Subagent Result Merge

Normalize parallel findings into one decision-ready report.

Worker Result Contract

Extend this contract for native and external workers; do not introduce a parallel report format. The launcher records execution facts independently of the worker's prose. This minimal blocked example is a shape example, not evidence of an installed or active integration:

{
  "task_id": "example-task",
  "goal": "Review the allowed synthetic snapshot",
  "base_commit": "unknown",
  "snapshot": {"state": "not_created", "sha256": "unknown", "files": []},
  "requested": {"provider": "explicit-local-mapping", "model": "explicit-local-mapping"},
  "observed": {"provider": "unknown", "model": "unknown", "source": "unknown", "verification": "unknown"},
  "status": "blocked",
  "findings": [],
  "checks": [],
  "artifacts": [],
  "open_questions": ["Runtime capability has not been verified"],
  "usage": "unknown"
}

Required fields:

• task_id, goal, base_commit, and snapshot: capture the base revision, uncommitted state, exact allowed files and content hash of the transferred snapshot. Missing revision or hash is unknown, never a fabricated value.
• requested: provider and model selected before launch, separately from observed provider/model returned by runtime or provider metadata.
• observed.source: identify the raw artifact and metadata field; observed.verification: runtime_metadata, provider_response, or unknown. Missing metadata remains unknown; a configured model is only requested.
• status: completed / blocked / failed. Include a reason for blocked or failed results. Completion is not proof that every check passed.
• findings: each has file_line, severity, evidence, limitations, and source reviewer. Link to raw artifacts; identify snapshot-relative locations.
• checks: each has the exact command, actual integer exit_code or null when not run, raw_artifact, and outcome (PASS, FAIL, SKIP, BLOCKED). A nonzero test exit cannot become PASS after output processing. Identify scanner errors separately from violations; do not infer scanner success solely from a summary. Skipped checks include a reason.
• artifacts: local paths to original safe output, findings, and proposed changes; open_questions: unresolved issues; usage: observed token/cost metadata with its source or unknown. Separate parent and worker usage.

Self-description such as “I am the requested model” is not route evidence. Runtime or provider metadata is not independent proof of a provider's internal execution. Preserve that limitation when reporting verification.

The portable scripts/worker-evidence.py helper implements consistency checks for this contract; see Worker Evidence. Captured snapshot entries use path, sha256, and bytes; findings use the string key source_reviewer. Artifact paths are relative to an explicit private artifact root. Usage is unknown or {duration_ms, tokens, cost_usd, method, source} with null for each unknown metric. Preserve contract_status separately from the result outcome, and filesystem SKIP when source/artifact roots were not checked. The helper checks recorded consistency, not metadata authenticity or sandboxing; the launcher remains responsible for independently recording execution facts.

Workflow

1. Validate result scope, snapshot freshness, requested versus observed route, and raw checks; quarantine malformed or contradictory records as failed.
2. Group findings by severity and affected behavior.
3. Deduplicate overlapping findings.
4. Preserve the strongest evidence and confidence level.
5. Separate confirmed issues from plausible/unverified risks.
6. Convert repeated themes into concrete follow-up actions. The main agent reviews proposed artifacts and owns applying changes.

Output

• findings ordered by severity
• duplicates collapsed with source reviewers noted
• open questions
• recommended minimal fixes
• validation gaps
• next skill to use, if any

Return a concise summary and artifact pointers, initially targeting at most 1500 tokens when material findings fit. Keep original event streams and logs outside tracked files; do not discard critical errors to meet the target.

Rule

Only when /quorum is active for this task, keep these original worker records and use the separately bound target, fixed roster and ballots described in Worker Quorum. Verify current source/artifact roots; preserve disagreement and veto evidence instead of averaging confidence. Unknown execution facts or missing verification cannot become an approval.

Without /quorum, merge findings without votes or majority decisions. /multimodel alone does not enable voting; ordinary same-model parallel work activates neither mode. Follow the task-scoped command contract.

Do not average confidence across reviewers. Prefer the most concrete evidence and mark disagreement explicitly.