This skill should be used after parallel or serial worker execution completes, when the user asks to "merge worker results", "combine agent outputs", "validate parallel execution", "integrate workstream changes", or when multiple workers have produced outputs that need to be combined into the source tree.
Collect, validate, and integrate worker outputs after orchestrated execution completes.
Workers produce isolated changes. Someone must verify each worker stayed within its assigned scope, detect any emergent conflicts, and apply changes in the correct order. This skill closes the loop between execution and delivery.
task-decomposer is available for reference.conflict-detector is available (may be absent for single-workstream DAGs that skipped conflict detection).Gather all worker outputs. Two sources:
Agent-tool workers return structured JSON:
{
"task_id": "a1b2c3d4",
"status": "success|partial|failed",
"summary": "what was done in 1-2 sentences",
"files_modified": ["path/to/file.ts"],
"key_decisions": ["chose X over Y because Z"],
"open_questions": []
}
safe-summon workers produce:
.safe-summon/claude-execution-log.jsonSerial/direct execution (Class B or Class D without orchestrator):
For each worker, extract: task_id, status, files touched, exit code.
For each worker, verify:
likely_paths.If any validation fails, flag the specific worker and issue. Do not silently proceed.
Even with pre-execution conflict analysis, workers can create new conflicts:
Scan for:
Use the dependency graph from the DAG:
For each integration step:
Produce a compact, structured report:
## Merge Summary
### Completed Tasks
- [task_id]: [summary] ([N files changed])
- ...
### Files Changed
- path/to/file.ts (by task_id)
- ...
### Key Decisions
- [decision from worker]
- ...
### Validation Results
- Build: [pass|fail|skipped]
- Lint: [pass|fail|skipped]
- Tests: [pass|fail|skipped]
### Risks
- [any flagged issues]
### Open Questions
- [anything workers flagged for human review]