subagent-dev

name	subagent-dev
description	This skill should be used when the user asks for "development with subagents", "parallel coding agents", "use pi/opencode/gsd2 workers", "subagent implementation", "delegate coding work", or wants explorer/planner/worker/tester/reviewer agents coordinated from Codex.
version	0.1.0

Subagent Development

Coordinate implementation work through external subagents while keeping Codex as the integrator. Use codex-subagent for execution, isolation, logs, and inspection. Keep write-capable work isolated by default.

Contract

• Use context-pack/v1 for each bounded task.
• Use role-card/v1 for each worker role.
• Require subagent-result/v1 structured output for subagent reports.
• Use --isolate worktree for write-capable agents.
• Assign disjoint file ownership to parallel workers.
• Treat subagent output as work product, not truth. Inspect diffs, logs, tests, and events before integration.
• Do not let subagents merge, publish, deploy, or close tasks unless the user explicitly authorized that phase.

For context and role details, read ../references/context-and-roles.md when needed.

Roles

• explorer: read-only repo mapping, dependency discovery, risk scan.
• planner: slices work into bounded tasks with file ownership.
• worker: implements one bounded task in an isolated worktree.
• tester: runs checks and reports failures with reproduction commands.
• reviewer: reviews worker diff against context and acceptance criteria.
• integrator: Codex role; merges or ports changes after inspection.

Workflow

1. Use an explorer for broad unknowns:

codex-subagent context build \
  --subject "Explore implementation surface" \
  --mode dev \
  --goal "Map files and risks; do not edit" \
  --rule AGENTS.md \
  --out explore-context.json

codex-subagent run opencode \
  --context-pack explore-context.json \
  --role-template explorer \
  --background

2. Create a bounded plan with file ownership.
3. For each write worker:

codex-subagent run opencode \
  --context-pack worker-context.json \
  --role-template worker \
  --isolate worktree \
  --background

4. Inspect every worker:

codex-subagent inspect <run-id>
codex-subagent events <run-id>
codex-subagent result <run-id> --structured

5. Review worker diff with a separate reviewer.
6. Integrate only after local verification in the main workspace.

Stop Conditions

Stop and ask the user when:

• file ownership overlaps between workers
• a worker needs unplanned write scope
• tests are absent or cannot run
• runtime output is empty, hung, or off-task
• integration requires merge, deploy, publish, or external side effects