agent-output-reconciler
Reconcile completed task results into one evidence-based report. This skill
describes what happened and what conflicts remain. It does not accept work,
merge changes, retry tasks, or write canonical memory.
Inputs
- the current plan and its task IDs, roles, scopes, dependencies, and success
criteria;
- each task packet, result summary, structured result, and raw-log path;
- the relevant
TaskCheckpoint and latest PolicyDecision;
- repository diff and test evidence.
Current plans use provider-neutral roles: primary-agent,
delegated-executor, reviewer, and synthesizer. Historical provider fields
may be parsed for migration only and must not influence current routing.
Workflow
- Validate the plan and checkpoint. If either is unreadable, stop and report a
blocker. Do not infer missing policy values.
- Inventory every planned task. Preserve
null, missing, cancelled,
declined, degraded, and failed results as distinct states.
- Filter unusable payloads before synthesis, while retaining a record of what
was filtered and why. Never turn a completed prose result into
null merely
because an optional structured-output step failed.
- Compare actual changed files with each task's declared scope. Record
justified transitive changes separately from unexplained drift.
- Verify claimed tests and artifacts from their evidence references. A result
without evidence remains unverified even if its prose says "success".
- Identify file collisions, incompatible assumptions, contradictory findings,
dependency-order violations, and unresolved risks.
- Compose a reconciliation report using
references/reconciliation_template.md.
- Hand the report to
agent-acceptance-gate. Do not convert a reconciliation
recommendation into an acceptance decision.
Output contract
The report must contain:
- plan and policy references;
- per-task role, status, evidence references, files changed, tests, and risks;
- filtered or missing result records;
- agreement that is supported by shared evidence;
- contradictions and file conflicts;
- recommended next action with explicit blockers.
Write coordination output to the configured scratch area by default. Promote a
report into a repository-owned evidence path only when the project explicitly
requires a shipping artifact or acceptance record. Scratch output is not
implicitly committable.
Policy and context
All numeric limits come from the canonical policy evaluated by
agent-collab policy evaluate. Do not copy default budgets into the plan or
this skill. Re-evaluate policy after a reconciliation cycle and before any new
spawn. Obey the decision's scope: a v2 action stop forbids repeating that action,
not safe primary-agent diagnosis; a goal stop preserves the hard gate. Persist
a checkpoint before continuing. An eligible v2 slice advances automatically
under the same goal, while v1 retains its original stopping behavior.
Safety rules
- Do not vote on factual truth. Prefer verified evidence and surface
disagreement.
- Do not hide failed or absent results with optimistic synthesis.
- Do not paste raw logs when a path and bounded excerpt are sufficient.
- Do not modify source files, canonical memory, history, or human decisions.
- Do not commit, merge, push, or approve the round.
1---2name: agent-output-reconciler3description: Use when multiple agent results must be checked for missing evidence, scope drift, contradictions, and merge conflicts before an acceptance decision.4metadata:5 source: https://skillmd.com/skills/wenyuchiou/agent-output-reconciler6---78# agent-output-reconciler910Reconcile completed task results into one evidence-based report. This skill11describes what happened and what conflicts remain. It does not accept work,12merge changes, retry tasks, or write canonical memory.1314## Inputs1516- the current plan and its task IDs, roles, scopes, dependencies, and success17 criteria;18- each task packet, result summary, structured result, and raw-log path;19- the relevant `TaskCheckpoint` and latest `PolicyDecision`;20- repository diff and test evidence.2122Current plans use provider-neutral roles: `primary-agent`,23`delegated-executor`, `reviewer`, and `synthesizer`. Historical provider fields24may be parsed for migration only and must not influence current routing.2526## Workflow27281. Validate the plan and checkpoint. If either is unreadable, stop and report a29 blocker. Do not infer missing policy values.302. Inventory every planned task. Preserve `null`, missing, cancelled,31 declined, degraded, and failed results as distinct states.323. Filter unusable payloads before synthesis, while retaining a record of what33 was filtered and why. Never turn a completed prose result into `null` merely34 because an optional structured-output step failed.354. Compare actual changed files with each task's declared scope. Record36 justified transitive changes separately from unexplained drift.375. Verify claimed tests and artifacts from their evidence references. A result38 without evidence remains unverified even if its prose says "success".396. Identify file collisions, incompatible assumptions, contradictory findings,40 dependency-order violations, and unresolved risks.417. Compose a reconciliation report using42 `references/reconciliation_template.md`.438. Hand the report to `agent-acceptance-gate`. Do not convert a reconciliation44 recommendation into an acceptance decision.4546## Output contract4748The report must contain:4950- plan and policy references;51- per-task role, status, evidence references, files changed, tests, and risks;52- filtered or missing result records;53- agreement that is supported by shared evidence;54- contradictions and file conflicts;55- recommended next action with explicit blockers.5657Write coordination output to the configured scratch area by default. Promote a58report into a repository-owned evidence path only when the project explicitly59requires a shipping artifact or acceptance record. Scratch output is not60implicitly committable.6162## Policy and context6364All numeric limits come from the canonical policy evaluated by65`agent-collab policy evaluate`. Do not copy default budgets into the plan or66this skill. Re-evaluate policy after a reconciliation cycle and before any new67spawn. Obey the decision's scope: a v2 action stop forbids repeating that action,68not safe primary-agent diagnosis; a goal stop preserves the hard gate. Persist69a checkpoint before continuing. An eligible v2 slice advances automatically70under the same goal, while v1 retains its original stopping behavior.7172## Safety rules7374- Do not vote on factual truth. Prefer verified evidence and surface75 disagreement.76- Do not hide failed or absent results with optimistic synthesis.77- Do not paste raw logs when a path and bounded excerpt are sufficient.78- Do not modify source files, canonical memory, history, or human decisions.79- Do not commit, merge, push, or approve the round.