Pipeline

CLoClo Pipeline

Chains SuperPowers skills (brainstorming, writing-plans, subagent-driven- development, verification) with autonomous reviews (Codex + CodeRabbit) and a PR-first auto-merge flow. Invokes underlying skills — does not reimplement them.

When to Use

/pipeline, "new feature", "build X", "implement Y". Multi-step work worth spec → plan → impl → verify gating. Not for one-liners, typos, or reads.

Invocation — no flags, ever

• /pipeline → read checkpoint.json (authoritative); if missing or corrupt, fall back to the artifact scan. If prior state exists, ask A/B/C in the terminal. Else run fresh from Phase 1.
• /pipeline <free text> → interpret text as a directive, skip the dialogue. Examples: passe au plan, le code est ecrit revois, refais tout, pas de codex, ship mode, pas de PR.

French, English, and mixed phrasing all accepted. 80% clear intent → act and log. Ambiguity → one clarifying terminal question. Full directive vocabulary: references/smart-resume.md.

The Pipeline Flow

Full per-phase execution: references/phases.md. GLM runs through the shared glm-companion runtime. It resolves a Z.ai key from ~/.glm.env, then ZAI_API_KEY / GLM_API_KEY, then the project .env. Missing key or failure of both the preferred and fallback models = skipped with a logged warning; Codex alone covers the review.

Step 0: Toggle Check

Before running the flagship, resolve the project root and check for a pause flag:

ROOT="${CLAUDE_PROJECT_DIR:-$(git rev-parse --show-toplevel 2>/dev/null || pwd)}"
[ -f "$ROOT/.cloclo-disabled" ] && echo "PAUSED"

If .cloclo-disabled exists, tell the user CLoClo is paused and ask whether to proceed anyway before starting Phase 1. Otherwise continue normally. (/toggle manages this flag.)

Confidence-First Principle (applies everywhere)

Autonomous does not mean guessing. Every decision — applying a finding, skipping it, interpreting a directive, picking a default — passes a confidence check. Below 95% confidence, ask the user in the terminal (never GitHub) with 2-3 concrete options, marking the recommended one and accepting free-form text.

Canonical decision flow, ask format, and the when / when-not-to-ask lists: references/confidence-first.md.

Auto-Integration — The 3 Gates

Every review phase (2, 4, 6, 6.5, 9) applies findings automatically under the same rule — this is the ONLY integration model; there are no interactive decision points in review phases. A finding is auto-applied only when ALL three gates pass:

1. Concrete patch or revision provided. Bot/reviewer gave a diff, an AI-Agent prompt with file:line+replacement, or for spec/plan a specific section + exact new text. Judgment-only findings ("consider refactoring X") are skipped.
2. Not auth / payments / data migration. Critical domains escalate to the user even with a concrete patch.
3. No conflicting patches across reviewers or findings. Different fixes at the same file:line → skip both, log [CONFLICT], ask.

Canonical statement, consensus and disagreement rules, iteration caps: references/review-chain.md.

Iteration caps: 3 rounds for code phases, 2 for spec/plan. After the cap, exit the loop and escalate any remaining critical findings.

Commit format for auto-applied fixes:

fix(phase-{N}): auto-apply {reviewer} findings ({N} fixes)

- [file:line] — description
- [CONSENSUS file:line] — description (2+ independent reviewers agreed)

Skipped (judgment-only or critical domain):
- [file:line] — reason

PR-First + Auto-Merge (Phase 9)

Phase 9 opens a Pull Request via superpowers:finishing-a-development- branch, waits for installed bots (bot-wait rule: first round until ≥1 bot has posted, max 10 min; re-review iterations max 5 min — see references/review-chain.md), auto-applies their concrete patches under the 3 gates, re-reviews, and auto-merges with --delete-branch when clean. The user stays in the terminal. When a Z.ai API key is available, Phase 9.5 runs a final GLM pass on the post-merge HEAD — preferring glm-5.2 and reporting any explicit glm-4.7 fallback. It is an independent safety net, non-blocking for P1/P2, and auto-escalates P0.

Default bots (no extra config once installed): CodeRabbit GitHub App

• Gemini Code Assist. Opt-in: Codex Cloud, Claude Code Action. Full stack + install URLs: references/bot-stack.md.

Consensus: any 2+ independent reviewers flag the same file:line → [CONSENSUS]. Full rule (severity escalation, when it overrides gate 1, conflict handling): references/review-chain.md.

Disagreement handling and escalation triggers: defined once in references/review-chain.md. Escalation is terminal-only with the standard ask format.

Branch lifecycle: on successful auto-merge, gh pr merge --squash --delete-branch --auto removes the branch locally AND on the remote. On escalation, the branch stays alive so the user can push manual fixes and resume.

Session Setup

1. Take the user message (or ask what to build if empty)
2. Create docs/cloclo-sessions/YYYY-MM-DD-<slug>/
3. Init session.log: [timestamp] CLoClo session started: <slug> [timestamp] Prerequisites: superpowers=OK, codex=OK|MISSING, coderabbit=OK|MISSING
4. Optional pipeline.config.md for project-specific verification commands

Supporting References

Heavy detail, variant tables, and lookups live in references/:

• review-chain.md — SHARED CONTRACT: 3 gates, consensus, disagreement, bot-wait rule, Phase 6.5 enable condition, escalation triggers
• smart-resume.md — checkpoint-first resume, artifact-scan fallback, A/B/C dialogue, directive interpretation table
• bot-stack.md — default + opt-in bot list with install URLs and costs
• phases.md — full per-phase execution (the longest reference, 450+ lines)
• prerequisites.md — SuperPowers/Codex/CodeRabbit/agent-browser auto-install + degraded modes
• model-policy.md — Opus/Sonnet/Haiku mixed policy, per-phase assignments
• retries.md — bounded retries, spike/dev/ship maturity levels
• session-files.md — session dir layout, checkpoint.json, handoff.md formats
• confidence-first.md — the 95% confidence gate, decision flow, ask format

Important Rules

1. Invoke underlying skills; never reimplement them. SuperPowers, Codex, CodeRabbit, and the finishing skill do the real work.
2. Phase 1 (brainstorming) is the only scheduled interactive phase. Every other phase auto-integrates under the 3 gates. BUT Confidence-First applies universally — ask when unsure, don't guess.
3. Questions stay in the terminal, never on GitHub. The user should never have to open the GitHub UI to answer the pipeline.
4. No flags, ever. Natural-language directives replace every flag use case. Escape hatch = edit the session dir before re-running.
5. Each phase outputs to the session dir with numbered filenames for traceability.
6. Checkpoint after every phase; handoff on exit, always. Session resumes via checkpoint.json after any crash.
7. SuperPowers specs/plans live in their own directories (docs/superpowers/specs/, docs/superpowers/plans/). CLoClo copies/symlinks them into the session dir.
8. Wiki ingest (Phase 8) is automatic and silent. Skipped when no wiki exists in the project.
9. Reviews NEVER auto-skip. Codex fails → Claude subagent fallback. CodeRabbit fails → warn and continue (static analysis is a safety net, not a blocker).
10. Branch deleted on auto-merge success (--delete-branch flag). On escalation the branch stays alive so the user can push manual fixes and rerun /pipeline.