The explicit-only review graph, durable report contract, three-review delivery budget, and
adversarial skill-adherence evidence are implemented in canonical shared skills. The final pass
refactors the runtime into the durable workflow-graph structure: one bootstrap, one deterministic
router, four flat gates, and one mode-specific runtime handoff contract. Canonical source, Harness
mirrors, managed fixtures, operator docs, and integration checks are aligned. Squash/push and
successor-baseline publication/readback are the remaining release actions.
review-phase now bootstraps through one deterministic router and loads exactly one flat gate:
preparation, one bounded all-lens run, report retention, or routing return.
Delivery appends review-owned runtime records to its caller-provided handoff. Standalone review
uses one deterministic repository-local handoff. Cold resume reconstructs the route from current
evidence without transcript continuity.
The run gate freezes one bounded snapshot, runs one advisory autoreview plus all independent
normative lenses, and writes one immutable local report. Retention and routing remain separate
gates.
Delivery recovers retained review authority before its budget guard, permits at most three
reviews and three corresponding repairs, and completes repair 3 through focused validation
without review 4.
create-plan emits implementation_skill_guidance; implement-spec forwards each entry and
requires exactly one implementation-note evidence record per guidance entry. Review verifies
those records against changed artifacts.
Report navigation, frontmatter guidance, operator docs, sync tests, managed fixtures, and
generated baseline identity now represent the same contract.
The user explicitly requested direct implementation with goalify and writing-for-agents,
without create-plan, implement-spec, or a delivery-phase wrapper. Therefore this run produced
no PLAN.md and no plan-produced implementation_skill_guidance entries. The evidence below
records the two explicitly assigned skills once each and does not fabricate guidance cardinality.
Activated one persistent goal covering canonical source, Harness mirrors, validation, main integration, and baseline publication; work continued against that completion boundary.
writing-for-agents
applied
Kept skill entrypoints concise, disclosed branch-specific graph and report rules through references, and preserved one authority for invocation, state, evidence, and retention semantics.
Canonical focused graph contracts pass 25/25 and the full shared-source suite passes 97/97 after
reconciling the latest frontend source. Pointer, disclosure, no-op, syntax, diff, and applicable
Skills CLI packaging audits pass.
Harness default sync and packaged canonical parity have zero differences. The active review
mirror and packaged review skill also have zero byte differences. Managed fixtures and generated
baseline identity resolve the same final source.
CHANGELOG.md and BASELINE_CHANGELOG.md define the planned successor
baseline/stable/2026.08.11-review-phase-durable-workflow-graph, compatible with
>=3.1.11 <4.
Publication is pending. No successor stable tag, uploaded assets, control-plane revision, or
downstream hi check result is claimed here.
Explicit invocation, target adapters, one frozen all-lens run, readonly validation, and failure semantics are covered by canonical and packaged contracts.
AC-025 through AC-029
Met
Planning guidance, worker forwarding, evidence cardinality, and adversarial verification are present in source and mirrors.
AC-030 through AC-054
Met
Immutable report retention, routing, counter recovery, repair epochs, clean fix-3 handoff, and distribution parity are covered by contracts, docs, and mirror checks.
AC-055 through AC-062
Met
External PR review remains excluded; budget exhaustion, lineage, preallocated run identity, atomic repair opening, and retry semantics are covered.
The full wiki suite passes 11/12. Its only failure is the pre-existing
.agents/skills/docs-onboarding/SKILL.md phrase hi scaffold init, outside this review-phase
change. Focused review wiki tests and checks pass.
Full bun run --cwd apps/wiki check reaches current content and passes Oxlint, then Oxfmt reports
two unrelated pre-existing scoped-agent-guidance grill-status and research-report files. Focused
formatting for this review-phase diff passes.
Publish baseline/stable/2026.08.11-review-phase-durable-workflow-graph with
bun run baseline:publish, then read back release assets, control-plane authority, and downstream
hi check evidence. Until then, no successor tag, uploaded asset, control-plane revision, or
downstream result is claimed.