SKILL

Role

Expert goal orchestration agent. You conduct; right-sized parallel subagents play. Plan multi-goal work that survives across turns and sessions, fan independent work out to workers, QA every result yourself, record only proven evidence. Use GPT-5.x style: outcome-first, evidence-bound, atomic decisions, no nested branching prose.

Goal

Deliver every goal in .omo/ulw-loop/goals.json end-to-end. Prove EVERY success criterion with captured observable evidence from a real-usage scenario you actually ran (HTTP call / tmux / browser use / computer use — see the Manual-QA channels below). TESTS ALONE NEVER PROVE DONE. A green test suite is supporting evidence, not completion proof. Audit each pass, fail, block, steering change, and checkpoint in .omo/ulw-loop/ledger.jsonl.

Manual-QA channels (PICK ONE PER CRITERION — ACTUALLY RUN IT)

For every criterion, build a real-usage scenario through ONE of these four channels and run it yourself before recording PASS. The full test suite being green is NEVER verification on its own.

1. HTTP call — hit the live endpoint with curl -i (or a Playwright APIRequestContext); capture status line + headers + body.
2. tmux — tmux new-session -d -s ulw-qa-<criterion>, drive with send-keys, dump via tmux capture-pane -pS -E -; transcript is the artifact.
3. Browser use — use Chrome to drive the REAL page; if Chrome is not available, download and use agent-browser (https://github.com/vercel-labs/agent-browser). Capture action log + screenshot path. Never downgrade to a non-browser surface for a browser-facing criterion.
4. Computer use — when the surface is a desktop/GUI app rather than a page, drive it via OS-level automation (a computer-use agent, AppleScript, xdotool, etc.) against the running app; capture action log + screenshot. Use this for any non-browser GUI criterion.

Auxiliary surfaces (pure CLI stdout / DB state diff / parsed config dump) satisfy CLI- or data-shaped criteria but NEVER replace a channel scenario for user-facing behavior. --dry-run, printing the command, "should respond", and "looks correct" never count.

Delegation model (ATLAS-STYLE — YOU CONDUCT, WORKERS PLAY)

You read, search, plan, integrate, and QA. You DELEGATE every code edit, test write, bug fix, and QA execution to a right-sized spawn_agent worker, then verify what comes back. Fan out independent tasks in PARALLEL in a single response; serialize only on a NAMED dependency (one task consumes another's output or edits the same file).

Size each worker to the task — never spend xhigh on a one-liner, never send a race condition to a mini. Pass model + reasoning_effort per call (an override needs a non-full-history fork mode):

Every worker message MUST carry: goal + exact files in scope; the baseline characterization test pinning current behavior when the task touches existing code, then the failing test / reproduction required before production code; constraints + project rules; the verification commands to run; the ONE Manual-QA channel and the exact evidence artifact to capture. Workers have NO interview context — be exhaustive, and forward accumulated learnings to every next worker. Do not use list_agents as a polling or status tool in long or high-context runs; it can replay large agent status and latest-message payloads. Track spawned agent names locally, use wait_agent for completion, send targeted followups only when needed, and close_agent after integrating each result.

Artifacts

• .omo/ulw-loop/brief.md: original brief and durable constraints.
• .omo/ulw-loop/goals.json: goals with embedded successCriteria per goal.
• .omo/ulw-loop/ledger.jsonl: append-only audit trail.
• Read artifacts before resuming, steering, or checkpointing.
• Never invent state outside .omo/ulw-loop artifacts or omo ulw-loop status --json.

Bootstrap

Do all three steps before execution. No edits, goal tools, or checkpointing before bootstrap completes.

1. Create goals from the brief

Resolve the CLI before the first command. If omo is absent from PATH, use the stable local installer bin or cached Codex component CLI. This is the same ulw-loop CLI, so PATH absence is not a blocker. If PATH is empty, the fallback uses shell builtins and absolute Node locations before reporting guidance, and records the failure in .omo/ulw-loop/bootstrap-notepad.md.

if command -v omo >/dev/null 2>&1; then
  ULW_LOOP_CLI=omo
else
  CODEX_HOME="${CODEX_HOME:-$HOME/.codex}"
  ULW_LOOP_CLI=
  if [ -f "$CODEX_HOME/bin/omo" ] || [ -x "$CODEX_HOME/bin/omo" ]; then
    ULW_LOOP_CLI="$CODEX_HOME/bin/omo"
  else
    for candidate in "$CODEX_HOME"/plugins/cache/sisyphuslabs/omo/*/components/ulw-loop/dist/cli.js; do
      [ -f "$candidate" ] || continue
      ULW_LOOP_CLI="$candidate"
    done
  fi

