Charon/.github/instructions/subagent.instructions.md

Subagent Usage Templates and Orchestration

This helper provides the Management agent with templates to create robust and repeatable runSubagent calls.

1. Basic runSubagent Template

runSubagent({
  prompt: "<Clear, short instruction for the subagent>",
  description: "<Agent role name - e.g., Backend Dev>",
  metadata: {
    plan_file: "docs/plans/current_spec.md",
    files_to_change: ["..."],
    commands_to_run: ["..."],
    tests_to_run: ["..."],
    timeout_minutes: 60,
    acceptance_criteria: ["All tests pass", "No lint warnings"]
  }
})

2. Orchestration Checklist (Management)

• Validate: plan_file exists and contains a Handoff Contract JSON.
• Kickoff: call Planning to create the plan if not present.
• Decide: check how to organize work into logical commits within a single PR (size, risk, cross-domain impact).
• Run: execute Backend Dev then Frontend Dev sequentially.
• Parallel: run QA and Security, DevOps and Doc Writer in parallel for CI / QA checks and documentation.
• Return: a JSON summary with subagent_results, overall_status, and aggregated artifacts.

2.1) Multi-Commit Slicing Protocol

• All work for a single feature ships as one PR with ordered logical commits.
• Each commit must have:
  • Scope boundary (what is included/excluded)
  • Dependency on previous commits
  • Validation gates (tests/scans required for that commit)
  • Explicit rollback notes for the PR as a whole
• Do not start the next commit until the current commit is complete and verified.
• Keep each commit independently reviewable within the PR.

3. Return Contract that all subagents must return

{
  "changed_files": ["path/to/file1", "path/to/file2"],
  "summary": "Short summary of changes",
  "tests": {"passed": true, "output": "..."},
  "artifacts": ["..."],
  "errors": []
}

4. Error Handling

• On a subagent failure, the Management agent must capture tests.output and decide to retry (1 retry maximum), or request a revert/rollback.
• Clearly mark the status as failed, and include errors and failing_tests in the summary.
• For multi-commit execution, mark failed commit as blocked and stop downstream commits until resolved.

5. Example: Run a full Feature Implementation

// 1. Planning
runSubagent({ description: "Planning", prompt: "<generate plan>", metadata: { plan_file: "docs/plans/current_spec.md" } })

// 2. Backend
runSubagent({ description: "Backend Dev", prompt: "Implement backend as per plan file", metadata: { plan_file: "docs/plans/current_spec.md", commands_to_run: ["cd backend && go test ./..."] } })

// 3. Frontend
runSubagent({ description: "Frontend Dev", prompt: "Implement frontend widget per plan file", metadata: { plan_file: "docs/plans/current_spec.md", commands_to_run: ["cd frontend && npm run build"] } })

// 4. QA & Security, DevOps, Docs (Parallel)
runSubagent({ description: "QA and Security", prompt: "Audit the implementation for input validation, security and contract conformance", metadata: { plan_file: "docs/plans/current_spec.md" } })
runSubagent({ description: "DevOps", prompt: "Update docker CI pipeline and add staging step", metadata: { plan_file: "docs/plans/current_spec.md" } })
runSubagent({ description: "Doc Writer", prompt: "Update the features doc and release notes.", metadata: { plan_file: "docs/plans/current_spec.md" } })

This file is a template; management should keep operations terse and the metadata explicit. Always capture and persist the return artifact's path and the changed_files list.