pipeline-autopilot

Run the full pipeline autonomously for a work item. Chains stages with context isolation, handles approval gates via interactive interview or auto-approve.

SKILL.md

GitHub

Pipeline Autopilot — Full Autonomous Pipeline Run

You are a pipeline orchestrator. You run the entire pipeline for a work item from its current state to completion, handling each stage in an isolated context and managing approval gates along the way. You are the conductor — the stage skills are the musicians.

Inputs

  • $ARGUMENTS — parsed as: <callsign> [flags]
    • callsign (required): WCP callsign (e.g., SN-3)
    • --from <stage> (optional): Start from a specific stage number (0-7). Default: auto-detect from artifacts.
    • --to <stage> (optional): Stop after a specific stage number (0-7). Default: run to completion.
    • --auto-approve (optional): Enable ludicrous speed mode — auto-approve checkpoints when there are no blocking open questions.

Parse flags from $ARGUMENTS. The callsign is always the first token. Flags follow.

Step-by-Step Procedure

Step 1: Read Work Item and Detect Starting State

wcp_get(CALLSIGN)

Detect the current pipeline stage using the same artifact inspection logic as /pipeline-advance:

1. prd.md missing? → Stage 0
2. No discovery reports (neither discovery-report.md nor any discovery-report-*.md)? → Stage 1
3. architecture-proposal.md missing? → Stage 2
4. architecture-proposal.md pending? → Architecture approval gate
5. gameplan.md missing? → Stage 3
6. gameplan.md pending? → Gameplan approval gate
7. test-coverage-matrix.md missing? → Stage 4
8. Incomplete milestones in progress.md? → Stage 5 (specific milestone)
9. review-report.md missing? → Stage 6
10. qa-plan.md missing? → Stage 7
11. All present → Complete (run /create-pr)

If --from was specified and the detected stage is earlier than the specified stage, skip to the --from stage. If the detected stage is later (artifacts already exist), warn the user but proceed from where the pipeline actually is.

Step 2: Configure Approval Mode

If --auto-approve was NOT passed, ask the user:

Use AskUserQuestion:

  • Question: "How should I handle approval gates (architecture review, gameplan review)?"
  • Options:
    • "Interview me at each gate" — walk through checklist items interactively (Recommended)
    • "Auto-approve if no blockers" — skip review when there are no blocking open questions
    • "Stop at gates" — pause and let me review manually

Save the choice as APPROVAL_MODE.

If --auto-approve was passed, set APPROVAL_MODE to "auto-approve" without asking.

Step 3: Present the Plan

## Pipeline Autopilot: CALLSIGN

**Starting at:** Stage [N] ([stage name])
**Running to:** Stage [M] ([stage name]) [or "completion"]
**Approval mode:** [Interview / Auto-approve / Stop at gates]

### Stages to Run

[List each remaining stage with its name and a one-line description]

Starting now...

Step 4: Pipeline Loop

Execute stages sequentially. Each stage runs in a Task agent for context isolation.

STAGE_ORDER = [
  { num: 0, name: "PRD",            skill: "prd",             gate: false },
  { num: 1, name: "Discovery",      skill: "discovery",       gate: false, per_repo: true },
  { num: 2, name: "Architecture",   skill: "architecture",    gate: true,  artifact: "architecture-proposal.md" },
  { num: 3, name: "Gameplan",       skill: "gameplan",        gate: true,  artifact: "gameplan.md" },
  { num: 4, name: "Test Generation",skill: "test-generation", gate: false, per_repo: true },
  { num: 5, name: "Implementation", skill: "implementation",  gate: false, multi: true, per_repo: true },
  { num: 6, name: "Review",         skill: "review",          gate: false, per_repo: true },
  { num: 7, name: "QA Plan",        skill: "qa-plan",         gate: false },
]

For each stage (starting from the detected or specified stage):

A. Check if we should stop:

  • If --to was specified and we've passed that stage → stop
  • If we're blocked → stop

B. If this stage has a gate (Stages 2 and 3), handle it AFTER the stage runs:

After Stage 2 or Stage 3 completes, handle the approval gate before advancing:

Interview mode:

  • Invoke /pipeline-approve via the Skill tool: Skill(skill="pipeline-approve", args="CALLSIGN")
  • This runs interactively — the user answers questions via AskUserQuestion
  • If approved → continue to next stage
  • If rejected → STOP. Report rejection reason.

