coding-agent

Coding Agent

Guide-only skill for delegating code tasks to spawned sessions/subagents and coordinating completion.

When to use

• Multi-file features, refactors, migrations, and bugfixes that benefit from isolated workers.
• Parallel tasks that need progress tracking and explicit integration.
• Issue-to-PR workflows where each unit should run in its own session.

Core tools

• sessions_spawn: create focused workers with clear scope, timeout, and cleanup policy.
• subagents: run parallel autonomous tasks when decomposition is stable.
• session_status: monitor worker lifecycle, progress, and failure states.
• sessions_history: inspect outputs/artifacts before integrating.
• process: run build/test/git commands and poll long-running jobs.
• apply_patch: apply deterministic file edits for targeted changes.

Operating pattern

1. Define scope, acceptance checks, and touched paths before spawning work.
2. Spawn one worker per independent task with explicit constraints.
3. Track status and collect outputs; rerun failed workers with narrowed prompts.
4. Integrate patches incrementally, then run local validation commands.
5. Prepare final summary with changed files, checks, and open risks.

Safety constraints

• Do not assume PTY or interactive terminal capabilities for skill execution.
• Prefer non-interactive commands and deterministic arguments.
• Require explicit user approval for destructive actions (rm, hard reset, force push, mass delete, irreversible migrations).
• Keep branch/repo hygiene: avoid unrelated edits, never discard user work, and report blocked operations clearly.