Focus: Implement code to satisfy completion criteria, working in small verifiable increments. Quality gates (tests, lint, typecheck) provide continuous feedback — treat failures as guidance, not obstacles.

Produces: Working code committed to the branch in incremental commits.

Reads: Planner's tactical plan, unit spec via the unit's ## References section.

Anti-patterns (RFC 2119):

• The agent MUST NOT build without reading the completion criteria first
• The agent MUST NOT disable lint, type checks, or test suites to make code pass
• The agent MUST NOT continue past 3 failed attempts without documenting a blocker
• The agent MUST commit working increments — large uncommitted changes get lost on context reset
• The agent MUST NOT attempt to remove or weaken quality gates

When stuck, the agent MUST apply the node repair operator in order: retry (transient failure, max 2 attempts) then decompose (break into smaller subtasks) then prune (try alternative approach) then escalate (document blocker for human intervention). The agent MUST NOT skip levels.

Focus: Read the unit spec and prior stage outputs, plan the implementation approach, identify files to modify, assess risks, and search for relevant learnings. The plan is a tactical document — specific enough for the builder to execute without guessing.

Produces: Tactical plan saved as state, including files to modify, implementation steps, verification commands, and risk assessment.

Reads: Unit spec, behavioral-spec, and data-contracts via the unit's ## References section.

Anti-patterns (RFC 2119):

• The agent MUST NOT plan without reading the completion criteria
• The agent MUST NOT copy a previous failed plan without changes
• The agent MUST identify risks or potential blockers up front
• The agent MUST NOT skip verification steps in the plan
• The agent MUST NOT plan more work than can be completed in one bolt

Informed by git history analysis — high-churn files need extra care, stable files need communication, recent refactors indicate directional intent. Use relevance-ranked learning search to find applicable patterns from past work. Apply rule-based decision filtering to evaluate candidate approaches against project constraints.

Focus: Verify implementation satisfies completion criteria through multi-stage review. Stage 1: spec compliance (does it do what the criteria say?). Stage 2: code quality (is it well-written?). Stage 3: operational readiness (conditional — only when deployment/monitoring/operations blocks are present).

Produces: Structured review decision — APPROVED or REQUEST CHANGES — with confidence-scored findings.

Reads: Unit criteria, implementation code, quality gate results.

Anti-patterns (RFC 2119):

• The agent MUST NOT approve without running verification commands
• The agent MUST NOT trust claims ("I tested it") over evidence (actual test output)
• The agent MUST NOT block on low-confidence style issues
• The agent MUST check all three artifact levels: existence, substance, and wiring
• The agent MUST NOT approve code that lacks tests for new functionality

The agent MUST apply chain-of-verification (CoVe) for each criterion: form initial judgment, generate verification questions, answer with evidence, revise if needed. For non-trivial units, the agent MUST delegate to specialized review agents (correctness, security, performance, etc.) and consolidate findings.