became an UNAVAILABLE verdict. With no
validator fallback configured the fallback ladder re-ran the SAME model
instantly, and fn_review_step answered with "code review remains
blocking; retry once" — bounding the loop with prompt text rather than
code. The tool's catch-all also swallowed the error into tool output, so
withRateLimitRetry, UsageLimitPauser and RetryStormError never fired.
- reviewer: throw ReviewerProviderError for usage-limit/transient errors
instead of laundering them into UNAVAILABLE, and never spend the
fallback budget (which bounds bad reviews) on an outage.
- reviewer: absorb flaky-network blips in-lane via withRetry with
jittered backoff; rate limits still escalate immediately.
- executor: re-raise the fatal after the prompt via
throwDeferredReviewerFatal — pi-agent-core converts tool throws into
tool_error results, so a tool cannot throw out of session.prompt().
- executor: give code review a real MAX_CODE_REVIEW_UNAVAILABLE_RETRIES
counter, mirroring the plan/spec limiter.
- reviewer: dedupe the model marker on text, so same-model retries stay
silent while a genuine model switch still emits.
Also fixes the run-on rendering: AgentLogType gains `status` for complete
engine messages. `text` means "streamed delta" and is re-glued with
join(""), which is why N standalone markers rendered as one string. The
split is at the type, not a separator — a separator would reintroduce the
FN-5787/5789/5803 streamed-spacing regression.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Plugin-contributed skills previously registered only a name for sessions and the dashboard, so their SKILL.md bodies were never actually loaded — fix threads real body paths through to both session creation and the Skills UI.
- Resolve each enabled plugin skill's body path via @fusion/core's resolvePluginSkillBodyPath and thread its body dir (plus parent dir) into every session-creating lane (executor primary/retry/verification-fix/step/child-agent, triage, reviewer, merger, agent-heartbeat, cron-runner) as additionalSkillPaths, unioned with existing CE skill dirs.
- Add collectPluginSkillNames/mergePluginSkills additionalSkillPaths plumbing in session-skill-context.ts so plugin skill discovery paths flow the same way as native/role-fallback skills.
- Update dashboard skills-adapter.ts to read plugin skill SKILL.md and reference files from disk (via the traversal-guarded reader) instead of returning a runtime-placeholder/"not found" response for plugin-sourced skills.
- Document the plugin skill body delivery mechanism in docs/PLUGIN_AUTHORING.md.
- Add regression coverage: plugin-skill-body-delivery.test.ts, expanded session-skill-context.test.ts and skills-adapter.test.ts.
- Add changeset fn-7857-plugin-skill-body-delivery.md (minor, fix).
Files changed:
.changeset/fn-7857-plugin-skill-body-delivery.md | 7 ++
docs/PLUGIN_AUTHORING.md | 3 +
.../dashboard/src/__tests__/skills-adapter.test.ts | 92 ++++++++++++++++------
packages/dashboard/src/skills-adapter.ts | 33 ++------
.../__tests__/plugin-skill-body-delivery.test.ts | 75 ++++++++++++++++++
.../src/__tests__/session-skill-context.test.ts | 84 +++++++++++++++++++-
packages/engine/src/agent-heartbeat.ts | 3 +-
packages/engine/src/cron-runner.ts | 2 +
packages/engine/src/executor.ts | 25 ++++--
packages/engine/src/merger.ts | 10 ++-
packages/engine/src/reviewer.ts | 2 +
packages/engine/src/session-skill-context.ts | 43 ++++++++--
packages/engine/src/step-session-executor.ts | 5 +-
packages/engine/src/triage.ts | 3 +-
14 files changed, 318 insertions(+), 69 deletions(-)
Fusion-Task-Id: FN-7857
Fusion-Task-Lineage: 9ba4c305-8b38-4ae8-85b3-4c87205ef767
Co-authored-by: Fusion (runfusion.ai) <noreply@runfusion.ai>
Ensure task-detail comments are delivered to live executor threads and preserved for the next step prompt when no step session is active.
- Forward steering comments through legacy, step-session, and workflow-step executor targets with delivery status logging.
- Keep step-session task details updated and include pending steering comments in full and reduced step prompts.
- Track delivered steering comment IDs so comments are injected or queued exactly once across active and subsequent step sessions.
- Update step-session executor tests for live steering, queued prompt fallback, and reduced prompt behavior.
Files changed:
.../src/__tests__/executor-step-session.test.ts | 467 ++++++---------------
.../src/__tests__/step-session-executor.test.ts | 63 ++-
packages/engine/src/executor.ts | 34 +-
packages/engine/src/step-session-executor.ts | 69 ++-
4 files changed, 283 insertions(+), 350 deletions(-)
Fusion-Task-Id: FN-6590
Fusion-Task-Lineage: 18fffd41-7632-4f29-8721-daaf3c239a74
Workflow context-limit recovery prompts now direct agents to read task attachments from the project root.
- Pass the project root into reduced workflow step prompts so attachment directories can be absolute when available.
- Replace the autonomous-agent "ask for context" wording with direct file-reading guidance for attachments.
- Cover root-aware, fallback, placement, and context-limit recovery attachment prompt behavior.
- Add a patch changeset for the published Fusion package.
Files changed:
.../fn-6375-workflow-attachment-recovery-prompt.md | 5 ++
.../src/__tests__/step-session-executor.test.ts | 70 +++++++++++++++++++---
packages/engine/src/step-session-executor.ts | 8 ++-
3 files changed, 72 insertions(+), 11 deletions(-)
Fusion-Task-Id: FN-6375
Fusion-Task-Lineage: ed9a5367-1623-49cb-9c7f-116c59ce5117
U4: seams stamp the governing node id into run context; a per-seam
resolveSeamColumnAgent feeds the core resolver and threads the effective
agent through resolveExecutorSessionModel, runtime hints, persona, memory
tools, and StepSessionExecutor attribution. Characterization tests pin the
no-binding path byte-identical. Gating/deferral principal moves in U5.
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
Adds a global `thinkingLogEnabled` setting that gates AI thinking log persistence across the engine (executor, reviewer, merger, triage, step-session) and exposes the control in the dashboard Settings modal, with tests verifying settings parity and modal behavior.
Fusion-Task-Id: FN-4062
