synapse-manager（突触管理器）

Synapse Manager

Orchestrate multi-agent work with structured delegation, monitoring, and quality gates.

When to Use

• Coordinating 2+ agents on related subtasks
• Monitoring progress of delegated work
• Verifying agent outputs (tests, file changes, integration)
• Sending targeted feedback with error details and fix guidance
• Orchestrating cross-review between agents
• Implementing a multi-phase plan (3+ phases or 10+ file changes)
• Planning agent assignment for multi-file changes

Workflow (7 Steps)

Step 1: Plan, Spec, and Setup

Prefer test-first when the scope is clear: create tests -> present/confirm spec -> then implement. For exploratory work, prototypes, or trivial fixes, writing tests after implementation is acceptable.

Edit plugins/synapse-a2a/skills/ first, then sync generated copies with sync-plugin-skills.

Check existing agents before spawning — reuse is faster (avoids startup overhead, instruction injection, and readiness wait):

synapse list --json

Review WORKING_DIR, ROLE, STATUS, TYPE. Only READY agents can accept work immediately. Other relevant statuses: SENDING_REPLY (brief outbound A2A send/reply POST; wait for the previous status to return), PROCESSING (busy), WAITING (permission prompt — use synapse approve/synapse reject), WAITING_FOR_INPUT (#538 — task paused asking for input, reply via synapse reply "<message>" or synapse reply "<message>" --to <sender_id>), RATE_LIMITED (#561 — LLM provider rate limit hit; wait for the window to reset before re-sending), DONE (demotes to READY ~10s later), SHUTTING_DOWN (do not send).

Spawn only when no existing agent can handle the task:

synapse spawn claude --worktree --name Impl --role "feature implementation"
synapse spawn gemini -w --name Tester --role "test writer"

Cross-model spawning (Claude spawns Gemini, etc.) provides diverse strengths and distributes token usage across providers, avoiding rate limits.

Wait for readiness using the helper script. Resolve it from the skill root so the command works from any working directory, whether you are in the plugin source or a synced copy:

cd plugins/synapse-a2a/skills/synapse-manager
scripts/wait_ready.sh Impl 30
scripts/wait_ready.sh Tester 30

# Synced copy example
cd .agents/skills/synapse-manager
scripts/wait_ready.sh Impl 30

See references/auto-approve-flags.md for per-CLI permission skip flags.

Step 2: Create Tests and Confirm the Spec

Assign the test/spec task and confirm scope before implementation starts:

synapse send Tester "Write the tests first and confirm the spec for auth module.
- Cover valid login, invalid credentials, token expiry, refresh flow
- Report any scope gaps before implementation starts" --attach synapse/server.py --force --wait

Use --attach to send reference files the agent should study. Use --wait while confirming tests/spec, then --silent or --notify once execution is unblocked.

Step 3: Delegate Implementation and Monitor

After tests/spec are confirmed, delegate the implementation task:

synapse send Impl "Implement auth module — tests/spec are confirmed.
- Add OAuth2 flow in synapse/auth.py
- Follow existing patterns" --attach synapse/server.py --force --silent

synapse list --json                       # AI-safe snapshot
synapse history list --agent Impl         # Completed work

Or use the aggregation script:

cd plugins/synapse-a2a/skills/synapse-manager && scripts/check_team_status.sh

If an agent stays PROCESSING >5 min, send an interrupt:

synapse interrupt Impl "Status update — what is your current progress?" --force

Step 4: Approve Plans

synapse approve <task_id>
synapse reject <task_id> --reason "Use refresh tokens instead of long-lived JWTs."

Step 5: Verify

Run the tests created in Step 2 first for fast feedback, then consider broader regression coverage if the changes touch shared modules or public interfaces:

# New tests first (fast feedback)
pytest tests/test_auth.py -v

# Full regression (catches cross-module breakage)
pytest --tb=short -q

Regression triage — distinguish new breakage from pre-existing issues:

cd plugins/synapse-a2a/skills/synapse-manager && scripts/regression_triage.sh tests/test_failing_module.py -v

• Exit 0 = REGRESSION (your changes broke it) → proceed to Step 6
• Exit 1 = PRE-EXISTING (already broken) → note it and continue

Step 6: Feedback

Concrete, actionable feedback saves iteration cycles:

synapse send Impl "Issues found — please fix:

1. FAILING TEST: test_token_expiry (tests/test_auth.py)
   ERROR: TypeError: cannot unpack non-iterable NoneType object
   FIX: Add None guard at the top of validate_token()

2. REGRESSION: test_existing_endpoint broke
   ERROR: expected 200, got 401
   CAUSE: auth middleware intercepts all routes
   FIX: Exclude health-check endpoints from auth" --force --silent

Save patterns for the team:

synapse memory save auth-middleware-pattern \
  "Auth middleware must exclude /status and /.well-known/* endpoints" \
  --tags auth,middleware --notify

After sending feedback, return to Step 3 (Monitor).

Step 7: Review & Wrap-up

Cross-review catches blind spots — each agent reviews the other's work:

synapse send Tester "Review implementation. Focus on: correctness, edge cases" --force \
  --attach synapse/auth.py --wait
synapse send Impl "Review test coverage. Focus on: missing cases, assertion quality" --force \
  --attach tests/test_auth.py --wait

Final verification and cleanup:

pytest --tb=short -q                      # All tests pass
synapse kill Impl -f && synapse kill Tester -f
synapse list --json                       # Verify cleanup

Killing spawned agents frees ports, memory, and PTY sessions. Orphaned agents may accidentally accept future tasks intended for other agents.

Decision Table

References