Yolo Skill

Full-auto AI-DLC pipeline. User provides a prompt; agent drives the entire lifecycle: Idea -> Elaboration -> Proposal -> Review -> Execute -> Verify -> Done.

Tool namespace: Chorus tools are exposed by the connected MCP server under a mcp__chorus__ prefix on dsh (e.g. mcp__chorus__chorus_pm_create_proposal). Bare names are used below for readability — prepend mcp__chorus__ when invoking. See chorus for the full rule.

dsh adaptations summarized (details inline below): (1) elaboration is self-answered with no user interaction; (2) reviewers run inline after each submit by loading the exact reviewer through the skill tool in a foreground subagent (run_in_background: false, wait for the verdict), with a read-only self-review fallback; (3) sessions are manual if you dispatch sub-agents; (4) task execution uses sequential main-agent waves.

Overview

yolo-chorus automates the complete AI-DLC workflow. You provide a natural language description of what you want built, and the agent handles everything:

1. Planning -- create project, idea, self-elaboration, proposal with docs & tasks
2. Proposal Review -- proposal-reviewer adversarial loop
3. Execution -- sequential, dependency-ordered task execution by the main agent
4. Verification -- task-reviewer adversarial loop + admin verify
5. Report -- completion summary

/yolo <prompt>
       |
       v
  Project + Idea + Elaboration (self-answered) + Proposal
       |
       v
  Proposal Reviewer (inline, up to maxProposalReviewRounds)
       |
       v
  Admin Approve --> Tasks materialize
       |
       v
  Sequential wave execution (main agent: loop chorus_get_unblocked_tasks)
       |  (implement task + task-reviewer per task)
       v
  Admin Verify each task --> unblock next
       |
       v
  Done. Report summary.

Escape hatch: interrupt at any time. All created entities (project, idea, proposal, tasks) persist in Chorus. Resume manually via develop-chorus or review-chorus.

Prerequisites

The API key needs write + admin on every resource it touches:

Check at startup:

perms = chorus_checkin().agent.permissions
need = { idea: ["write"], proposal: ["write","admin"],
         task: ["write","admin"], project: ["write"] }

for resource, actions in need:
  missing = [a for a in actions if a not in (perms[resource] or [])]
  if missing: ABORT "/yolo needs {resource}: {missing}. Use an Admin-preset API key."

Input

/yolo <natural language prompt>
/yolo <prompt> --project <project-uuid>

• <prompt> -- what you want built (becomes the Idea content)
• --project <uuid> -- optional; use an existing project instead of creating a new one

Workflow

Phase 1: Planning

Step 1.1: Resolve Project

Parse the arguments for --project <uuid>.

If --project is provided:

chorus_get_project({ projectUuid: "<uuid>" })

Verify it exists and proceed.

If not provided, search for a suitable existing project first:

# 1. Search for projects matching the prompt topic
chorus_search({ query: "<key terms from prompt>", entityTypes: ["project"] })

# 2. Or list recent projects to find a match
chorus_list_projects()

Review the results. If a project clearly matches the user's intent (same topic, active, relevant scope), use it. If no suitable project exists, create a new one:

chorus_admin_create_project({
  name: "<short title derived from prompt>",
  description: "<1-2 sentence summary of the prompt>"
})

Step 1.2: Create Idea

chorus_pm_create_idea({
  projectUuid: "<project-uuid>",
  title: "<concise title derived from prompt>",
  content: "<full user prompt as-is>"
})

Then claim it:

chorus_claim_idea({ ideaUuid: "<idea-uuid>" })

Step 1.3: Self-Elaboration

In /yolo mode, the agent generates elaboration questions and answers them itself -- no user interaction at all. When CHORUS_DAEMON_HEADLESS=1, ask_user_question is prohibited, and yolo deliberately does not prompt the user; it self-answers to preserve an audit trail without interrupting the run.

Self-elaboration is still a loop. If answering your own questions surfaces a new question, contradiction, or gap, loop back to chorus_pm_start_elaboration for another self-answered round before resolving — don't force a resolve over unresolved ambiguity. There is no human gate in YOLO, so the loop exits on your judgment that nothing material is left open (round cap 10). Steps 1–2 are one round; repeat them as needed, then resolve once in Step 3.

