| name | auto |
| description | Use to run the full execute-review-fix-verify pipeline automatically with clean context between steps. Orchestrates sequential subagents for unattended execution. Use when someone says 'auto wave', 'auto milestone', 'run full pipeline', 'automated execution', 'hands-free', 'run everything', 'unattended wave', or 'auto run'. |
| argument-hint | [wave N | milestone N] |
What This Skill Does
Orchestrates the full execute-review-fix-verify pipeline using sequential subagents with clean context. Each step runs as an independent agent that only knows what's in its prompt and what it reads from disk — preventing bias from prior work. The pipeline is fully unattended and resumes from where it left off if interrupted.
Modes
auto wave N — Runs the full pipeline for a single wave: execute all tasks, review, generate fix plan, apply fixes, verify fixes, then merge the feature branch to main and clean up branches.auto milestone N — Loops over all waves in milestone N (in order). For each wave: runs the full wave pipeline, merges, cleans up, then advances to the next wave. Stops if a wave fails.
Prompt
Context Loading
Before spawning any subagent, the orchestrator (you) must build context:
- Read
workflow/plan/PLAN.md — find all waves under the target milestone. Determine milestone number M from task numbering (task X.Y belongs to Milestone X). For auto wave N, identify all tasks in wave N. For auto milestone N, identify all waves and their tasks.
- Read
workflow/plan/PROGRESS.md — check current task statuses.
- Check for existing review artifacts in
workflow/plan/reviews/:
- Task review files:
task-X.Y.md
- Wave review file:
wave-mM-N.md (preferred) or wave-N.md (legacy)
- Check git branch state:
git branch to see which feature/task/fix branches exist.
Artifact-Based Resume
Check existing artifacts to determine where to start. Check in this order:
1. Is the wave already merged?
All tasks in PROGRESS.md are done AND no feature/mM-waveW-* branch exists → wave is complete. Skip it (milestone mode) or report "already done" (wave mode).
2. Does a wave review file exist?
No → start at Step 1 (execute) if task review files are also missing, or Step 2 (review) if task review files exist for all tasks in the wave.
3. Does ## Security Review exist in the wave review?
No → start at Step 2b (security review). The code review is already done (the wave review file exists).
4. Does the wave review have issues?
Read the wave review file. Check ### Issues Found sections under both ## Task X.Y headings (code review) AND ## Security Review (security review). Count BLOCKER and SUGGESTION entries (ignore NITs).
- Zero BLOCKERs and SUGGESTIONs across both reviews → skip fix cycle, go to Step 6 (merge).
- Issues exist → continue checking.
5. Does ### Fix Plan exist in the wave review?
No → start at Step 3 (fix-plan-analysis).
6. Does ### Fix Results exist in the wave review?
No → start at Step 4 (execute fixes).
7. Does ### Fix Verification exist in the wave review?
No → start at Step 5 (verify-fixes).
Yes → read it. If all issues are marked "Fixed" → go to Step 6 (merge). If any are "Not Fixed" or "Regression" → STOP and report (see Failure Handling).
Pipeline Steps
Run steps sequentially. After each step, verify the expected artifact exists before proceeding.
Agent prompt principle: Every subagent prompt below invokes a skill (/agentic-dev:execute, /agentic-dev:review). The skills contain all necessary templates, output formats, and instructions. Keep agent prompts minimal — state the context (milestone, wave, task) and invoke the skill. Do NOT rewrite or paraphrase skill instructions in the agent prompt. Custom prompts cause agents to miss critical templates (task review file format, fix result format, etc.).
Step 1 — Execute Wave
Read the wave's task list from PLAN.md. Check the Depends on: field for each task. Classify tasks as independent (no intra-wave dependencies) or dependent (depends on another task in this wave).
Agent role matching: For each task, read the Role: field from PLAN.md. Map it to the corresponding agent type when launching:
| Plan Role | Agent Type |
|---|
| backend-engineer | backend-engineer |
| frontend-developer | frontend-developer |
| data-engineer | data-engineer |
| devops-engineer | devops-engineer |
| security-engineer | security-engineer |
| qa-engineer | qa-engineer |
| senior-engineer | senior-engineer |
For the single-agent wave case, use the role of the first task (or senior-engineer if the wave mixes roles). For per-task agents, each agent gets its task's specific role.
If the wave has 1 task OR all tasks form a dependency chain: spawn a single agent (set agent type to the task's role):
You are running the execute step of an automated pipeline for milestone M, wave W.
Invoke /agentic-dev:execute wave W.
