auto

name	auto
description	End-to-end autonomous execution: figure-out → define → do, chained without manual approval gates. Use when you want to define and execute without intervention during planning, when the user asks for autonomous or end-to-end work, says just build it, or asks to tend or babysit a PR.
argument-hint	[task] [--babysit <pr-url>] [--verification per-gate|consolidated|self] [--verifier-model <model>]
user-invocable	true

Chain manifest-dev:figure-out --autonomous (when the transcript lacks shared understanding) → manifest-dev:define --autonomous → manifest-dev:do on a single task. The --autonomous flag on figure-out makes the model self-answer with recommended answers instead of waiting on the user (see figure-out/references/autonomous.md). Surface define's Summary for Approval for visibility but don't wait — treat as approved and proceed to /do.

Task text comes from $ARGUMENTS; if empty, infer from conversation context (summarize the discussed task into a concrete description). Fresh session with no context and no args → halt: No task description provided and no conversation context to infer from. Usage: /auto <task description> | /auto --babysit <pr-url>.

Verification policy. Parse only top-level option uses of --verification and --verifier-model as /auto flags; quoted or topic mentions remain task text. Omitted --verification means consolidated. After resolving that default, load the matching sibling /do reference under ../do/references/ and apply its policy validation before /define; the reference, not /auto, owns mode-specific model support and evidence provenance. Remove parsed flags from the task before /define, and forward them only to /do. Never write either option into the Manifest. Carry the selected mode, model provenance, and the reference's required evidence/provenance wording in the full-chain completion contract.

Babysit mode (--babysit <pr-url>) skips fresh synthesis. Invoke manifest-dev:define with --babysit <pr-url> --autonomous, then /do with the parsed verification options. PR-lifecycle platform auto-detects from PR URL host (github.com → github composition); non-github host → halt. Multi-repo manifest produced by /define → single /do invocation navigates all repos.

Failure handling. /define returns no manifest path → stop, report. /do escalates (BLOCKED criterion or other blocker) → surface the escalation verbatim to the user with the action it requests.

Unattended launch. At the start of a standalone run, before chaining, establish a durable full-chain goal-setting backstop. If the active harness exposes a goal-setting or continuation capability, set the goal directly; otherwise print the copy-pasteable completion contract below. Its completion condition spans the full chain, so continuation does not stop after the first phase. /auto owns this backstop as the chain entrypoint. The terminal success condition is outcome-gated: manifest written, /do reports /done, and every Acceptance Criterion and Global Invariant has fresh PASS evidence under the selected verification mode in a manifest gate ledger. When figure-out --autonomous runs, the parent contract also carries the autonomous Read bar as a phase checkpoint before /define, not as the post-hoc terminal success criterion for a chain whose /do evidence already passes. figure-out --autonomous suppresses its standalone Read-level backstop because this parent owns the checkpoint, /define only emits the manifest handoff, and /do operates under the existing full-chain contract instead of setting or printing a narrower manifest-only goal. Run the autonomous chain for <task> until /define has written the manifest and /do has reported /done with every Acceptance Criterion and Global Invariant carrying fresh PASS evidence under <selected verification mode and provenance> in a manifest gate ledger. If figure-out runs first, complete a full-anatomy Read checkpoint before /define: every load-bearing branch pressed; Evidence Ledger explicit; assumptions separated from verified and inferred claims; independent re-derivation run or explicitly unavailable; rival set no longer moving; confidence, evidence, and overturn conditions stated. For diagnosis-shaped work, the Read checkpoint is not complete if it only localizes where the symptom concentrates: it must name the concrete mechanism — the variable, difference, or sequence that produces the symptom, including why this case differs when the question is comparative — or earn an underdetermined Read by naming the surviving explanations and showing which feasible probes that could distinguish them were run, what they showed, or why they were blocked. Treat a missing or weak Read checkpoint as a phase defect to repair before /define, not as a post-hoc terminal failure after /do has fresh all-gate PASS evidence. During /do, maintain a ledger for every AC/GI with gate id, phase, verify.instructions source, selected verification mode, evaluator provenance, explicit or inherited verifier model, latest verdict, evidence, and freshness relative to the last relevant change to its subject; unverified, FAIL, stale, BLOCKED/actionable, or escalation-pending gates are non-terminal, and a substantive change to a gate's subject after a PASS makes it stale until re-evaluated, while re-reading, re-examining, and cosmetic or no-op edits do not. Drive every phase to its real bar yourself; do not accept unevidenced self-attestation, "looks done", or summary claims in place of the selected mode's required evidence. Escalate only a blocker that genuinely needs me. Record compact progress checkpoints after the Read checkpoint when present, manifest path, verification/repair cycles, and blockers. Stop after N turns if it stalls.