1. Generate and submit questions:

   chorus_pm_start_elaboration({
     ideaUuid: "<idea-uuid>",
     depth: "standard",
     questions: [
       {
         id: "q1",
         text: "<question about scope, architecture, etc.>",
         category: "functional",
         options: [
           { id: "a", label: "<option A>" },
           { id: "b", label: "<option B>" }
         ]
       }
       // ... 5-8 questions covering functional, technical, scope aspects
     ]
   })
   

2. Answer immediately (agent selects best options based on the prompt — no user prompt):

   chorus_answer_elaboration({
     ideaUuid: "<idea-uuid>",
     roundUuid: "<round-uuid>",
     answers: [
       { questionId: "q1", selectedOptionId: "a", customText: "Rationale: ..." },
       // ...
     ]
   })
   

3. Resolve — in YOLO mode the agent resolves elaboration autonomously, with no human-confirmation gate (the human-confirmation requirement that applies to the interactive idea-chorus flow is explicitly waived under yolo-chorus automation):

   chorus_pm_validate_elaboration({
     ideaUuid: "<idea-uuid>"
   })
   

   chorus_pm_validate_elaboration requires idea:admin. yolo-chorus already mandates an Admin-preset key in Prerequisites, so this is satisfied. To open another self-elaboration round instead of resolving, just call chorus_pm_start_elaboration again.

Step 1.4: Create Proposal

1. Detect OpenSpec mode (inline). Load the openspec-aware-chorus skill and run its §1 inline three-check detection (CHORUS_OPENSPEC_MODE != "off", an openspec/ directory at the project root, and the openspec CLI on PATH).

   dsh note: there is no Claude Code SessionStart hook to precompute CHORUS_OPENSPEC_ACTIVE. You must run the three checks yourself, inline, here. This is mandatory — yolo runs unattended, so silently picking the wrong mode is exactly the failure scenario the detection contract exists to prevent.

   • All three checks pass → spec-driven branch (sub-step 2a below).
   • Any check fails (or CHORUS_OPENSPEC_MODE=off) → free-form branch (sub-step 2b below).

2. Create the empty proposal container. In OpenSpec mode, the description MUST contain the literal line OpenSpec change slug: <slug> (use the $SLUG you'll pick in 2a); in free-form mode, omit that line.

   chorus_pm_create_proposal({
     projectUuid: "<project-uuid>",
     title: "<feature name>",
     description: "<summary>\n\nOpenSpec change slug: <slug>",   // OpenSpec mode
     // description: "<summary>",                                 // free-form mode
     inputType: "idea",
     inputUuids: ["<idea-uuid>"]
   })
   

   Then branch:

   2a. OpenSpec mode (all three checks pass). Follow openspec-aware-chorus §3 end-to-end:

   • Pick $SLUG, run openspec new change "$SLUG" (§3.1–§3.2).
   • Author proposal.md, design.md, and one specs/<capability>/spec.md per capability locally on disk (§3.3). ADDED Requirements only; per-spec fallback to free-form Markdown if MODIFIED/REMOVED is needed.
   • Define json_encode_file, chorus_check_response helpers (§3.4, §6).
   • Mirror each local file via "$CHORUS_MCP_CALL" chorus_pm_add_document_draft "$PAYLOAD" (§3.6) — one call per file, with the document type from openspec-aware-chorus §5.

   ⛔ Do not invoke chorus_pm_add_document_draft / chorus_pm_update_document_draft / chorus_pm_update_document from the MCP harness with a hand-typed content field in this branch. Re-typing the markdown body wastes 20k+ tokens per proposal and breaks byte-equality with the local files. See openspec-aware-chorus §2 Rule 1.

   Then continue to step 3 (task drafts).

   2b. Free-form mode (any check fails). Add a tech design document draft directly via MCP, content authored inline:

   chorus_pm_add_document_draft({
     proposalUuid: "<proposal-uuid>",
     type: "tech_design",
     title: "Tech Design: <feature>",
     content: "<markdown tech design covering architecture, data model, API, module contracts>"
   })
   

3. Add task drafts incrementally (use returned draftUuid for dependency chaining). acceptanceCriteriaItems is required on every draft — at least one non-blank criterion, or the call is rejected:

   # First task
   result1 = chorus_pm_add_task_draft({
     proposalUuid: "<proposal-uuid>",
     title: "<module name>",
     description: "<what to build, referencing tech design>",
     priority: "high",
     storyPoints: 3,
     acceptanceCriteriaItems: [
       { description: "<testable criterion>", required: true },
       // ...
     ]
   })
   
   # Second task, depends on first
   chorus_pm_add_task_draft({
     proposalUuid: "<proposal-uuid>",
     title: "<dependent module>",
     description: "...",
     priority: "medium",
     storyPoints: 2,
     acceptanceCriteriaItems: [...],
     dependsOnDraftUuids: ["<result1.draftUuid>"]
   })
   

4. Validate:

   chorus_pm_validate_proposal({ proposalUuid: "<proposal-uuid>" })
   

   Fix any errors, then proceed.

