| name | deep-auto |
| description | Fully autonomous pipeline orchestrator. Chains deep-plan (with auto-resolved Q&A), deep-execute, deep-review, and optionally deep-complete into a single unattended run. All critic questions are resolved using codebase evidence and codified best practices. Resumable if interrupted. Use when the user says "deep auto", "/deep-auto", or wants to run the full pipeline autonomously.
|
| imports | ["compaction-resilience"] |
| invariants | ["Preflight infrastructure check is MANDATORY — fix or abort, never skip.","Auto-resolved Q&A must reflect genuine trade-off analysis, not path of least resistance.","deep-review MUST be invoked — never skip review unless --skip-review was explicitly passed.","0 test failures for final PASS — inherits deep-review's no-exceptions rule."] |
Deep Auto Skill
Fully autonomous pipeline that chains plan, execute, review, and optionally complete
without human intervention. Auto-resolves all critic Q&A using codebase evidence and
a codified decision framework.
Behavioral directive: Load and follow the intellectual-honesty skill. Even in autonomous mode, auto-resolved decisions must reflect genuine trade-off analysis, not the path of least resistance.
Invocation
/deep-auto "feature description" # Full pipeline: plan + execute + review (3 rounds)
/deep-auto --resume {slug} # Resume interrupted pipeline
/deep-auto --rounds 5 "feature description" # Custom review rounds
/deep-auto --complete "feature description" # Include deep-complete phase
/deep-auto --complete --auto-commit "feature description" # Complete + auto-commit
/deep-auto --skip-review "feature description" # Plan + execute only (no review)
/deep-auto --mode bug "bug description" # Bug-fix pipeline: validate + red + fix + verify + regression
References
This skill has reference files in references/:
references/auto-resolve-framework.md — Decision framework for auto-resolving critic Q&A
references/infrastructure-preflight.md — Infrastructure checks for specific stacks
Step 0: Preflight — Infrastructure Health Check (Auto-Fix)
This step is MANDATORY and runs before ANYTHING else. The pipeline MUST fix infrastructure
issues automatically. No user input. No "please start X manually." Fix it or die trying.
Check Process
Run checks in order. Each check follows: detect → auto-fix → verify → abort only if unfixable.
- Read
references/infrastructure-preflight.md for stack-specific checks (database, build tools, dev server, etc.)
- Run each check following the detect → auto-fix → verify pattern
- Abort only if auto-fix fails after reasonable retries
Preflight Summary
Print on success (showing what was auto-fixed):
[deep-auto] Preflight checks PASSED
[deep-auto] {service}: UP {+ " [auto-started]" if it was started}
[deep-auto] {service}: CONNECTED
[deep-auto] {service}: AVAILABLE
Step 0b: Initialize
-
Parse user input:
- Extract feature description (everything not a flag)
- Parse flags:
--resume, --rounds N (default 3), --complete, --auto-commit, --skip-review, --mode bug
-
Generate slug from feature description (kebab-case, e.g., add-tooltip-to-stat-cards)
-
Check for docs/plans/{slug}/meta.md:
- Exists → Resume mode. Read meta, find first incomplete phase using the System column.
- Doesn't exist → Fresh start.
-
Save verbatim prompt (immutable) — Write the user's exact input (unprocessed, unedited) to
docs/plans/{slug}/prompt.md (or docs/bugs/{slug}/prompt.md for bug mode).
This preserves the original request for pipeline evaluation.
If the file already exists, NEVER overwrite it — the first writer wins. Since deep-auto
writes this before invoking deep-plan/deep-bug, downstream skills will see it exists and skip.
---
timestamp: { ISO timestamp }
source: deep-auto
flags: { parsed flags, e.g., --rounds 5 --complete }
---
{exact user prompt, verbatim — no summarization, no slug extraction, no cleanup}
-
Read references/auto-resolve-framework.md for Q&A auto-resolution context.
-
Write invariants (compaction resilience): Write this skill's invariants: from
frontmatter to docs/plans/{slug}/invariants.md. See compaction-resilience skill.