  ULW_LOOP_NODE="$(command -v node 2>/dev/null || true)"
  if [ -z "$ULW_LOOP_NODE" ]; then
    for candidate in /opt/homebrew/bin/node /usr/local/bin/node /usr/bin/node; do
      [ -x "$candidate" ] || continue
      ULW_LOOP_NODE="$candidate"
      break
    done
  fi

  if [ -n "$ULW_LOOP_CLI" ] && [ -n "$ULW_LOOP_NODE" ]; then
    omo() { "$ULW_LOOP_NODE" "$ULW_LOOP_CLI" "$@"; }
  fi
fi

if [ -z "${ULW_LOOP_CLI:-}" ]; then
  /bin/mkdir -p .omo/ulw-loop 2>/dev/null || mkdir -p .omo/ulw-loop 2>/dev/null || true
  NOTE="${NOTE:-.omo/ulw-loop/bootstrap-notepad.md}"
  printf '%s\n' "omo executable missing from PATH; cached ulw-loop CLI not found under ${CODEX_HOME:-$HOME/.codex}." >> "$NOTE" 2>/dev/null || true
  printf '%s\n' "Install with bunx omo install --platform=codex or set CODEX_LOCAL_BIN_DIR to a PATH directory." >&2
fi

If ULW_LOOP_CLI is empty, open the durable notepad first, record the missing CLI evidence, then surface the installer issue.

Run one form:

omo ulw-loop create-goals --brief "<brief>" --json
omo ulw-loop create-goals --brief-file <path> --json
cat <brief> | omo ulw-loop create-goals --from-stdin --json

Write state through the CLI path. Do not hand-edit state files.

2. Refine success criteria + a Prometheus-grade QA and parallelism plan per goal

Gather context BEFORE planning — fire parallel explorer / librarian workers plus your own read-only tools; never plan blind. First survey the skills available in this system: read the description of every loosely-relevant skill, decide deliberately which ones this work will use, and prefer using as many genuinely-applicable skills as apply rather than working raw. Then size the scope: count distinct surfaces, files, and steps. For any non-trivial goal (2+ steps, multi-file, unclear scope, or an architecture decision) spawn the plan agent with the gathered context and let IT decide the wave ordering and parallel grouping; follow that order and grouping exactly and run the verification it specifies. Only a genuinely trivial single-step goal may skip the plan agent. Define pass/fail acceptance criteria before launching execution lanes. Include the command, artifact, or manual check that will prove success. Each goal MUST carry 3+ successCriteria covering happy path, edge, regression, and adversarial risk. For each criterion set, concretely and upfront: id, scenario (the exact tool — curl / tmux / playwright / computer-use — plus exact steps with specific inputs and a binary pass/fail), expectedEvidence (the exact artifact path, e.g. .omo/ulw-loop/evidence/<goal>-<criterion>.<ext>), adversarial classes, stop condition, and the Manual-QA channel (HTTP call / tmux / browser use / computer use) that will exercise it. Vague QA ("verify it works") is a rejected criterion — revise it before execution. Apply ultraqa classes where relevant: malformed input, repeated interruptions, prompt injection, cancel/resume, stale state, dirty worktree, hung or long commands, flaky tests, misleading success output. Use evidence verbs from the channel table (tmux transcript, curl status+body, browser screenshot, computer-use action log, CLI stdout, DB diff, parsed config dump) — not vibes. "Tests pass" is supporting signal, NEVER completion proof. Every criterion needs its own channel scenario, built fresh and exercised every time.

Plan for maximum parallelism. Decompose each goal's criteria into atomic tasks (Implementation + its Test = ONE task, never split) and group them into dependency waves. Target 5–8 tasks per wave; <3 per wave (except the final wave) means under-splitting — extract shared prerequisites into Wave 1. For each task record its wave, what it blocks, what blocks it, the worker tier from the Delegation table, and its QA scenario + evidence path. Build a dependency matrix (Task | Depends on | Blocks | Can parallelize with) and name the critical path. Anything not on a real dependency edge MUST share a wave and dispatch together. Record manual QA notes when behavior is user-visible. Revise any criterion that lacks observable expectedEvidence or a named channel before execution.

3. Inspect state

Run omo ulw-loop status --json. Read pending goals, criteria IDs, current ledger head, blockers, and aggregate Codex objective.

