Orchestration Protocols

Runtime patterns for delegated agents.

Active Steering

Intervene early on:

When redirecting, explain why + how:

"Don't modify libs/data/src/lib/product.ts — shared across features. Add the new query in libs/data/src/lib/reviews.ts."

Sub-agents: Catch problems early (5 min in saves an hour). Background agents: Steer post-hoc — invest in prompt specificity, partition constraints upfront.

Background Agents

Run autonomously in isolated Git worktrees. Reserve for well-scoped tasks >5 min with clear acceptance criteria.

• Spawn: Delegate Session → Background → Select agent → Enter prompt
• Auto-compaction: At 95% token limit; use --resume to continue
• No real-time monitoring: Invest in specific prompts, strict partition constraints, acceptance criteria checklists upfront

Parallel Research Protocol

Spawn multiple research sub-agents in parallel when 3+ independent questions need answers before implementation. Spawn if: ≥3 independent questions AND answers span multiple codebase areas — else handle sequentially.

Spawn Strategy

Prompt template:

Research: [specific question]
Scope: [files/directories to search]
Return: key findings, relevant file paths (with line numbers), patterns, unanswered questions

Result Merge Protocol

1. Collect all results into single context
2. Checkpoint: verify every researcher returned a result (no timeout/error); re-run failures before proceeding.
3. Deduplicate (same file/pattern counts once)
4. Resolve conflicts — specific evidence beats general observations
5. Synthesize into concise context block for implementation prompts
6. Checkpoint: confirm synthesized block covers every original question; mark unanswered as blockers.

Batch Reviews

• Group by domain (UI, data); run fast reviews in parallel for independent outputs
• Review sequentially when outputs share partition boundary
• Combine related artifacts into one panel question when sharing acceptance criteria

Context Compaction

Summarize prior phase output before passing to next agent. Extract: files changed, key decisions, verification (pass/fail), blockers. Discard: raw tool output, reasoning traces, failed attempts.

Template:

### Prior Phase Output
**Phase [N] — [Agent Name] — [Task Title]**
- Files changed: [list]
- Decisions: [key decisions affecting downstream work]
- Verification: [lint ✅ | types ✅ | tests ✅]
- Blockers: [none | list]

Concrete example:

### Prior Phase Output
**Phase 2 — Researcher A — Find usages of `calculateTotal()`**
- Files changed: none (read-only research)
- Decisions: `calculateTotal` lives in `libs/cart/src/lib/total.ts`; new logic should live in `libs/cart/src/lib/discounts.ts`
- Verification: lint ✅ | types ✅ | unit smoke test (cart total) ✅
- Blockers: design question on rounding behavior (see `docs/rounding.md`)

Health & Recovery Reference

Agent Health Monitoring, Error Recovery Playbook, Circuit Breaker tables moved to REFERENCE.md. See REFERENCE.md for thresholds, recovery steps, escalation flows.

CLI examples (spawn & monitor)

Use OpenCastle CLI (npx opencastle or bin/cli.mjs):

opencastle run --file convoy.yml --dry-run
opencastle run --file convoy.yml --verbose
opencastle run --resume
opencastle run --status
opencastle run --retry-failed

Post-run verification (copy-paste checks):

if [ $? -ne 0 ]; then
  echo "opencastle run failed — inspect .opencastle/convoy.log" \
	 && tail -n 200 .opencastle/convoy.log && exit 1
fi

npx opencastle run --status

grep -i "error\|failed" .opencastle/convoy.log || echo "no obvious errors in logs"

Validation & Verification Checkpoints