5. Submit:

   chorus_pm_submit_proposal({ proposalUuid: "<proposal-uuid>" })
   

   Immediately proceed to Phase 2 and run the proposal reviewer inline — dsh has no PostToolUse hook to remind you.

Phase 2: Proposal Review Loop

dsh difference: there is no PostToolUse hook injecting a "spawn the reviewer" reminder. Run the reviewer inline, right after chorus_pm_submit_proposal.

Obtain an independent VERDICT on the proposal:

• Preferred — spawn a reviewer sub-agent (foreground). Use the dsh subagent tool to spawn a sub-agent with run_in_background: false (foreground — the call waits and returns the result inline; the approve/reject decision depends on the verdict) whose task tells it to call the skill tool with proposal-reviewer-chorus, then review the proposal. The authoritative result is the newest VERDICT: comment on the proposal. Set run_in_background: true (a continuable/background sub-agent whose settlement notice you collect later) only when you deliberately want to fan out and don't need the verdict before your next step.

  Load and run the proposal-reviewer-chorus skill to review proposalUuid <uuid>. This is review round <N>. Read the proposal, its documents, the idea, and the elaboration; classify findings as BLOCKER/NOTE; post your VERDICT comment on the proposal when done.

• Fallback — review it yourself. If subagent is unavailable (e.g. spawning disabled by policy), do the review yourself as a focused, read-only pass following the proposal-reviewer-chorus skill's procedure (read proposal + comments + idea + elaboration; check doc completeness, task granularity, AC↔requirement coverage, the DAG, and integration checkpoints; classify BLOCKER/NOTE) and record the result via chorus_add_comment ending with a VERDICT: line. Do not modify drafts during the review pass.

Then:

1. Read the reviewer's VERDICT:

   chorus_get_comments({ targetType: "proposal", targetUuid: "<proposal-uuid>" })
   

   Look for the most recent comment containing VERDICT:.

2. Act on the VERDICT:

   • PASS or PASS WITH NOTES --

     chorus_admin_approve_proposal({
       proposalUuid: "<proposal-uuid>",
       reviewNote: "PASS from reviewer. <brief summary of notes if any>"
     })
     

     Tasks and documents materialize automatically. Proceed to Phase 3.

   • FAIL -- Read the BLOCKERs from the reviewer comment. Then:

     chorus_pm_reject_proposal({
       proposalUuid: "<proposal-uuid>",
       reviewNote: "FAIL from reviewer. Fixing BLOCKERs: <list>"
     })
     

     Revise the drafts (chorus_pm_update_document_draft, chorus_pm_update_task_draft) to address each BLOCKER, then resubmit:

     chorus_pm_submit_proposal({ proposalUuid: "<proposal-uuid>" })
     

     After resubmission, run the reviewer inline again for Round 2 (same as above).

3. Max rounds: Loop up to maxProposalReviewRounds (from plugin config, default 3). If exhausted:

   STOP: "Proposal review failed after {maxRounds} rounds.
          Remaining BLOCKERs: <list>. Human review needed.
          Proposal UUID: <uuid>"
   

4. No new VERDICT comment after a spawned reviewer returns? It exhausted its turn budget. Respawn it ONCE with a concise-budget hint: "Stay within turn budget. Skip deep source verification. Fetch proposal + comments + idea only, skim for obvious BLOCKERs, and post your VERDICT within the first 10 turns." If still no VERDICT, fall back to reviewing manually and post the VERDICT yourself — the pipeline cannot loop forever on a silent reviewer.

Phase 3: Task Execution (Sequential Waves)

After proposal approval, tasks exist in open status. Execute them in dependency-ordered waves.

dsh difference: dsh has no Agent Teams / TeamCreate primitive. Run waves sequentially as the main agent: loop chorus_get_unblocked_tasks, implement each ready task yourself, verify it, then loop again for the next wave. Do NOT call TeamCreate — it does not exist on dsh. (Under the Claude Code plugin, each wave can be dispatched in parallel via TeamCreate; that is a Claude-Code-only optimization that degrades to the sequential loop here.)

wave = 1

loop:
  # 1. Find ready tasks (all dependencies done/closed)
  unblocked = chorus_get_unblocked_tasks({ projectUuid: "<project-uuid>" })

  if no unblocked tasks and all tasks done/closed:
    break  # All complete

  if no unblocked tasks and some tasks not done:
    # Stuck -- tasks failed review and can't proceed
    break with escalation report

