[jhihweijhan]

Summary

Truncated review material can be mistaken for exhausted correction rounds. Define bounded recovery separately from finding correction while retaining delivery-only synchronization and evidence reuse.

delivery -> local impact filter -> necessary synchronization -> invalidated feature review
missing material -> bounded reads -> at most two recovery redispatches
demonstrated finding -> at most two correction/re-review cycles

Keep one independent reviewer responsible for each complete feature scope. Preserve raw attempts outside the final passing record, bind test reuse to recorded inputs, and persist unattended request budgets across generated commits.

Evidence

• Before: A fresh baseline reviewer on b17df83 selected safe recovery but explicitly found correction counting and reviewer continuity ambiguous. This did not reproduce premature stopping.
  After: Four added raw cases and separate rubrics cover material recovery, retry limits, partial/changing inputs and repeated preparation. Independent initial nine-case evaluation matched the intended decisions; follow-up findings were corrected and the final focused four-case review passed.
• Final candidate: d6cdd124d9191890afd1b9f06a6c04fef24d1706, compared with b17df8308b674704f9c0aab13cd4db434609e2c4.
• PYTHONDONTWRITEBYTECODE=1 npm test: 20 Node tests passed; 52 Python tests completed successfully with 13 local skips. Required CI ran all 20 Node and 52 Python tests successfully without skips, with pinned parsers and both Python versions.
• Claude Code Opus 5.5 reviewed raw cases before any rubric. Nine-case review: Orca dispatch ctx_b76df59fc8ff. Final focused verification: ctx_4f92871842f3; three source SHA-256 identities matched the candidate. No blocking findings remain.
• Non-blocking suggestion retained for future consideration: explicitly prohibit listing the manifest itself in policy_inputs. No such configuration was introduced here; the suggested self-invalidation case was not reproduced. This change excludes review outputs while preserving behavior-bearing test/configuration inputs.
• No target-project hooks, unattended model runner, runtime equivalence or token savings were measured. Instructions and decision cases do not establish those integrations.

Raw case	Observed independent decision	Source within skills/flow-docs-sync/
WIP	Read the trusted baseline and pending diff; defer generation/review and retain pending delivery.	SKILL.md, bootstrap and Maintain
Ready and later change	Prepare at ready delivery; revalidate subsequent candidates/base changes; reuse matching evidence.	SKILL.md, Maintain; contract, delivery triggers
Comments	Refresh raw MD5 at delivery; retain flows and semantic evidence when bound AST inputs match.	contract, change filtering
Hook	Do not impose the strict hook under PR delivery; preserve and explicitly resolve existing strict policy.	contract, delivery triggers
Installation	Installation does not configure GitHub generation; report missing runner and use agent preparation.	contract, delivery triggers
Material recovery	Two ended incomplete attempts use zero correction cycles and one recovery redispatch; one recovery remains.	reviewer prompt, correction and material recovery
Retry limits	Preserve findings; retain budgets across fingerprints and task-owned pushes; unrelated additions do not justify retry.	reviewer prompt, recovery; contract, unattended request identity
Partial and changing inputs	Reject combined partial passes and mixed candidates; use one authoritative base and stable full-feature material.	reviewer prompt, preparation and acceptance
Repeated preparation	Batch at delivery; reuse unchanged evidence; avoid preparation agents, per-chunk fan-out and unnecessary redraws.	SKILL.md, Maintain; reviewer prompt, preparation

Final focused cases independently confirmed stale/timeout recovery, agent pushes under user credentials, attempt-history storage, and exclusion of review outputs without excluding behavior-bearing tests.

Merge Danger

Door: two-way

Instruction-only change; no checker code, schema, dependencies or workflow changes.

Blast Radius: guidance

Adopting agents change review preparation and bounded recovery behavior. Required final candidate checks and repository human approvals remain in force.