operate

OpenPlanr Operate — Claude Code-native workflow

Run the full workflow for the user. They do not write an orchestration prompt or run lifecycle commands. A bare command invocation means run one complete cycle through the review gate.

1. Run planr operate inspect --json.
2. Bind the cycle to claude-code; never invoke Codex or Cursor.
3. When context is absent or stale, execute the returned bootstrap harness and follow procedures/operate/bootstrap.md. Research the workspace before asking questions. Use AskUserQuestion for one compact review of genuine authority decisions. Never dump a questionnaire.
4. Preview and start one cycle, then follow only current handoff.next[].argv. The slash command authorizes reversible local continuation through reviewable, blocked, or failed only.
5. Dispatch the independent operating-<role> agents with their v1.4 mandates and procedures/operate/advisor.md, preferably in parallel. Record each advisor's exact result the instant it returns through planr operate harness record — one role at a time against the shared session, but never waiting on a sibling to return first; the handoff authorizes a record action for every pending role, so a stalled lens can never strand a completed one. Retries must replay identical bytes for the same role and idempotency key. If any role is still running as the lease window approaches, issue planr operate harness heartbeat to renew the session before it expires. If a dispatched lens genuinely never returns — it exceeded its budget even after a heartbeat, or crashed — record it terminal with planr operate harness abandon --role <role> --reason "<why>" so the board can consolidate without it. That lens is marked 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; abandon it honestly instead.
6. Dispatch Chair with procedures/operate/chair.md, finalize, and present the report following procedures/operate/review.md. If the runtime itself stops before abandoning a stalled lens, an operator can reach a reviewable cycle without discarding it: once the session lease has lapsed, planr operate cycles abandon-role <cycleId> --reason "<why>" --yes marks the still-unrecorded lenses not_evaluated and unblocks consolidation of the rest of the board. That --yes is the operator's explicit confirmation of a governed action; never infer it.

Agents inspect the project directly using current Claude Code session permissions. Planr grants no permissions and sends no repository JSON body. Return rich analysisMarkdown plus citation-bearing claims, actions, gaps, and conflicts. Uncited opinion cannot create work.

Local research is automatic. Connected research requires a per-cycle preview and explicit consent. Classify context as observed, inferred, hypothesis, owner-confirmed, or unknown. If the user says “find it from the project,” keep researching; unknown context is a confidence/gap concern, not a blocker.

The binding remains runtime: claude-code, runtimeBinding: required, and crossRuntimeFallback: false. Preserve all handoff binding values exactly. Legacy adapter commands are compatibility-only; new flows use harness.

Qualified actions may create reversible proposed Quick Tasks, Specs, Epics, decisions, or agent artifacts. List them at review, but never approve or execute them. Stop before finding acceptance, route application, draft approval, PLAN, SHIP, deployment, publication, spending, customer contact, credential changes, or destructive work.

When the user requests a public read-only subcommand, perform only it. The dashboard is optional and never the sole report.

Completion is an explicit phase discipline, not a judgement call. A cycle advances A (inspect/bootstrap) → B (cycle start and runtime binding) → C (dispatch and incremental recording) → D (Chair consolidation) → E (report and drafts materialization) → F (review gate and stop). It is done only when the on-disk artifacts of each phase exist: the expected role records and their Markdown, the Chair result, the final report, the actions file, the draft and provenance results, a cycle state of reviewable, blocked, or failed, no leftover scratch, a correct runtime binding, and no unauthorized effect. A cycle that merely started, prepared, launched advisors, or left temporary result files is not complete. planr operate review — the CLI, not this workflow — is authoritative on whether a cycle has reached the review gate: when it reports an earlier phase and the artifacts still missing, that is the truth of where the cycle stands, and the cycle is not yet reviewable.