-
If --mode bug:
- Instead of
docs/plans/{slug}/, use docs/bugs/{slug}/
- Check for
docs/bugs/{slug}/meta.md for resume state
- The bug pipeline uses
/deep-bug instead of /deep-plan + /deep-execute
Print:
[deep-auto] Starting autonomous pipeline for: "{feature description}"
[deep-auto] Slug: {slug}
[deep-auto] Mode: {fresh | resume from Phase N}
[deep-auto] Review rounds: {N}
[deep-auto] Completion: {yes | no}
Bug Mode Branch
If --mode bug:
- Invoke
/deep-bug "{description}" (this handles all phases internally: reproduction, intent validation, clarification, red, fix, verification, regression review, and completion)
/deep-bug internally invokes /deep-review --rounds 1 --slug {slug} --output-dir docs/bugs/{slug}/review
/deep-bug internally invokes /deep-complete --slug {slug}
- After
/deep-bug completes all phases, the pipeline is done — skip Steps 1-4 entirely
Skip Steps 1-4 entirely -- /deep-bug handles its own planning, execution, review, and completion.
Step 1: Planning Phase
Skip if: meta.md shows all plan-system phases (Research through Plan Verification) completed.
Invoke the deep-plan skill with auto-resolve and research:
/deep-plan --auto-resolve --research "{feature description}"
Or if resuming:
/deep-plan --resume {slug}
IMPORTANT: The --auto-resolve flag tells deep-plan to auto-resolve the Q&A phase without
presenting questions to the user. It uses the auto-resolve decision framework to make decisions
based on codebase evidence and codified best practices.
The --research flag ensures autonomous runs include deep internet research (market analysis,
competitor landscape, UX theory, user pain points, and industry trends) alongside codebase
research. Vision alignment within deep-research auto-defers non-aligned items to
docs/backlog/out-of-vision-scope.md — no prompts in auto mode.
Gate Check
After deep-plan completes, verify:
docs/plans/{slug}/final-plan.md exists and is non-empty
docs/plans/{slug}/task-graph.md exists and is non-empty (Plan Verification output)
meta.md shows Plan Verification as completed
On failure:
- Try
--resume {slug} once to recover from where it left off
- If still failing, stop and report (see Error Handling)
Print:
[deep-auto] Planning COMPLETE
[deep-auto] Final plan: docs/plans/{slug}/final-plan.md
[deep-auto] Task graph: docs/plans/{slug}/task-graph.md
[deep-auto] Auto-resolved questions: {count} ({high-confidence}/{medium}/{low})
Step 2: Execution Phase
Skip if: meta.md shows Implementation as completed.
Invoke the deep-execute skill:
/deep-execute {slug}
Or if resuming:
/deep-execute --resume {slug}
Gate Check
After deep-execute completes, verify:
- Run lint — must pass with 0 errors
- Run build — must pass with 0 errors
meta.md shows Implementation as completed
On failure:
- If lint fails: run lint fix, then re-check
- If build fails: read the error output, attempt to fix errors
- Try
--resume {slug} once after fixing
- If still failing, stop and report
Print:
[deep-auto] Execution COMPLETE
[deep-auto] Tasks completed: {done}/{total}
[deep-auto] Lint: PASS | Build: PASS
Step 3: Review Phase (MANDATORY)
Skip if: --skip-review flag is set, or meta.md shows all review phases completed.
CRITICAL: This step MUST invoke /deep-review --rounds {N} --slug {slug}.
Do NOT substitute manual test runs or filtered tests.
The review phase runs the FULL test suite and spawns blind reviewers.
What this is NOT:
- Running filtered test subsets is NOT a review
- Manually fixing test selectors is NOT a review round
- Running only tagged tests is NOT sufficient
What this IS:
/deep-review --rounds {N} --slug {slug} spawns blind reviewers per round
- Each round runs the full test suite
- Each round can find different issues (round 1: most, round 2: regressions, round 3: polish)
Invoke the deep-review skill:
/deep-review --rounds {N} --slug {slug}
Where N is the --rounds value (default 3).
Gate Check (STRICT)
After deep-review completes:
- File MUST exist: Check the test results file:
- Feature mode:
docs/plans/{slug}/review/test-results.md
- Bug mode:
docs/bugs/{slug}/review/test-results.md
- If it does NOT exist → the review phase did NOT run the full test suite. HARD FAILURE. Re-run.
- Check for coverage gaps: If
docs/plans/{slug}/execution/coverage-gaps.md exists:
- The completeness reviewer found MISSING flows too large for the fixer
- Invoke
/deep-execute --resume {slug} to implement the coverage gaps
- After execution completes, reset all review phases to pending in
meta.md
- Re-invoke
/deep-review --rounds {N} --slug {slug} from Round 1
- This loop (review → execute → review) can happen at most once
- Read the file and check:
Meets Standards: line — must be YES
Failed: 0
- If standards NOT met:
- Do NOT proceed to completion phase
- Re-run
/deep-review --rounds 2 --slug {slug} for additional fix cycles
- If STILL not met after recovery round, stop and report as unrecoverable failure
- Only proceed to Step 4 when the file exists AND
Meets Standards: YES
Print:
[deep-auto] Review COMPLETE
[deep-auto] Rounds: {N} | Findings fixed: {count} | Out-of-scope: {count}
[deep-auto] Test results: docs/plans/{slug}/review/test-results.md — Meets Standards: YES
[deep-auto] Verdict: PASS
Step 4: Completion Phase (conditional)
Skip if: --complete flag was NOT passed, or meta.md shows Completion as completed.