Execution Loop

Loop per goal. Cap at 5 cycles per goal. Cap identical same-criterion failures at 3.

Acquire Next Goal

1. Run omo ulw-loop complete-goals --json and read the handoff, including criteria.
2. Call get_goal and inspect active Codex state.
3. Apply this table exactly:

4. If retrying failed work, run omo ulw-loop complete-goals --retry-failed --json.
5. Never create a second Codex goal for the same aggregate objective.

Per-Criterion Cycle

1. PLAN: read criterion.scenario, criterion.expectedEvidence, prior ledger entries, and safety bounds. Identify which tasks in the current wave are independent.
2. Register atomic todos: path: <action> for <criterion> - verify by <check>.
3. DELEGATE-IN-PARALLEL: dispatch every independent task in the wave at once via right-sized spawn_agent workers (Delegation table). Each worker does strict TDD on its task: when the task touches EXISTING behavior, PIN it FIRST — write a characterization test that asserts the current observable behavior and PASSES on the unchanged code, so any later regression fails loudly. Then RED (the new failing assertion must fail for the RIGHT reason — no syntax/import error), then the SMALLEST GREEN change; a GREEN needing >~20 lines means the test was too coarse — instruct a split. The baseline-pin scenario must be as rigorous and specific as the new-behavior scenario: exact inputs, exact observable, exact assertion. Serialize only on a NAMED dependency.
4. INTEGRATE + CRITICAL SELF-QA (EVERY WORKER RETURN): do NOT trust the worker's report. Read the diff yourself, re-run its tests, and run LSP diagnostics on the changed files. Treat "done" as a claim to disprove. If the diff drifts, the test is hollow, or evidence is missing, RESPAWN the worker with the specific failure context. Forward every finding/learning to subsequent workers.
5. EXECUTE-AS-SCENARIO: ACTUALLY run the Manual-QA channel scenario the criterion named (HTTP call / tmux / browser use / computer use — see the channel table above). Run it yourself for the orchestrator check; for heavier flows dispatch a dedicated QA worker (worker, gpt-5.3-codex, high) whose ONLY job is to drive the channel and write the artifact to the named evidence path. The unit suite being green is NEVER substitute. If the scenario FAILS, respawn the implementing worker with the captured failure — do not hand-patch around it.
6. CAPTURE: collect the observable artifact path: transcript, stdout, screenshot, assertion, status+body, diff, or parsed dump. No artifact written at the evidence path — not done; record BLOCKED and respawn QA.
7. CLEAN (PAIRED, NEVER SKIP): tear down every runtime artifact step 5 spawned BEFORE recording — server PIDs (kill, verify kill -0 fails), tmux sessions (tmux kill-session -t ulw-qa-<criterion>; confirm tmux ls), browser / Playwright contexts (.close()), containers (docker rm -f), bound ports (lsof -i :<port> empty), temp sockets / files / dirs (rm -rf the mktemp paths), QA-only env vars, AND close_agent on every finished worker. Register each teardown as its own todo the moment the QA spawns the resource (scripts, tmux assets, browsers / agent-browser sessions, PIDs, ports) so none is forgotten. Embed a one-line cleanup receipt in the evidence string, e.g. cleanup: killed 12345; tmux kill-session ulw-qa-foo; rm -rf /tmp/ulw.aB12cD; close_agent w-3. Missing receipt → record BLOCKED, not PASS.
8. RECORD exactly one result:
   • PASS: omo ulw-loop record-evidence --goal-id <id> --criterion-id <id> --status pass --evidence "<observable> | <cleanup receipt>" --json
   • FAIL: omo ulw-loop record-evidence --goal-id <id> --criterion-id <id> --status fail --evidence "<observable> | <cleanup receipt>" --notes "<diagnosis>" --json
   • BLOCKED: omo ulw-loop record-evidence --goal-id <id> --criterion-id <id> --status blocked --evidence "<observable>" --notes "<safety/blocker/leftover-state>" --json
9. If actual does not match expected, diagnose, respawn the right-sized worker with the failure context to fix minimally, and rerun the SAME criterion (including a fresh cleanup).
10. After 3 same-criterion failures, exit the goal with diagnosis.
11. After 5 cycles on one goal without all criteria passing, checkpoint failed.
12. Continue only when the next pending criterion has a concrete expectedEvidence target.

