Deep Build

Require a Plan and independent PASS gates. Use config.example.yaml and external-dependencies.json.

For external guidance, read references/external-skills.md. Record mode, evidence, failures, and fallback in the Plan; never waive gates.

Boundaries

Implementation authorizes scoped local edits and a new Plan. Do not overwrite Plans, commit, push, open a PR/MR, deploy, update issues, or mutate services. Preserve unrelated changes. Show and confirm destructive migrations, rewrites, dependency installs, or permission changes first.

Workflow

1. Recover and map. Reuse requirements/IDs. Inspect instructions, callers, persistence, authorization, failures, tests, and worktree. Record assumptions/decisions.
2. Write the Plan. Create <plan_directory>/<slug>.md with the ledger, evidence, scope/non-goals, slices, files/changes, dependencies, checks, risks/rollback, destructive actions, and review criteria. Do not implement first.
3. Plan Review and user-approval gate. Read references/review-criteria.md for independence and model selection. Self-review cannot satisfy either gate. Immediately report each result: Plan path, reviewer context, verdict, blocking findings, revisions, risks, and next decision. On FAIL, stop at pending user plan decision; revise only after user direction, then re-review. On PASS, stop at pending user plan approval. Implement only after recorded explicit user approval.
4. Implement slices. Follow the approved Plan. Establish observable signals, make small coherent changes, run checks, inspect diffs, and update the ledger. A new interface, migration, dependency, permission change, or out-of-scope caller returns to the Plan gate.
5. Code Review gate. Obtain independent review of the actual diff, surrounding code, and tests using references/review-criteria.md. Record criteria, reviewer/model selection, findings, and PASS/FAIL in the Plan; report the result to the user. The implementer fixes Critical/Required findings, then requests independent re-review. Without PASS, do not hand off as complete.
6. Hand off. After both gates pass, fill assets/build-handoff-template.md with evidence; request user acceptance. Git delivery is separate.

Plan record format

# Build Plan: <change>
## Requirement ledger
| ID | Requirement | Acceptance evidence | Status |
## Codebase evidence and scope
## Slices and verification
| Slice | Requirement IDs | Files/change | Checks | Risk/rollback |
## Predicted review criteria
## Plan Review — Round N
Reviewer/context: <agent | human | external>; Model/effort: <known value | unknown | N/A>
Selection rationale and reviewed version: <evidence>; Verdict: PASS / FAIL
Findings and revisions: <evidence>
## User Plan Decision — Round N
Plan path: <path>; Decision: APPROVED / CHANGES_REQUESTED
Decision evidence: <the user's explicit response>
## Code Review — Round N
Reviewer/context, model/effort, selection rationale, reviewed version: <evidence>
Actual-diff criteria: <list>; Verdict: PASS / FAIL
Critical/Required findings, resolution, and checks: <evidence>

Failure behavior

Without an independent reviewer, report pending external review; never bypass its gate. Preserve tool-failure evidence; report unverified work and recovery. The worktree overrides stale state. Return failed acceptance to its ledger item.