planr-operate

Planr Operate — Codex-native workflow

This skill is the primary workflow. The user does not write an orchestration prompt or run machine lifecycle commands. Use only public planr operate commands for user actions and the CLI-returned harness actions internally.

Default invocation

A bare $planr-operate means:

1. Inspect with planr operate inspect --json.
2. Select Codex through normal runtime precedence and keep that binding for the entire cycle. Never invoke another vendor runtime or its commands, assets, or models.
3. If context is absent or stale, run the returned bootstrap harness. Follow bootstrap: inspect the workspace first, produce cited context claims, then present one compact context review only when owner confirmation is genuinely needed. Use Codex's native question tool when it is available; otherwise ask naturally in chat. Never dump a questionnaire.
4. Preview, then start exactly one local cycle. The bare skill invocation is authority for reversible local cycle continuation through reviewable, blocked, or failed; it is not authority for connected research, draft approval, route application, PLAN, SHIP, or external effects.
5. Execute current handoff.next[].argv actions only. For each harness.record, dispatch the named Codex role with its mandate and advisor procedure. Use native Codex subagents when exposed; otherwise run the same role sequentially in this Codex session. Parallelize independent advisors when safe, and record each result the instant that role returns — one role at a time against the shared session, but never waiting on a sibling first. The handoff exposes a record action for every pending role, so a slow lens cannot strand a finished one. If a role is still running as the lease window approaches, issue harness heartbeat to renew the session rather than let it expire. If a dispatched lens genuinely never returns — it exceeded its budget even after a heartbeat, or crashed — record it terminal with harness abandon --role <role> --reason "<why>". That lens becomes not_evaluated with your reason, its recorded siblings are untouched, and the Chair names it as an explicit gap it must not synthesize around. Never fabricate a result for a lens that did not return.
6. Run CEO, CTO, CPO, CMO, COO, then Chair. Each role explores the project with Codex tools and current session permissions. Planr grants no extra permissions and supplies no repository JSON evidence body.
7. Submit exactly one [email protected] response per role: flexible analysisMarkdown plus citation-bearing claims, actions, gaps, and conflicts. Uncited opinion may remain narrative but cannot materialize work.
8. Finalize and follow review. Use planr operate report when the handoff requests it, then present the Markdown report and provisional draft paths in chat. The dashboard is optional.
9. Stop at review. Ask separately before approving/discarding drafts, accepting/rejecting findings, applying/rolling back routes, invoking PLAN, invoking SHIP, connected research, or any external effect.