/cauto — Semi-Auto Implementation Pipeline
Orchestrate the full implementation pipeline after spec review. Runs /ctdd, /simplify, /cverify, /cupdate-arch, /cdocs, then creates a PR — with flexible phase resume and autonomous decision-making.
When to Use
- After
/creview(or/creview-spec) approves a spec — this runs the entire implementation pipeline - When you want hands-off execution from TDD through PR creation
- When resuming a pipeline that was interrupted (escalation, context overflow, hard stop)
- Not for: features that don’t have an approved spec yet (run
/cspecand/creviewfirst)
How It Fits in the Workflow
/cauto is the pipeline orchestrator. Instead of manually running /ctdd → /cverify → /cdocs → PR, /cauto runs them all in sequence, handling failures, retries, and escalation. It respects the same workflow state machine as manual invocation — skills still run in separate agents with context: fork. The orchestrator invokes skills; skills do not auto-continue on their own.
What It Does
- Phase gate — reads the current workflow phase and computes remaining pipeline steps. Accepts any active phase from
reviewthroughdocumented. Rejectsspecandmodel(spec must be reviewed first). - Install freshness check — detects stale hooks before any skill runs. Warns if source files changed since last setup.
- Invoke /ctdd — spawns a sub-agent for the full RED → GREEN → QA → mini-audit cycle. If TDD triggers a spec-update, escalates to the human.
- Commit before /simplify — creates a clean revert point (“TDD complete”).
- Invoke /simplify — runs Claude Code’s built-in simplification. Post-simplify validation reverts if tests fail or
.correctless/paths were modified. - Invoke /cverify — verifies implementation matches spec.
- Invoke /cupdate-arch (high+ intensity only) — updates architecture documentation.
- Invoke /cdocs — updates project documentation.
- Consolidation — scoped commit and push. Only stages known pipeline output paths. Belt-and-suspenders guard prevents committing artifacts.
- Create PR — via
gh pr create(default), skip, or custom command. PR body includes summary, test plan, QA findings, and verification status.
Escalation
When a failure persists after N attempts (3 at standard, 2 at high/critical), the pipeline stops and writes a structured escalation file. Re-running /cauto after the human resolves the issue resumes from where it left off.
Phase 2: Tiered Decision Architecture
Decisions are routed through a tiered hierarchy: Policy (deterministic) → Worker (self-resolution) → Decision Agent (ephemeral) → Supervisor (lightweight) → Hard Stop (human). Token and time budgets are enforced with warn at 75% and hard stop at 100%.
Phase 3: Spec-to-PR Orchestration
When no workflow exists, /cauto can start from scratch — running /cspec interactively or accepting a provided spec file. Autonomous review with supervisor triage incorporates findings automatically. A mandatory spec approval gate always requires human confirmation before implementation begins.
Example
Agent: Reading workflow state... phase is "review". Full pipeline ahead.
Checking install freshness... all hooks current.
Reading preferences... pr_creation: gh, defaults for all others.
Starting /ctdd...
[RED → test audit → GREEN → /simplify → QA → mini-audit]
/ctdd complete. Committing TDD changes.
Running /simplify... accepted (tests pass, no protected paths modified).
Starting /cverify...
Verification complete — 5/5 rules covered, 0 smells.
Skipping /cupdate-arch (standard intensity).
Starting /cdocs...
Documentation updated.
Consolidation: staging pipeline outputs...
Pushed to origin/feature/my-feature.
Creating PR via gh...
PR #42 created: https://github.com/user/repo/pull/42
Pipeline Summary:
┌──────────────┬──────────┬────────┬────────┐
│ Step │ Duration │ Tokens │ Result │
├──────────────┼──────────┼────────┼────────┤
│ ctdd │ 8m 12s │ 45,200 │ pass │
│ simplify │ 0m 34s │ — │ pass │
│ cverify │ 1m 05s │ 8,300 │ pass │
│ cupdate-arch │ — │ — │ skip │
│ cdocs │ 2m 18s │ 12,100 │ pass │
└──────────────┴──────────┴────────┴────────┘
What It Reads / Writes
| Reads | Writes |
|---|---|
Workflow state (.correctless/artifacts/workflow-state-{slug}.json) | All artifacts from invoked skills |
.correctless/config/workflow-config.json | .correctless/artifacts/audit-trail-{slug}.jsonl |
.correctless/preferences.md | .correctless/artifacts/escalation-{slug}.md (on failure) |
.correctless/config/auto-policy.json (Phase 2) | .correctless/meta/overrides/{slug}-{date}.json |
| Approved spec | PR on GitHub |
Intensity Levels
| Aspect | Standard | High | Critical |
|---|---|---|---|
| Pipeline | ctdd → simplify → cverify → cdocs → PR | + cupdate-arch | + cupdate-arch |
| Failure threshold | 3 attempts | 2 attempts | 2 attempts |
| Supervisor activations | Same | Same | Same |
Common Issues
- “Run /creview first”: The spec must be reviewed before
/cautocan run. Run/creviewor/creview-specfirst. - Escalation during /ctdd: Usually a spec issue — the spec contradicts implementation reality. Fix the spec and re-run.
- Push failure: Auth or rejected push. The local commit is preserved. Fix the issue and re-run — the phase gate routes to PR-only.
- “gh CLI not installed”: Set
pr_creation: skipin.correctless/preferences.mdor installgh. - Context overflow: Each skill runs in a fresh context via
context: fork. If the orchestrator itself overflows, re-run — the checkpoint system saves progress.