Auto

/ouroboros:auto

Run the full-quality auto pipeline from a single task description.

Dispatch requirement

This skill must be executed by invoking MCP tool ouroboros_start_auto. Do not manually inspect repositories, run shell commands, query GitHub, edit files, or otherwise emulate the auto pipeline as a substitute. Full auto runs routinely exceed interactive MCP tool-call timeouts, so the background starter is the supported default: it returns job_id and auto_session_id quickly. Retain both. When response.meta.job_observer is present, delegate its read-only wait/result contract to exactly one independent child session. The main session keeps only start and explicit on-demand status responsibility.

If ouroboros_start_auto is unavailable, or if any required job polling/result MCP tool is unavailable, stop and report that the required MCP tool is unavailable. A manual fallback is not an ooo auto run.

If a started auto job later returns detached, blocked, failed, or another auto-session status, report that auto-session status and the tool's blocker. detached is non-terminal tracked background work; surface the job/Ralph handles and keep observing them through the same owner. Do not label a blocked or failed outcome as MCP dispatch failure; dispatch failure means the MCP tool could not be invoked.

If the active runtime routes ooo auto through a background starter such as ouroboros_start_auto, do not stop after returning the job_id. Keep ownership of the conversational UX: retain the returned job_id, auto_session_id, and cursor, then delegate monitoring when the host supports child sessions. Only the fallback path monitors with ouroboros_job_wait / ouroboros_job_status in the main session. The user should not have to poll the job manually.

Usage

ooo auto "Build a local-first habit tracker CLI"
ooo auto --resume auto_abc123
ooo auto "Build a local-first habit tracker CLI" --skip-run
/ouroboros:auto "Build a local-first habit tracker CLI"

ooo auto does not accept a parent Seed ID/path or infer lineage from goal prose. Mentioning an existing Seed in the goal is ordinary context, not an authorized derivative operation. Use ooo evolve for typed parent evolution, or ooo run to execute an existing immutable Seed. Do not claim preserved lineage or AC edit scope from an ooo auto goal alone.

CLI flag → MCP arg translation

When the user types ooo auto with CLI-style flags inside chat, translate to MCP arguments before invoking ouroboros_start_auto:

--max-generations is not a flag for ooo auto; it belongs to ooo ralph. The chained Ralph started by the run job is bounded by execution.auto_evolve_max_generations.

--complete-product is deprecated and ignored: the run job owns run → evaluate → ralph, so a single Auto invocation no longer drives Ralph itself. Follow the run job's chain with ooo status or the job tools.

--pipeline-timeout-seconds is accepted only when starting a session. Passing it with --resume is rejected because the original deadline is preserved across process restarts.

Before a fresh Auto start, if the user did not already choose an efficiency policy, first check the persistent default: when execution.default_policy in ~/.ouroboros/config.yaml is efficient or quality_first, do not ask — omit both arguments and the server applies the configured default (the handoff still reports the resolved policy). Otherwise ask in outcome language: Efficient execution maps to adaptive/observe; Quality-first execution maps to quality_first/off. strict assurance is a separate explicit opt-in because it may spend extra work on proof. Never infer strict from the efficiency choice. On resume, do not ask or send either argument; Auto restores the persisted contract.

Behavior

1. Starts an auto session.
2. Runs bounded Socratic interview rounds with source-tagged auto answers.
3. Generates a Seed.
4. Reviews and repairs until A-grade or blocked.
5. Starts execution only after A-grade.
6. Auto reaches COMPLETE as soon as the run has a durable handle. The run keeps going as a background job, and that job carries its own run → evaluate → ralph chain governed by execution.auto_evaluate / execution.auto_evolve (both default true, Ralph bounded by execution.auto_evolve_max_generations). Auto does not evaluate the run itself; the chained evaluate job does.

Preflight blockers and recovery

seed_preflight_unexecutable is a start-new-session boundary, not a plain --resume boundary. The persisted Seed artifact is immutable for that session, so answering an open question cannot silently replace the artifact. Use the displayed questions to revise the goal/Seed contract, then start a new ooo auto session; the blocked session remains available for audit.

Transient interview, evaluator, and lateral-tool exhaustion is different: its resume capability is durable and can be retried with the same session after the external dependency or provider is healthy again. Do not treat a seed_preflight_unexecutable status as a retryable provider outage.

Background monitoring UX

When an auto start response includes response.meta.job_id:

1. Briefly acknowledge that auto started and keep the handles in local state: job_id, auto_session_id / session_id, and cursor from response.meta if present. Show response.meta.dashboard_url when available; otherwise mention ouroboros tui open once as the live view. Tell the user that an observer will post meaningful progress/attention/completion events here and that this conversation remains available for requirement refinement, read-only inspection/review, explicit control, or unrelated isolated work. Include the resolved runtime_backend, llm_backend, efficiency_mode, and frugality_assurance when present. Say that the exact active model and first parallel level will be announced from the first configuration/plan events rather than guessing them.
2. If MCP metadata is unavailable, recover job_observer from the final <!-- ouroboros-job-observer-v1 base64 ... --> content sentinel. Fail closed unless the bounded payload passes canonical v1 validation and its job/session identity matches the visible start receipt. Visible IDs are identity anchors, not a source for reconstructing tools or arguments. Reject validation failure or any mismatch between structured and inline surfaces.
3. If the structured or recovered job_observer is present and the host supports independent child sessions, spawn exactly one read-only observer and pass the contract unchanged. Codex uses spawn_agent, OMP uses one native Task child, and Claude Code uses one Task/Agent child. The observer owns the cursor, waits until terminal, fetches the result, and follows downstream IDs named by follow_result_job_keys. It must not edit files, control execution, or spawn implementation workers. The main session must not poll the same job. On Codex, call spawn_agent exactly once with task_name="run_observer"; a wait call is not a spawn, and the handoff may claim an observer only after the spawn result returns a live child ID/path. Once acknowledged, keep the parent turn open with wait_agent calls of at most 60 seconds while the child remains active. A child send_message only queues a mailbox event; it cannot revive an ended parent turn. Relay meaningful updates and wait again until the observer sends its terminal summary. On OMP, submit exactly one native Task child named RunObserver, require its live agent/job ID, and use the host wait/inbox relay until terminal. User input may interrupt the wait; handle it and resume waiting while observation remains active unless the user asks to stop live observation or replaces the active request. Then end only the relay loop, keep the durable job running, and offer next-turn or explicit- status catch-up. If the observer child fails, is cancelled, or exits before a terminal summary, use that same fallback instead of waiting indefinitely. This parent relay loop must not call job wait/result or take cursor ownership. If spawn fails, do not promise live proactive relays. The detached worker continues after the stdio turn; say that durable progress will be caught up on the next parent turn or an explicit status request. Keep the current turn open in the fallback polling loop only when the user asked for live watching.
4. If child-to-parent progress messages are supported, the observer relays only meaningful phase_changed, progress_advanced, attention_required, and terminal events in at most 1-2 lines. During interview, it may use ouroboros_session_status(session_id=<auto_session_id>) to surface a new pending question or newly answered rounds. Otherwise, keep the main session available and use on-demand status only when the user asks. Surface attention_required immediately when a blocker needs human judgment. Suppress unchanged heartbeats and raw tool output. Interpret execution relays explicitly: run_configuration reports current runtime/harness/model policy; execution_plan reports total ACs, total dependency/parallel levels, and first scheduled ACs; discovery_summary reports bounded targets and purpose; level/routing/harness/verified subtypes report only material changes. Never relay raw commands or reasoning. Before the main session writes to the active auto workspace, check for overlap with worker files or move the unrelated work to an isolated worktree.
5. When no independent child session exists and the user explicitly asks to keep watching in this turn, enter the fallback low-noise loop. Otherwise end the turn safely and resume from the durable cursor on the next interaction:
   • ouroboros_job_wait(job_id=<job_id>, cursor=<cursor>, timeout_seconds=120, view="summary", stream="linked", wait_for="attention_or_ac_change")
   • update cursor = response.meta.cursor after every wait/status response
   • treat response.meta as the source of truth; use response text only as a human-readable hint
6. Relay only meaningful changes: status changes, phase changes, new execution/session/lineage handles, progress counters, blocker/error text, or a terminal state. If response.meta.changed is false, continue silently unless the user asked for heartbeat updates. Synapse delivery events are meaningful: distinguish queued/delivering from applied/completed, and surface rejected/delivery_uncertain immediately. Render the relay in the user's current conversation language; preserve raw event codes only when exact diagnostics help. When the user adds implementation intent during an executing auto run, the main session first reloads Synapse schemas with tool discovery query: "+ouroboros session signal", then calls ouroboros_session_signal_targets with the observed execution_id, selects the semantically matching AC from ac_content and current activity, then sends ouroboros_session_signal with that target's exact IDs. Never ask the user for internal IDs. Ask a short clarification only when multiple live ACs remain genuinely tied, and never route shared goal/AC/ constraint changes to one worker. Send additive implementation refinements with exact target guards, contract_effect="additive", source="user", mode="redirect", and explicit fallback_mode="after_turn". Use mode="inform" for a read-only AC question or assurance request, omit fallback_mode entirely in that mode, and relay the bounded reply from the completed event.
7. During the interview phase, surface the live Q&A — not just the round counter. Whenever the relayed phase is interview (e.g. progress reads interview round N/50), call ouroboros_session_status(session_id=<auto_session_id>) and relay to the user: (a) the current meta.pending_question (the question the interview is asking right now), and (b) the meta.auto_answer_log entries — each is {round, source, question, answer}, i.e. what the auto-answerer answered and why (source: conservative_default = safe-default policy, inference = model reasoning, assumption = auto-answerer fallback). Show this so the user sees what the interview is converging on, not a bare counter. Note: this Q&A lives in the auto-session state, so session_status surfaces it even though ouroboros_query_events returns nothing for the auto session, and it shows only the last 3 answers (each truncated). Keep it low-noise: relay the pending question and any newly answered rounds, not the same 3 entries every poll.
8. If the job status is non-terminal (queued, running, or another active status), keep waiting. Do not tell the user to call job tools themselves.
9. When the job reaches a terminal status, the polling owner calls ouroboros_job_result(job_id) and summarize the final auto-session outcome. If the final auto result is detached, keep tracking the surfaced downstream job/Ralph handles when available instead of presenting detached as completion.
10. If response.meta.status == "delegated_to_plugin" and response.meta.job_id is None, report that OpenCode plugin mode delegated the work to the child Task/session. Do not call job wait/result without a real job id; follow the host Task widget/session lifecycle.

Use short progress relays; the goal is “I am still watching this for you,” not a wall of logs.