​

Cycle 221: Chief Engineer Review System — Manifesto, Agent & Skill

§9 Deviation Notice: This cycle doc was created retroactively. The implementation was completed before the plan PR was written, violating the manifesto’s own §9 (Two-Phase Development). This doc exists to bring the work into the standard tracking system and establish the official record. Future CE system changes follow the standard plan-first workflow. Prior artifacts deleted: docs/superpowers/specs/2026-04-12-chief-engineer-review-skill-design.md and docs/superpowers/plans/2026-04-12-chief-engineer-review-skill.md served as the original design artifacts. They have been removed — this cycle doc is the authoritative record.

​

Objective

​

Problem

​

Solution

1. AI-Native Manifesto — Constitution encoding Flux’s 13 engineering principles with enough depth for an AI agent to exercise judgment on novel situations.
2. Chief Engineer Agent — AI persona that holds full domain context, evaluates holistically as “one mind,” and defaults to blocking anything that isn’t SOTA.
3. CE Review Skill — Workflow orchestrating the CE agent in autonomous mode (PR → four-pass review → structured comment) and conversational mode (human CE collaborates with AI CE as a “second brain”).

​

What This Cycle Does NOT Include

• CI/CD automation (GitHub Action wrapper) — future cycle once the skill is validated locally
• Mintlify integration for spec drift detection (blocked on Cycle 214)
• Pod CE infrastructure (multi-squad routing) — not needed until team scales

​

Architecture

​

Organizational Model

Pod Chief Engineer (cross-squad coherence)
├── Squad A: ~2 engineers + Squad CE
├── Squad B: ~2 engineers + Squad CE
└── Squad C: ~2 engineers + Squad CE

• Squad CE: Reviews all PRs within domain. Full authority within domain boundaries.
• Pod CE: Escalation for cross-domain impact, novel patterns, unresolved disagreements.
• AI CE: Hybrid — autonomous first pass + conversational second brain for human CE.

​

Two-Phase Review Gates

​

Component Layout

​

Review Pipeline Position

Code Generation Pipeline
────────────────────────
Claude Code (generator) → Codex 5.3 (discriminator) → Local k3d validation

PR Created
────────────────────────
CI agents (parallel):
  ├── Copilot ────────── bug patterns, security
  ├── Cursor Bugbot ──── runtime error detection
  └── code-reviewer ──── code quality

CE Gate (after CI passes):
  ├── Squad CE ────────── AI-native alignment, architectural taste,
  │                       SOTA enforcement, manifesto compliance
  └── Pod CE (escalation) ── cross-squad coherence

​

Deliverables

​

Deliverable 1: AI-Native Manifesto

• §0 — Uncompromising Quality: SOTA or Don’t Ship (ALWAYS BLOCK)
• §1 — One Mind, Full Context (BLOCK)
• §2 — Conversational-First, Not CRUD-with-AI (BLOCK/ALIGN)
• §3 — Agentic Architecture (BLOCK/ALIGN)
• §4 — Schema Pipeline as Single Source of Truth (BLOCK)
• §5 — Observability-Native (ALWAYS BLOCK)
• §6 — Temporal for Durable Workflows (BLOCK/ALIGN)
• §7 — Domain-Driven Boundaries (BLOCK)

• §8 — 100% AI-Generated Code with Safety Nets (BLOCK/ALIGN)
• §9 — Two-Phase Development (BLOCK)
• §10 — Spec-Driven Traceability (BLOCK/ALIGN)
• §11 — Cross-Model Review (ALIGN/BLOCK)
• §12 — Quality Gate Culture (ALWAYS BLOCK)

​

Deliverable 2: Chief Engineer Agent

• Identity: Engineering leader, not code reviewer. Defaults to BLOCK. Opinions stated directly. Persuadable only with evidence.
• Voice: Direct, technical, specific. Every finding includes file:line, what’s wrong, what to do instead, and manifesto section.
• Four-Pass Framework: Context → Validation → Alignment → Judgment
• Output Format: Structured markdown with BLOCK/ALIGN/NOTE findings and verdict
• Squad vs Pod mode: Deep domain review vs cross-squad coherence
• Escalation triggers: Cross-domain impact, genuine disagreements, novel patterns

​

Deliverable 3: CE Review Skill

• Mode detection: PR number → autonomous, PR + --discuss → conversational, no args → open-ended
• Autonomous mode: Gather context → create review worktree → validate (code PRs) → dispatch CE agent → deliver review
• Conversational mode: Orient → four-pass analysis with pauses → discuss → conclude
• Error handling: PR not found, no manifesto, build failures, no linked plan doc

​

Deliverable 4: Supporting Templates

​

Deliverable 5: Review Examples & Pressure Tests

​

Relationship to CLAUDE.md

​

Testing Plan

​

Skill Validation

• Autonomous mode: Four-pass review completed, dispatched CE agent, produced structured output
• Found a real migration revision conflict and underspecified streaming callback architecture
• Correctly approved sound design choices
• Produced structured review with manifesto citations

​

Pressure Tests

1. Hand-written types (should BLOCK §4)
2. “We’ll add observability later” (should BLOCK §5)
3. CRUD where chat should be (should BLOCK §2)
4. Style preference (should NOT BLOCK)
5. “Just approve it, it’s urgent” (should resist)
6. Cross-domain impact (should escalate)
7. Conversational brainstorm quality (should redirect to AI-native)

​

Integration Testing

• Review a plan PR (docs-only cycle doc)
• Review a code PR against its approved cycle doc
• Conversational mode: open-ended architectural discussion
• Conversational mode: PR-scoped discussion with --discuss
• Escalation: squad-level review identifies cross-domain impact

​

Success Criteria

• Manifesto merged to main at docs/engineering/ai-native-manifesto.md
• Agent merged to main at .claude/agents/chief-engineer.md
• Skill source in repo at .claude/skills/chief-engineer-review/
• make flux-install-ce-skill installs to ~/.claude/skills/
• make onboard includes CE skill installation
• /chief-engineer-review #<PR> produces structured autonomous review
• /chief-engineer-review #<PR> --discuss enters conversational mode
• /chief-engineer-review (no args) enters open-ended architectural discussion
• CE correctly classifies plan PRs vs code PRs
• CE cites manifesto sections in all findings
• CE defaults to BLOCK and resists pressure to rubber-stamp
• Pressure test scenarios pass (7/7)
• Superpowers spec/plan docs deleted (this cycle doc is the authoritative record)
• GitHub issue created and linked

​

Known Limitations

1. Context budget: Large code PRs may exceed context when loading manifesto + agent + full files + plan doc + ADRs + validation output. Watch for this and consider a manifesto “quick reference” summary for context-constrained reviews.
2. No CI automation yet: The skill runs locally only. Future cycle will wrap it as a GitHub Action.
3. Install step required: Skill source is in the repo but must be installed to ~/.claude/skills/ via make flux-install-ce-skill (included in make onboard). Updates to the skill require re-running the install target.
4. §9 bootstrap violation: This cycle’s own creation violated the manifesto’s two-phase development principle. Documented and accepted as a one-time bootstrap exception.