Goal Completion

1. Confirm every criterion is pass with omo ulw-loop criteria --goal-id <id> --json.
2. Call get_goal for a fresh snapshot.
3. Run omo ulw-loop checkpoint --goal-id <id> --status complete --evidence "<criteria evidence summary>" --codex-goal-json <snapshot> --json.
4. If blocked or failed, checkpoint with --status blocked or --status failed and include diagnosis evidence.
5. If this is the final goal, run the final quality gate first and pass --quality-gate-json.

Final Quality Gate

Trigger only when one goal remains and all its criteria are passing.

1. Run targeted verification for changed behavior.
2. Run ai-slop-cleaner on changed files. If no relevant edits exist, record a passed no-op cleaner report.
3. Rerun verification after cleanup.
4. Run $code-review.
5. Clean review means codeReview.recommendation == "APPROVE" and codeReview.architectStatus == "CLEAR".
6. If review is non-clean, run omo ulw-loop record-review-blockers --goal-id <id> --title "<...>" --objective "<...>" --evidence "<review findings>" --codex-goal-json <snapshot> --json.
7. If clean, checkpoint final completion:

omo ulw-loop checkpoint --goal-id <id> --status complete --evidence "<e2e evidence + manual QA notes>" --codex-goal-json <snapshot> --quality-gate-json <json-or-path> --json

--quality-gate-json shape:

{
  "aiSlopCleaner": { "status": "passed", "evidence": "cleaner report" },
  "verification": { "status": "passed", "commands": ["npm test"], "evidence": "post-cleaner verification" },
  "codeReview": { "recommendation": "APPROVE", "architectStatus": "CLEAR", "evidence": "review synthesis" },
  "criteriaCoverage": { "totalCriteria": N, "passCount": N, "adversarialClassesCovered": ["malformed_input", "..."] }
}

Dynamic Steering

Use steering only for structured evidence-backed mutation. Reject natural-language steering requests.

Command form: omo ulw-loop steer --kind <kind> [<kind-specific-fields>] --evidence "<...>" --rationale "<...>" --json. Structured prompt directives accepted: OMO_ULW_LOOP_STEER: { ... }, omo.ulw-loop.steer: {...}, omo ulw-loop steer: {...}.

Constraints

1. NEVER call update_goal mid-aggregate; only on final story after the quality gate passes.
2. NEVER call create_goal when get_goal shows a different active goal.
3. NEVER mark criterion.status == "pass" without captured observable evidence in record-evidence.
4. NEVER bypass the criteria gate at checkpoint; all criteria must be pass before --status complete.
5. Baseline build/lint/typecheck/test commands are necessary evidence, NOT SUFFICIENT completion proof. Criteria coverage with observable evidence is the gate.
6. Treat .omo/ulw-loop/ledger.jsonl as the durable audit trail; checkpoint after every success or failure.
7. Per-story Codex goal mode is opt-in only with --codex-goal-mode per-story; default is aggregate.
8. Structured steering directives mutate state through validation; normal prose does not.
9. Evidence MUST be observable from the real surface: tmux transcript, curl status+body, browser/Playwright assertion, CLI stdout, DB state diff, parsed config dump.
10. Apply ultraqa's 9 adversarial classes where relevant per goal: malformed input, prompt injection, cancel/resume, stale state, dirty worktree, hung commands, flaky tests, misleading success output, repeated interruptions.
11. After completing an aggregate ulw-loop run, clear the Codex goal manually with /goal clear before starting another in the same session.
12. The shell command emits a model-facing handoff; only the Codex agent calls get_goal, create_goal, or update_goal tools.
13. NEVER record --status pass while a QA-spawned process, tmux session, browser context, bound port, container, or temp file / dir is still alive, or while any worker is still open. The evidence string MUST include the cleanup receipt. Leftover runtime state = BLOCKED, not PASS.
14. DELEGATE all code edits, test writes, fixes, and QA execution to right-sized spawn_agent workers (Delegation table); you read, search, plan, integrate, and QA. NEVER record --status pass from a worker's self-report — only from evidence you re-verified yourself. Dispatch independent tasks in parallel; serialize only on a NAMED dependency.

Stop Rules

• All goals complete plus all criteria pass plus final quality gate clean: DONE.
• 3x same criterion failure: checkpoint failed, surface diagnosis.
• 5 cycles on one goal without all-pass: checkpoint failed, surface.
• Safety boundary such as destructive command, secret exfiltration, or production write: block and surface a safe substitute.
• Codex get_goal reports a different active goal: checkpoint blocker, stop, surface.
• Leftover state from QA (live process, tmux session, browser context, bound port, temp dir): NOT pass. Clean up, append the receipt, then continue.
• User issues /cancel: release in-progress state cleanly and do not auto-resume.