  # 2. Implement each unblocked task, in order, AS THE MAIN AGENT:
  for each task in unblocked:
    chorus_claim_task({ taskUuid: task.uuid })
    chorus_update_task({ taskUuid: task.uuid, status: "in_progress" })

    # ... read task + proposal + project documents for context,
    #     write code, run tests ...

    chorus_report_work({ taskUuid: task.uuid, report: "...", status: "to_verify" })
    chorus_report_criteria_self_check({ taskUuid: task.uuid, criteria: [...] })
    chorus_submit_for_verify({ taskUuid: task.uuid, summary: "..." })

    # 3. Proceed to Phase 4 (verification) for THIS task before moving to the next.

  wave += 1

Optional sub-agent dispatch: if your dsh host supports generic worker sub-agents (not Agent Teams), you may hand one task to a sub-agent at a time. Because there is no SubagentStart hook, the worker prompt must include the manual session instructions explicitly — see develop-chorus "Optional: sub-agent dispatch". The main agent still owns review + verification. This does not change the sequential, wave-by-wave structure above.

Phase 4: Verification

After each task is submitted (Phase 3 step 3), verify it before moving on:

for the just-submitted task:
  # 1. Check task status
  task = chorus_get_task({ taskUuid: "<task-uuid>" })

  if task.status != "to_verify":
    # implementation may have failed; handle or skip
    continue

  # 2. Run the task-reviewer INLINE (no hook on dsh):
  #    - Preferred: use the subagent tool to spawn a sub-agent with run_in_background:false (foreground, wait for the verdict) whose
  #      task says: "Call the skill tool with task-reviewer-chorus, verify taskUuid
  #      <uuid> (round <N>), and post the VERDICT comment." Let it finish, then
  #      read the newest VERDICT comment on the task.
  #    - Fallback (subagent unavailable): review it yourself as a focused read-only
  #      pass following the task-reviewer-chorus procedure (read task + proposal + docs + code,
  #      run read-only tests, classify findings BLOCKER/NOTE) and post the VERDICT via
  #      chorus_add_comment.

  # 3. Read task-reviewer VERDICT
  comments = chorus_get_comments({ targetType: "task", targetUuid: "<task-uuid>" })
  # Find the most recent comment containing "VERDICT:"

  # 4. Act on VERDICT — three possible outcomes:
  if VERDICT is "PASS":
    chorus_mark_acceptance_criteria({
      taskUuid: "<task-uuid>",
      criteria: [
        { uuid: "<ac-uuid>", status: "passed", evidence: "<from reviewer>" },
        // ...
      ]
    })
    chorus_admin_verify_task({ taskUuid: "<task-uuid>" })
    # Task is now "done" -- unblocks dependents for the next wave

  if VERDICT is "PASS WITH NOTES":
    chorus_mark_acceptance_criteria({ ... })
    chorus_admin_verify_task({ taskUuid: "<task-uuid>" })

  if VERDICT is "FAIL":
    # BLOCKERs found. Do NOT verify. Reopen for rework.
    chorus_admin_reopen_task({ taskUuid: "<task-uuid>" })
    # Fix the BLOCKERs in a later pass (the task returns to in_progress/open)

After verifying the wave's tasks, return to Phase 3's loop to pick up newly unblocked tasks. Remember: only done (not to_verify) unblocks dependents.

Max rounds per task: Tracked by maxTaskReviewRounds from plugin config (default 3). If a task has been reopened maxRounds times, skip it and flag for human escalation:

ESCALATE: "Task '{title}' failed review after {maxRounds} rounds.
           Last BLOCKERs: <list>. Manual intervention needed.
           Task UUID: <uuid>"

Continue with remaining tasks -- do not halt the entire pipeline for one stuck task.

No new VERDICT comment after a spawned task-reviewer returns? It exhausted its turn budget. Respawn it ONCE with a concise-budget hint: "Stay within turn budget. Skip deep verification. Fetch task/proposal/comments, run only the core tests, and post your VERDICT within the first 12 turns." If still no VERDICT, fall back to reviewing manually and post the VERDICT yourself — do not loop indefinitely.