Follow the skill's instructions completely. Execute all tasks in the wave.
OVERRIDE: Do not stop for human review or approval. Complete all tasks and their task review files.
When done, confirm which tasks were executed and their status.
If the wave has multiple independent tasks: launch one agent per independent task. Send each agent in a separate sequential message with isolation: "worktree" and run_in_background: true. This staggers worktree creation to avoid git config lock contention. All agents run concurrently once launched — only the launch is serialized. Do NOT wait for each agent to fully complete before launching the next — that eliminates parallelism. Tasks that depend on other tasks in the wave must wait for their dependency to complete first.
Per-task agent template (set agent type to the task's role from PLAN.md):
You are running the execute step of an automated pipeline for milestone M, task X.Y.
Invoke /agentic-dev:execute task X.Y.
Follow the skill's instructions completely. The feature branch is feature/mM-waveW-short-description.
OVERRIDE: Do not stop for human review or approval. Complete the task and its task review file.
When done, confirm what was implemented and the task status.
Wait for all task agents to complete. Then merge each task branch into the feature branch (in dependency order if any exist). After merging, delete task branches and worktree tracking branches — the latter are platform artifacts from isolation: "worktree":
git branch -d task/X.Y-short-title # for each merged task
git branch -D worktree-agent-* 2>/dev/null || true
Proceed to Step 2.
Post-check: Verify workflow/plan/reviews/task-X.Y.md exists for each task in the wave. Verify PROGRESS.md shows those tasks at status review. If any task review is missing, STOP and report.
Step 2 — Review Wave
Spawn an independent agent (code-reviewer role) with this prompt:
You are running the review step of an automated pipeline for milestone M, wave W.
Invoke /agentic-dev:review wave W.
Follow the skill's instructions completely. Write the review to workflow/plan/reviews/wave-mM-N.md.
OVERRIDE: Do not stop for human review or approval. Complete the full review.
When done, summarize the issues found (or confirm no issues).
Post-check: Read the wave review file. Verify it exists and has ## Summary. Continue to Step 2b — security review is always performed regardless of whether the code review found issues.
Step 2b — Security Review
Spawn an independent agent (security-reviewer role) with this prompt:
You are running a security review of an automated pipeline for milestone M, wave W.
Invoke /agentic-dev:review security wave W.
Follow the skill's instructions completely. Append your findings to the wave review file at workflow/plan/reviews/wave-mM-N.md.
OVERRIDE: Do not stop for human review or approval. Complete the full security review.
When done, summarize the security findings (or confirm no issues).
Post-check: Read the wave review file. Check for ## Security Review section. Now count BLOCKERs and SUGGESTIONs across BOTH the code review (### Issues Found under ## Task X.Y sections) and security review (### Issues Found under ## Security Review):
- Zero BLOCKERs and SUGGESTIONs total → short-circuit to Step 6 (merge).
- If issues exist → verify
### Fix Plan was also generated by the code review step. If fix plan is missing, continue to Step 3 which will generate one as fallback. If fix plan exists → continue to Step 3 for analysis. The fix plan must cover issues from both reviews.
Step 3 — Fix Plan Analysis
Spawn an independent agent (software-architect role) with this prompt:
You are running the fix-plan-analysis step of an automated pipeline for milestone M, wave W.
Invoke /agentic-dev:review fix-plan-analysis wave W.
Follow the skill's instructions completely. Analyze the fix plan in workflow/plan/reviews/wave-mM-N.md.
OVERRIDE: Do not stop for human review or approval. Complete the analysis.
When done, confirm the analysis was written.
Post-check: Read the wave review file. Verify ### Fix Plan Analysis exists under ## Review Discussion. If missing, STOP and report.
Step 4 — Execute Fixes
Spawn an independent agent with this prompt (the execute skill determines the correct role(s) in fix mode — see execute skill step 3):
You are running the fix execution step of an automated pipeline for milestone M, wave W.
Invoke /agentic-dev:execute fix the issues according to workflow/plan/reviews/wave-mM-N.md.
Follow the skill's instructions completely. Apply all fixes from the fix plan.
OVERRIDE: Do not stop for human review or approval. Apply all fixes and report results.
IMPORTANT: Only apply fixes and report results. Do NOT verify fixes yourself — that is Step 5's job with a separate agent and clean context. The person who writes the fix must not be the one who verifies it.
When done, confirm which fixes were applied.
