oa-teams-coordination

name: oa-teams-coordination description: "Coordinate multi-agent teams using staging patterns and shared result directories. Use when managing a group of workers writing to shared output, sequencing phases, or using oa team commands. Activates for: oa team, staging area, shared results, phase coordination, L-005." user-invocable: false allowed-tools: Bash(oa *)

Critical Rules

• ALWAYS write worker output to a staging directory first — because merging directly to main files causes race conditions and overwrites between parallel agents (L-005).
• NEVER skip the Verify phase — because agents may produce incomplete or truncated output that silently corrupts the merged result.

Stage → Merge → Verify → Cleanup Pattern (L-005)

Phase 1 — Stage: Workers write to /tmp/staging/<agent-name>/
Phase 2 — Merge: Orchestrator reads all staging files → merges to main files
Phase 3 — Verify: Check completeness, format, cross-references
Phase 4 — Cleanup: Remove staging area after verified merge

Phase Overlap (L-011)

Start the next phase early when there are no inter-phase dependencies:

• Combiner can start when first 2 of 3 researchers are done (if outputs are independent)
• Verify can run per-file as files land in staging (don't wait for all workers)
• Cleanup runs immediately after each verified file

Instructions

1. Create a shared results directory before spawning workers:

   mkdir -p /tmp/shared-results/staging
   

2. Spawn workers, each writing to their own staging subdirectory:

   oa run "Task A. Write output to /tmp/shared-results/staging/worker-a/result.md" \
     --name worker-a --model claude/sonnet --direct
   oa run "Task B. Write output to /tmp/shared-results/staging/worker-b/result.md" \
     --name worker-b --model claude/sonnet --direct
   

3. Monitor workers until done:

   oa status
   

4. Merge staging outputs to main files (orchestrator does this, not an agent):

   cat /tmp/shared-results/staging/worker-*/result.md > /tmp/shared-results/final.md
   

5. Verify completeness before cleanup:

   • Check line count / section headers match expected structure
   • Cross-check agent outputs for consistency

6. Cleanup staging area:

   rm -rf /tmp/shared-results/staging/
   

Team CLI Commands

Patterns

Pattern 1: Full Stage → Merge → Verify → Cleanup

# Stage
mkdir -p /tmp/staging/
oa run "Write analysis to /tmp/staging/worker-a.md" --name worker-a --model claude/sonnet --direct
oa run "Write analysis to /tmp/staging/worker-b.md" --name worker-b --model claude/sonnet --direct

# Wait for completion
oa status  # poll until both are done

# Merge
cat /tmp/staging/worker-a.md /tmp/staging/worker-b.md > /tmp/final-analysis.md

# Verify
wc -l /tmp/final-analysis.md  # expect > 0 lines

# Cleanup
rm -rf /tmp/staging/

Pattern 2: Team with shared results dir

oa team create my-team worker-a worker-b
oa task create my-team "Research and compile report"
# Agents share: /tmp/shared-results/ as communication channel

Anti-Patterns

• Bad: Workers writing directly to the same output file — parallel writes corrupt the file.
• Good: Each worker writes to /tmp/staging/<worker-name>/output.md; orchestrator merges.
• Bad: Cleaning up staging before Verify completes — losing the source of truth for re-runs.
• Good: Keep staging until Verify passes; only then cleanup.
• Bad: Waiting for ALL workers before starting Merge — wastes time when outputs are independent.
• Good: Apply phase overlap (L-011) and merge each file as it lands.

References

• Related: oa-orchestration-spawn, oa-prompting-delegation, oa-web-dashboard