Invoke the deep-complete skill:
/deep-complete {slug}
Or with auto-commit:
/deep-complete --auto-commit {slug}
Print:
[deep-auto] Completion phase DONE
[deep-auto] Commits: {proposed | created}
[deep-auto] Documentation: {count} files created/updated
Final Summary
After all phases complete, print:
[deep-auto] Pipeline COMPLETE for "{feature description}" ({slug})
Phases:
Planning: COMPLETE ({auto-resolved-count} questions auto-resolved)
Execution: COMPLETE ({task-count} tasks implemented)
Review: COMPLETE ({round-count} rounds, {fix-count} fixes)
{Completion: COMPLETE (if --complete)}
Artifacts:
- Final plan: docs/plans/{slug}/final-plan.md
- Flow spec: docs/plans/{slug}/user-flow-spec.md
- Task log: docs/plans/{slug}/execution/task-log.md
- Review: docs/review/summary.md
{- Commits: docs/plans/{slug}/complete/commit-plan.md (if --complete)}
Low-confidence auto-resolved questions: {count}
{list any low-confidence decisions for optional human review}
Next steps:
- Review auto-resolved decisions in docs/plans/{slug}/questions.md
- {If no --complete: Run `/deep-complete {slug}` to finalize commits and docs}
- Run `/deep-improve` periodically for pipeline self-improvement
Error Handling
Recovery Strategy Per Phase
| Phase | Failure | Recovery | Max Retries |
|---|
| Plan | final-plan.md missing | /deep-plan --resume {slug} | 1 |
| Execute | lint/build fails | lint fix, fix errors, --resume | 1 |
| Review | verdict NEEDS FIXES or test-results.md missing/failing | /deep-review --rounds 2 --slug {slug} (additional) | 1 |
| Complete | agent output missing | /deep-complete --resume {slug} | 1 |
Unrecoverable Failure Protocol
If a phase fails after its recovery attempt:
-
Update meta.md with current state (mark failed phase as failed)
-
Write failure details to docs/plans/{slug}/auto-failure-log.md:
# Auto Pipeline Failure
- **Slug:** {slug}
- **Failed Phase:** {phase name}
- **Timestamp:** {ISO}
- **Error:** {error output}
- **Recovery Attempted:** {what was tried}
## Resume Command
/deep-auto --resume {slug}
-
Print a clear error message:
[deep-auto] FAILED at {phase name}
[deep-auto] Error: {brief description}
[deep-auto] Failure log: docs/plans/{slug}/auto-failure-log.md
[deep-auto] Resume with: /deep-auto --resume {slug}
Resume Logic
On /deep-auto --resume {slug}:
- Read
docs/plans/{slug}/meta.md
- Walk the phase table top to bottom:
- Find the first phase with status !=
completed
- Use the System column to determine which skill to invoke
- Phase-to-skill mapping:
| System | Phases | Skill |
|---|
| plan | Research through Plan Verification | /deep-plan --auto-resolve --resume {slug} |
| execute | Task Breakdown + Implementation | /deep-execute --resume {slug} |
| review | Review rounds + Full E2E Suite | /deep-review --rounds {N} --slug {slug} |
| complete | Completion | /deep-complete --resume {slug} |
| bug | All bug phases | /deep-bug --resume {slug} |
- Jump to that step and continue the pipeline from there
- For bug mode resumes:
- Check
docs/bugs/{slug}/meta.md (not docs/plans/{slug}/meta.md)
- If the System column is
bug, review, or complete, invoke /deep-bug --resume {slug}
- After
/deep-bug completes, the pipeline is done
If auto-failure-log.md exists, read it first to understand what previously failed.
Important Notes
- No human intervention required — All decisions are made autonomously using the auto-resolve framework
- Fully resumable — Can be interrupted and resumed at any point via
--resume
- Idempotent — Running twice with the same slug is safe (completed phases are skipped)
- Low-confidence decisions are flagged — User can review
questions.md after completion to audit auto-resolved decisions
- Existing skills are unchanged — Only deep-plan gets the
--auto-resolve flag addition