Post-check: Read the wave review file. Verify ### Fix Results exists under ## Review Discussion. If ### Fix Verification also appears (the fix agent verified its own work), this is a process violation — proceed to Step 5 anyway for independent verification. If ### Fix Results is missing, STOP and report.
Step 5 — Verify Fixes
Spawn an independent agent (qa-engineer role) with this prompt:
You are running the verification step of an automated pipeline for milestone M, wave W.
Invoke /agentic-dev:review verify-fixes wave W.
Follow the skill's instructions completely. Verify all fixes in workflow/plan/reviews/wave-mM-N.md.
OVERRIDE: Do not stop for human review or approval. Complete the full verification.
When done, state clearly how many issues were resolved vs unresolved.
Post-check: Read the wave review file. Find ### Fix Verification. Check each issue's status:
- All marked "Fixed" → proceed to Step 6.
- Any marked "Not Fixed" or "Regression" → STOP and report (see Failure Handling).
Step 6 — Merge and Cleanup
The orchestrator performs this step directly (no subagent needed):
-
Merge fix branches into feature branch (if any exist):
git checkout feature/mM-waveW-*
git merge fix/* # merge each fix branch
git branch -d fix/* # delete fix branches
-
Merge feature branch to main:
git checkout main
git merge feature/mM-waveW-short-description
-
Clean up branches:
git branch -d feature/mM-waveW-short-description
Delete any remaining task branches (task/X.Y-*) and worktree tracking branches (worktree-agent-*) for this wave:
git branch -D worktree-agent-* 2>/dev/null || true
-
Verify: Confirm you are on main and the feature branch no longer exists.
If a merge conflict occurs at any point, STOP and report the conflict. The user must resolve it manually.
Report: "Wave W complete. Feature branch merged to main. All branches cleaned up."
Milestone Mode: Wave Loop
For auto milestone N:
- Parse PLAN.md for all
### Wave W: headings under ## Milestone N.
- For each wave (in order):
a. Check completion: All tasks in the wave are
done in PROGRESS.md AND no feature/mM-waveW-* branch exists. If complete, log "Wave W: already complete" and skip.
b. Run the full wave pipeline (Steps 1-6) using artifact-based resume.
c. If the wave fails (unresolved issues, merge conflict, missing artifact) → STOP the milestone. Report which wave failed and what went wrong.
d. If the wave succeeds → log "Wave W complete" and advance to next wave.
- After all waves complete, run Step 7 — Verify & Retrospective.
Step 7 — Verify & Retrospective (milestone mode only)
Spawn an independent agent (qa-engineer role) with this prompt:
You are running end-to-end verification for milestone M after all waves have been merged.
Invoke /agentic-dev:verify.
Follow the skill's instructions completely. Run the full test suite, verify README.md, and generate the retrospective.
OVERRIDE: Do not stop for human review or approval. Complete the full verification and retrospective.
When done, confirm the retrospective file location and overall pass/fail status.
Post-check: Verify workflow/retro/RETRO-MN.md exists. If the verify step found test failures, report them — the user decides whether to fix before the retrospective is finalized.
Report: "Milestone N complete. All waves merged. Retrospective at workflow/retro/RETRO-MN.md."
Failure Handling
When the pipeline stops, report clearly:
PIPELINE STOPPED — Wave W, Step N (step-name)
Reason: {{what went wrong}}
Artifact: {{which file to check}}
Next action: {{what the user should do to resolve}}
To resume after resolution, run the same command:
/agentic-dev:auto wave W
/agentic-dev:auto milestone N
Specific failure cases:
- Missing artifact after subagent: The subagent completed but didn't produce the expected file/section. Report which artifact is missing.
- Unresolved issues after verify-fixes: List the unresolved issues. The user should fix manually or investigate, then re-run.
- Merge conflict: Report the conflict. The user resolves it, then re-runs.
- Subagent error: Report the error. The user investigates and re-runs.
No automatic retries. The user resolves the issue and re-runs the same command. Artifact-based resume picks up from the right step.
Constraints
- Fully unattended — no human gates between steps within a wave pipeline.
- Clean context per step — each step is an independent agent. No reasoning leaks between steps.
- Orchestrator reads, agents write — the orchestrator checks artifacts between steps but never modifies review files directly.
- NITs do not block — only BLOCKERs and SUGGESTIONs trigger the fix cycle.
- Sequential waves — milestone mode runs waves in order. No cross-wave parallelism.
- Direct merge — uses
git merge (solo project default). For team PR workflows, run wave mode and handle PRs manually.
- Append-only — same constraint as the review skill for
## Review Discussion sections.