Phase 4.5: Code-Review Gateway (mandatory pre-ship)

Once every task of the idea's proposal is verified (done) — Phase 3 finds no more unblocked tasks and none remain non-terminal — run the final ship-time code-review gateway before the Phase 5b completion report. It reviews the whole Idea's aggregate code change across all tasks (not a single task) and posts its verdict on the idea. Inline (no hook on dsh), same mechanism as Phase 4:

• Preferred — spawn a reviewer sub-agent (foreground). Use subagent to spawn a sub-agent with run_in_background: false (foreground — the call waits and returns the verdict inline; the ship decision depends on it; do NOT detach — set run_in_background: true only to deliberately fan out) whose task tells it to call the skill tool with code-reviewer-chorus and follow it against the idea. Example task prompt: Load and run the code-reviewer-chorus skill to review the aggregate code for ideaUuid <uuid> (round <N>); post your VERDICT comment on the idea when done.
• Fallback — review it yourself. If subagent is unavailable, perform the review as a focused read-only pass following the code-reviewer-chorus procedure (read the idea, its approved proposals + documents + tasks; infer the aggregate diff from task reports + git log/diff; review cross-task integration, architecture, security, regression, feature-level coverage; run the project build/test) and post the VERDICT: comment on the idea yourself.

Act on the VERDICT:

• PASS / PASS WITH NOTES — the feature is cleared to ship. Proceed to Phase 5 / 5b.
• FAIL — do NOT ship. Read the BLOCKERs, then fix them via the quick-dev workflow (quick-dev-chorus): call chorus_create_tasks with proposalUuid set to the current approved proposal so the fix tasks attach to it — do not reopen the already-verified tasks or apply untracked fixes. Group related small BLOCKERs into one cohesive task by default; split only materially large or independently testable fixes. Drive every fix task through Phase 3 → Phase 4, including AC self-check, independent task review, and admin verification. Re-run the gateway only after every fix task is successfully done; a failed or cancelled fix task, stop the automatic loop and escalate. Loop bounded by maxCodeReviewRounds (default 3; 0 = unlimited).

ESCALATE: "Idea '{title}' failed code review after {maxCodeReviewRounds} rounds.
           Last BLOCKERs: <list>. Manual intervention needed. Idea UUID: <uuid>"

No VERDICT comment after the code-reviewer returns? It exhausted its turn budget (larger than the task-reviewer's because it reviews the whole feature). Respawn ONCE with a concise-budget hint; if still silent, fall back to a manual read-only pass and post the VERDICT yourself.

The gateway is behavioral like the other two reviewers: its verdict is advisory and does not change the Idea's stored status; the orchestrator honors it. It runs before the completion report so the report is never written while a FAIL is outstanding.

Phase 5: Report

After all waves complete, output a markdown summary:

## /yolo Complete

**Project:** <project-name> (<project-uuid>)
**Proposal:** <proposal-title> (<proposal-uuid>)
**Idea:** <idea-title> (<idea-uuid>)

### Tasks
| Task | Status | Review Rounds |
|------|--------|---------------|
| <title> | done | 1 |
| <title> | done | 2 |
| <title> | ESCALATED | 3 (max) |

### Summary
- Total tasks: N
- Completed: X / N
- Escalated: Y (need human review)
- Waves executed: W

Phase 5b: Idea Completion Report (mandatory)

A successful yolo-chorus run always finishes the Idea — call chorus_create_report once with proposalUuid set to the last verified proposal. The content parameter's description carries the section template; follow it. Surface the returned documentUuid in the Phase 5 summary. Skipping is a protocol violation.

Order: write the completion report only after the Phase 4.5 code-review gateway returns PASS / PASS WITH NOTES — never while a code-review FAIL is outstanding.

OpenSpec archive: if you ran in OpenSpec mode (Step 1.4 branch 2a), the last verified task also triggers the archive flow. dsh has no PostToolUse hook to remind you — after verifying the final task, run openspec-aware-chorus §3.9 yourself (openspec archive <slug> --yes, then mirror each emitted openspec/specs/<capability>/spec.md back via §3.8).

Error Handling

Tips

• Keep the initial prompt detailed -- the more context you provide, the better the auto-generated proposal quality
• The proposal-reviewer is your quality gate -- if it keeps FAILing, the prompt may be too vague
• Watch the wave count -- if tasks keep getting reopened, consider stopping and reviewing the feedback manually
• All audit trail is preserved: elaboration Q&A, reviewer VERDICTs, work reports. Check Chorus UI for full history
• For small/simple tasks, consider quick-dev-chorus instead -- it skips the Idea->Proposal overhead
• Sub-agents (if you dispatch any) share your API key; ensure it has the permissions listed in Prerequisites before starting

Next

• To manually review proposals: review-chorus
• To manually develop tasks: develop-chorus
• To create quick standalone tasks: quick-dev-chorus
• For platform overview: chorus