Auto-approve mode:

  1. Read the artifact that was just produced
  2. Find the Open Questions section
  3. Check: are there any questions marked as "blocking"? Are there any open questions with no recommendation?
    • If no blocking questions and all have recommendations → auto-approve:
      • Update the Approval Checklist: Status = "Approved", Reviewer = "autopilot", Date = today
      • Update pipeline_approved_at in frontmatter
      • Reattach the artifact via wcp_get_artifact + modify + wcp_attach (you have these tools)
      • Log: wcp_comment(CALLSIGN, "pipeline/autopilot", "Auto-approved [artifact] — no blocking open questions.")
    • If blocking questions exist → fall back to Interview mode:
      • Tell the user: "Auto-approve blocked — [N] blocking open questions need resolution. Switching to interactive review."
      • Invoke /pipeline-approve via Skill tool

Wait — you do NOT have wcp_attach in your allowed tools for this skill. If you need to auto-approve, use a Task agent that has the necessary tools:

Task(
  subagent_type="general-purpose",
  description="Auto-approve pipeline artifact",
  prompt="Read the artifact [artifact-name] from work item CALLSIGN via wcp_get_artifact.
  Update the Approval Checklist section: set Status to 'Approved', Reviewer to 'autopilot', Date to [today].
  Update pipeline_approved_at in YAML frontmatter to [timestamp].
  Reattach via wcp_attach. Log a comment via wcp_comment."
)

Stop-at-gates mode:

  • Report: "Approval gate reached for [artifact]. Pipeline paused."
  • Tell the user to run /pipeline-approve CALLSIGN manually, then /pipeline-autopilot CALLSIGN --from [next-stage] to resume.
  • STOP.

C. Run the stage (for non-gate stages, or after gate is cleared):

Launch a Task agent:

Task(
  subagent_type="general-purpose",
  description="Run pipeline Stage [N] ([name])",
  prompt="You are running a pipeline stage. Invoke the skill:

Skill(skill='[skill-name]', args='CALLSIGN [extra-args]')

After the skill completes, summarize:
1. What artifact was produced
2. Whether it succeeded or had errors
3. Any issues flagged for human attention
4. What the skill said to do next"
)

D. Handle Stage 5 (Implementation) — multi-milestone:

Stage 5 is special: it runs once per milestone. After each milestone:

  1. Re-read progress.md to find the next incomplete milestone
  2. If another milestone exists → run /implementation CALLSIGN [next-milestone] in a new Task agent
  3. If all milestones complete → advance to Stage 6

This means Stage 5 may spawn multiple Task agents sequentially (one per milestone). Each gets a fresh context.

E. Handle per_repo stages (multi-repo projects):

For stages marked per_repo: true, detect whether this is a multi-repo project by checking the work item's artifact list for multiple discovery-report-*.md artifacts.

If multi-repo:

  1. Stage 1 (Discovery): Determine all repos involved (from PRD body or work item). For each repo without a discovery-report-{repo-name}.md artifact, run /discovery CALLSIGN --repo <path> in a Task agent. Run these sequentially (each may need significant context).
  2. Stage 4 (Test Generation): Read the gameplan to identify all target repos. For each repo that doesn't yet have a test-coverage-matrix-{repo-name}.md, run /test-generation CALLSIGN --repo <path>.
  3. Stage 5 (Implementation): When iterating milestones, read each milestone's **Repo:** field from the gameplan. Pass --repo <path> to each /implementation invocation. Milestones are already ordered by cross-repo dependency in the gameplan.
  4. Stage 6 (Review): For each repo that has a progress-{repo-name}.md artifact, run /review CALLSIGN --repo <path>.

If single-repo: Behave as before (no --repo flag).

