mo-orchestrate-orca

Orchestrate a feature through Orca

Start only when the user names mo-orchestrate-orca; a generic implementation, planning or review request does not activate this lifecycle.

Read Feature lifecycle, Backend contract, Portable review protocol, Orca native mechanics, and Маршрутизация подтверждённой внешней работы, Обратная связь о методологии, and Purpose and architecture contract completely. Resolve one Orca binary and read its non-empty version-matched orchestration and orca-cli guides.

The session the user asked to orchestrate is the orchestrator. An unset orchestrator role does not block the start and is not a reason to ask which model this session should use; --force and another model generation stay forbidden. Keep a role's terminal hot while its proven idle stays under the methodology's threshold, and prefer a fresh session past it.

Read role selections with bundled scripts/mo-models.mjs --show --project <root>; require exact user-approved values and never select a model, effort, placement or posture fallback. Prove provider-native auth, account freshness and one real launch with requested/effective identity equal. Task bytes wait for a publicly proven normal agent prompt, never trust UI or shell.

Use titles <feature>:orchestrator, <feature>:executor, <feature>:review:<vendor> and <feature>:e2e:<n>. Preserve exact Run/task/Dispatch/terminal identities in reasoning, not a state store. Before actors, record project registrations/resources; every actor must remain in the same Orca project and cleanup may touch only exact-owned delta.

This lifecycle also runs from an ordinary shell. One predicate decides placement, the terminal's own handle: ORCA_TERMINAL_HANDLE present and orca terminal read --terminal "$ORCA_TERMINAL_HANDLE" --screen --json readable means inside, and the variable absent or that read refused means outside. Where the shell stands is not a second sign: a directory inside a registered worktree owns no terminal, and a terminal Orca created keeps its handle from any directory. A handle-free terminal read resolves the worktree's focused terminal, not the caller's, and --terminal accepts only a runtime-issued handle, never a relative word; neither decides placement. Then a worktree selector is id:<repo>::<path> or path:<path> and never current or active, the Run is created without --from, waiting is check --run <id> --wait, and no coordinator title is set. If no watchdog sees this session, say once that the limit is accepted — work stops at the limit until a human returns — and continue. Probe the watchdog helper, jq and flock only when the user asks for the watchdog; their absence is then an ordinary typed gap, and starting the watchdog stays a human boundary. Temporary specifications, brief drafts and intermediate coordinator files live in the project's .orca/; where .orca/ is not ignored, write no temporary file into a tracked path and return needs_attention.

A project that cannot answer the readiness questions is not ready, and this skill does not prepare it. Read the knowledge layer of every exact SHA with bundled scripts/mo-knowledge-layer.mjs --candidate <sha> and keep its one line. not_enabled means the project has no knowledge layer to close: drop only the MO-BACKLOG/1 closure proof from G0, GC, G1 and G2. The remote-head equality of G1 and G2, G2's CI evidence and every other gate still apply. needs_attention — a missing MO-BACKLOG/1 command, papercut document or identifier-history gate in a project that had or half-declared them — is reported with the helper's reason and its missing= field as printed, and running the setup skill stays the human's own step. Read the layer of the committed HEAD before intake migration writes anything, so that a needs_attention request ends with no spec, ledger or executor.

With enabled, run G0 through the project's MO-BACKLOG/1 command after intake migration and before substantive work. The one hot executor owns all product/spec edits, regression tests, invariant comments and coherent commits. The orchestrator does not inspect or edit product code. The executor's task asks for the methodology's cleanroom self-review before every handoff: PASS from a subagent that receives only the specification, the diff and the candidate SHA, the same hot subagent for its own findings, and a new one after findings from you or the reviewers. Accept the executor's self_review_unavailable as its harness's typed answer and never as PASS; do not run the self-review yourself.

Use one run-wide blocking waiter: 600000 ms for executor-only and 300000 ms when reviewer/E2E is active. Demultiplex exact handles, process complete event batches, re-arm quiet timeouts without narration, and allow only one same-arm retry after transport failure.

Call mo-review-orca for the first deep pair and remediation follow-ups. Give the executor both immutable reports only through its verified pair paths/sizes; wait for exact Review-Handoff-Ack before cleanup. Keep remediation reviewers hot, and use a fresh final same-SHA pair. Public updates contain only candidate, pair verdict and summed authored P0–P3 census.

Confirmed friction in a Meta-O skill, reference or script is ISS-16: search the repository named by this skill's metadata.repository, comment on a match or create one sanitized Issue, and never write it anywhere else. Without that field there is no addressee, so the observation goes to the human as needs_attention. A worker reports friction to you as one ordinary Methodology-Friction: message; in an eval or E2E run nothing is written outside at all.

For a confirmed out-of-scope technical defect, apply the bundled Issue-routing decision table: route to project/upstream only after verified ownership, search open+closed with explicit complete limits, then reread and redact the final title and every allowlisted field of the private file-safe body with the same closed policy. Use native gh/glab. A technical Issue with established root cause/owner is pre-authorized; credentials, product disputes, subscriptions, irreversible or unknown effects and watchdog remain human boundaries. Never retry an unknown write effect.

Run the project's declared QC in the foreground in a clean checkout of exactly the candidate, naming the command and its exit code, and call mo-e2e for applicable scenarios; reviewers never run tests, linters or QC. After knowledge harvest and removal of the implemented spec, require GC on the exact candidate. Immediately before any agent-owned MR/PR create run G1 and verify remote source head equality. Immediately before merge run G2: the closure proof, the remote head, CI-Coverage/1 covered and a successful run of every covering CI QC job whose printed checkout SHA equals the candidate, then bind the provider write to that head/required policy. Any other coverage outcome, a missing or failed run, another checkout SHA or unsupported integration candidate proof is needs_attention; a locally proven candidate is still delivered, but never merged by the agent without that CI evidence.

Return only when one unchanged full SHA has QC, two fresh vendor-diverse PASS reports, applicable E2E and, with the knowledge layer enabled, an empty GC. Otherwise return honest needs_attention. Never finish with unacknowledged reports, unresolved backlog, foreign cleanup or a required question.

Meta-O calls

• mo-review-orca
• mo-e2e