/deep-review - Merge Gate Review

Merge-gate review. The bug hunt (correctness, regressions, design, performance) is delegated to the built-in code-review skill, which adversarially verifies its findings. This skill layers what code-review doesn't cover — does the change deliver what was scoped, is it complete across all affected locations, is it adequately validated — and renders a merge verdict.

Usage

/deep-review [target] [level] [--fix]
/deep-review branch --scope .work/dev/5096/
/deep-review branch max --fix --scope .work/dev/5096/

• --scope <path>: When provided, read goals.md from the given directory to understand the original intent, success criteria, and user experience requirements. Use this to evaluate whether the implementation actually delivers what was scoped — not just whether the code is correct in isolation.
• [level] (low | medium | high | xhigh | max): forwarded to the built-in code-review pass as its effort level. Default: high. If ultra is passed, do not launch it (user-triggered and billed) — tell the user to run /code-review ultra themselves, then re-run /deep-review, which will reuse those findings.
• --fix: forwarded to code-review — it applies its confirmed findings to the working tree after its review. Gate findings (goals/completeness/validation) are reported, never auto-applied.
• No argument: review staged changes (git diff --cached)
• all: all uncommitted changes
• branch: changes vs base branch (git diff main...HEAD)
• pr: current PR changes (gh pr diff)
• commit:SHA: specific commit

Positioning

Step 1 — Delegate the bug hunt

Invoke the built-in code-review skill (Skill tool, skill: "code-review") on the target diff, forwarding the user's arguments: the effort level (default high) and --fix if provided.

• If the user already ran /code-review (or ultra) on this change set in this session, do NOT re-run it — use those findings.
• NEVER launch ultra yourself — it is user-triggered and billed. For large or high-risk change sets, suggest the user run /code-review ultra instead of the inline pass.
• commit:SHA targets don't map to code-review's diff model: review that commit's diff yourself with the same lenses (correctness, regressions, design, performance) and say you did. --fix doesn't apply to committed diffs — say so if it was passed.
• With --fix, code-review mutates the working tree — run Steps 2-4 on the POST-fix state, and list the applied fixes in the report.

Carry the confirmed findings into this review: a blocking code-review finding blocks the merge verdict here (unless --fix already resolved it — then note it as fixed).

Step 2 — Goals alignment (when --scope given)

For each success criterion and UX requirement in goals.md, mark it delivered / partial / missing with evidence (file:line or traced behavior). A missing or partial success criterion is a finding — severity per user impact. Without --scope, state the inferred intent of the change in one sentence and evaluate against that.

Step 3 — Completeness

Trace every modified/added/deleted symbol to ALL its consumers — read the modified files in full, not just the diff hunks:

• Import statements, call sites, subclass overrides, interface implementations
• Both environments: desktop/node AND browser/web paths per AGENTS.md
• Per-operation git providers: packages/git/src/providers/, packages/git-cli/src/providers/, host integrations (src/plus/integrations/host/providers/)
• Protocol/type changes: webview IPC pairs, serialization, telemetry, logging
• Decorator-wrapped methods: check AGENTS.md § Decorator System — @gate(), @memoize(), @sequentialize() alter runtime behavior significantly
• Adjacent features not accidentally broken

If a change touches shared abstractions, audit all consumers. If it alters async flows, explicitly inspect cancellation, deduplication, sequencing, and stale-state behavior. When you can't fully trace a path, document it in "Open questions" — don't silently skip it.

Step 4 — Validation

• Assess whether build, typecheck, lint, and relevant tests adequately validate the change. Missing or weak validation is a finding, not a neutral detail. If tests are missing, name the specific cases that should exist.
• Run pnpm run build to confirm the change compiles, type-checks, and lints cleanly; report related failures as findings.

Output Format

Findings (ordered by severity)

Then: what is wrong, why it matters, the best fix (not just a workaround).

Required Sections

1. Findings — gate findings plus code-review's confirmed findings, rolled up and ordered by severity
2. What is good — well-implemented aspects worth calling out
3. Open questions — things that need validation or that you couldn't confirm
4. Verdict — one of:
   • Safe to merge
   • Safe with follow-ups (list them)
   • Should be reworked before merge (explain why)

If no issues found, say so explicitly, then list residual risks and validation gaps.