kenoxa/do-build

Automated build-review-polish loop composing run-* phase skills: scope → run-implement → run-review ↔ run-implement (fix) → run-polish → finalize.

Entry Gate

Requires do-design recommendation or clear task with approved direction.

Depth

Classify at entry. Controls review fanout. Default standard.

| Level | Heuristic | Review behavior | |-------|-----------|----------------| | focused | Single partition, 1-3 files | Inline review (no subagent batch) | | standard | Default | Full run-review dispatch |

Phases

• Session: reuse input session when available; log at every phase boundary.
• Phase Trace: finalize verifies rows for scope, implement, review, review-gate, polish, finalize.

| Phase | Mechanism | Reference | |-------|-----------|-----------| | Scope | mainthread | — | | Implement | invoke /run-implement | — | | Review | invoke /run-review | — | | Review gate | mainthread | build-review-gate.md | | Polish | invoke /run-polish | — | | Finalize | mainthread | build-finalize.md |

1. Scope

Main thread. Read plan or consult recommendation, classify depth, partition work. Output scope_artifact: target_files, partitions, input_source, input_excerpt, risk_level, blocking_questions (must be empty). Ask user when blocking questions non-empty.

Doc-awareness: when the task changes user-visible behavior, APIs, or configuration, include affected doc files (README, CHANGELOG, docs/) in target_files. Keep doc files in the same partition as owning code. Compose with use-writing skill.

Risk classification (advisory — user may override):

• low — config changes, single-module edits, no trust boundaries
• medium — multi-file changes, dependency updates, API surface changes
• high — auth/payment/migration, trust boundary changes, shared middleware

Risk level feeds run-review's risk scaling (security probe at high, testing-depth at medium).

2. Implement

Invoke /run-implement with scope_artifact. Returns implement_artifact with files_modified.

3. Review

Invoke /run-review scoped to files_modified at scope_artifact.risk_level. Returns findings artifact.

Review gate: Load build-review-gate.md. Read run-review output. Verdict: ITERATE or ACCEPT.

• ITERATE → invoke /run-implement with fix_context (blocking findings) → re-invoke /run-review → re-gate. Cap 5 iterations; freeze on cap.
• ACCEPT → proceed to polish.

Stuckness advisory — after 2 consecutive review-loop iterations where the same error signature repeats (same file + same finding category), surface an advisory to the user. Fires once per build session; informational only — does not auto-dispatch or modify caps.

Suggest agents based on failure pattern:

• Same code approach failing repeatedly → suggest @consultant for approach reset
• Cross-cutting issue (same finding across multiple files) → suggest @envoy for cross-provider second opinion
• Missing external knowledge (API, library, framework) → suggest @navigator for upstream research

4. Polish

Invoke /run-polish scoped to files_modified. Repeat until no E2+ actions remain. Cap 3 iterations.

5. Finalize

Main thread. Load build-finalize.md unconditionally. Sole completion authority.

Anti-Patterns

• Bypassing run-* skills to dispatch agents directly
• Exceeding review cap (5) or polish cap (3) — freeze and surface to user
• Skipping finalize after ACCEPT
• Skipping polish after review passes
• "I'll review after all partitions are done" — bugs compound across partitions; review each
• "This fix is obvious, skip re-review" — obvious fixes break adjacent code; re-review catches regressions
• "Polish is cosmetic, ship now" — polish catches complexity and convention drift missed by review

Completion

• Phase Trace rows exist for scope, implement, review, review-gate, polish, finalize [E3]
• All blocking review findings resolved or user-deferred [E2]
• Token counts verified for any modified skill files [E3]