F. Check for stage failure:

After each Task agent returns, check the result:

  • If the stage reported success → continue to next stage
  • If the stage reported a failure (test failures after 5 attempts, spec gaps, missing prerequisites) → STOP
    • Report the failure
    • Log it as a WCP comment
    • Tell the user what to fix

G. Log progress after each stage:

wcp_comment(
  id=CALLSIGN,
  author="pipeline/autopilot",
  body="Autopilot: Stage [N] ([name]) complete. Advancing to Stage [N+1]."
)

H. Report stage completion to user:

After each stage, print a brief progress update:

--- Stage [N] ([name]): DONE ---
[1-line summary from Task agent]

Step 5: Completion Report

When the loop ends (either all stages complete, blocked, or --to reached):

## Autopilot Complete: CALLSIGN

### Stages Executed
| Stage | Name | Result |
|-------|------|--------|
| 0 | PRD | [Done / Skipped / Failed] |
| 1 | Discovery | [Done (N repos) / Done / Skipped] |
| 2 | Architecture | [Done / Skipped] |
| 2.5 | Architecture Approval | [Approved / Auto-approved / Rejected / Skipped] |
| 3 | Gameplan | [Done / Skipped] |
| 3.5 | Gameplan Approval | [Approved / Auto-approved / Rejected / Skipped] |
| 4 | Test Generation | [Done / Skipped] |
| 5 | Implementation | [M1 Done (wcp-cloud), M2 Done (wcp-ios) / M2 Failed] |
| 6 | Review | [Done (N repos) / Done / Skipped] |
| 7 | QA Plan | [Done / Skipped] |

### Approval Decisions
[List any approval decisions made, with notes captured]

### Open Issues
[List any failures, spec gaps, or blockers encountered]

### Next Steps
- [If complete] "Run `/create-pr CALLSIGN` to push and create PR."
- [If blocked] "[What needs to happen before pipeline can resume]"
- [If --to reached] "Pipeline paused at Stage [M]. Run `/pipeline-autopilot CALLSIGN --from [M+1]` to continue."

Step 6: Final Log

wcp_comment(
  id=CALLSIGN,
  author="pipeline/autopilot",
  body="Autopilot run complete. Stages [N]-[M] executed. Result: [success/blocked/paused]. [Summary of approvals and issues.]"
)

Error Recovery

The pipeline is designed to be idempotent — re-running a stage is safe because:

  • Document stages (0-3) overwrite their artifacts (same content, no side effects)
  • Code stages (4-7) check pre-flight conditions before acting
  • Approval gates check current status before prompting

If autopilot stops mid-run (session timeout, crash, error), just run it again:

/pipeline-autopilot CALLSIGN

It will detect the current state from artifacts and resume from where it left off.

What NOT To Do

  • Do not run stages inline. Every stage MUST run in a Task agent for context isolation. The autopilot's context window must stay small — it's an orchestrator, not an executor.
  • Do not auto-approve when blocking open questions exist. Even in auto-approve mode, blocking questions trigger the interview flow.
  • Do not continue past a blocker. Spec gaps, test failures, and rejection are hard stops.
  • Do not modify artifacts directly (except for auto-approve). Stage skills own their artifacts.
  • Do not skip the progress report. The user needs to see what happened at each stage.
  • Do not run without confirming approval mode. The user must choose (or pass --auto-approve).

When You're Done

The completion report (Step 5) serves as the final output. Make sure it includes:

  1. Every stage that was executed and its result
  2. Every approval decision made (with notes)
  3. Any issues or blockers encountered
  4. Clear next steps

Success Criteria

  • Detects starting stage correctly from artifact inspection
  • Runs each stage in an isolated Task agent
  • Handles approval gates in all three modes (interview, auto-approve, stop)
  • Auto-approve falls back to interview when blocking questions exist
  • Handles Stage 5 multi-milestone iteration
  • Stops gracefully on failures (doesn't continue past blockers)
  • Resumes correctly when re-run after a stop
  • Reports clear progress after each stage
  • Comprehensive completion report with all decisions logged
  • Context window stays bounded regardless of how many stages run