---
name: cross-model-review
description: Reviews implemented code, commits, and diffs using external frontier models (Claude CLI, Codex CLI, Antigravity CLI, Kiro CLI) or local Ollama models with evidence-grounded triage.
---

# Cross-Model Review

## Overview

Cross-Model Review conducts Shift-Right independent code reviews on implemented software commits and diffs using external frontier models (Claude CLI, Codex CLI, Antigravity CLI, Kiro CLI) or local models (Ollama).

Operating on the principle of **Shift-Right Verification**, it catches code quality issues, missing test coverage, subtle concurrency regressions, architectural drift, and spec deviations before code is merged to main.

**Announce at start:** "I'm using the cross-model-review skill to conduct an independent code review."

## Two Modes of Operation

### 1. In-Lifecycle Mode (Stage 3 Review Choice after Gate 4)

Invoked after all tasks of an epic are implemented and Gate 4 (machine acceptance) returns PASS.

- **Trigger**: Developer selects Option 2 ("Cross-model review") or Option 3 ("Local model review") at the Review Choice prompt.
- **Scope**: Evaluates all task commits, diffs against base branch, audit disclosures, and integration test traceability in `.devtool/epic/<epic_dir>/`.
- **Execution**:
  - External CLI (`--auto-quota` or explicit `--cli` following priority `Claude -> Codex -> Kiro -> Argon`):
    ```bash
    python3 skills/dev-implementation/resources/scripts/run_external_review.py run \
      --epic-dir <epic_dir> --base <base_branch> [--auto-quota | --cli <claude|codex|kiro|agy>]
    ```
  - Local Ollama:
    ```bash
    python3 skills/dev-implementation/resources/scripts/run_external_review.py run \
      --epic-dir <epic_dir> --base <base_branch> --cli ollama --model <model>
    ```
- **Persistence**: Results recorded into `.devtool/epic/<epic_dir>/review_results.md`.

### 2. Ad-Hoc Mode (Arbitrary Git Branch or Commits)

Invoked on-demand to review a feature branch, pull request diff, or specific commit range without requiring an epic directory.

- **Trigger**: Developer invokes `/cross-model-review --base <base> [--commits <range>] [--auto-quota | --cli <claude|codex|ollama|kiro|agy>] [--model <model>]`.
- **Scope**: Evaluates the commit range and diff against `--base` (or `--commits <range>`).
- **Execution**:
  ```bash
  python3 skills/dev-implementation/resources/scripts/run_external_review.py run \
    --base <base_branch> [--commits <range>] [--auto-quota | --cli <claude|codex|kiro|agy>] [--model <model>]
  ```
- **Streaming & Visibility**:
  - The review output and formatted triage table stream directly to stdout.
  - Ephemeral review logs are written to `build/` (git-ignored) for live inspection.
  - No epic directory is required, and no `review_results.md` is created.

## Preconditions & Safeguards

1. **Clean Working Tree & Fast Mechanical Pre-Checks (< 5s, 0 Credits on Failure)**:
   - The runner enforces a clean git working tree before invoking any model (`commit or stash your changes before a review`).
   - Before invoking any external reviewer CLI, `run_external_review.py` runs fast mechanical pre-checks:
     - **Max 3 Rounds Guard**: Refuses a 4th round unless `--force` is passed (`[PRECONDITION BLOCKED] — Max review rounds (3) reached`).
     - **Gate 4 Epic Audit (`audit_epic.py`)**: Including the **Anti-Stub Enforcer** (`! git grep -n -E "def .+_stub\(|raise NotImplementedError|# STUB" src/`).
     - **Python Mechanical Gates** (when `.agents/quality/python-quality.json` is present): `anti_stub`, `importlinter_coverage` (`verify_importlinter_coverage.py`), and `self_probe` (`pre_review_self_probe.py`), plus Blast-Radius Scoped `Integration results:` and `check_bdd_traceability.py tests --results`.

2. **Quota-Aware Provider Selection (`check-quota`), 3-Round Convergence Funnel & Shared Context**:
   - **Provider Priority (`Claude -> Codex -> Kiro -> Argon`)**: Configured in `.devtool/review_config.json` (`provider_priority: ["claude", "codex", "kiro", "agy"]`, `max_rounds: 3`). With `--auto-quota`, `run_external_review.py` selects the highest-priority ACTIVE provider meeting thresholds: Claude/Codex (`5H >= 15%` AND `7D > 2%`), Kiro (`Monthly Credits >= 3%` AND `>= 15.0 credits`), or Argon (`Argon Weekly >= 10%` or `5H >= 15%`).
   - **3-Round Scope-Narrowing Convergence Funnel (Max 3 Rounds)**:
     - **Round 1 (`Scope: full`)**: Full audit of `base...HEAD` across touched modules (enforcing the **Exhaustive Per-File Pass** anti-drip-feed rule).
     - **Round 2 (`Scope: delta`)**: Verifies `fixed in <sha>` items and scopes diff to `Round1_SHA..Round2_SHA`; prohibits new `Minor` findings on untouched code.
     - **Round 3 (`Scope: convergence`)**: Strict Blocker-Only Convergence Mode (incomplete fixes or new `Critical`/`Important` regressions in `prev_sha..curr_sha`).
   - **Shared Context & 4 Anti-Fix-Loop Prompt Strategies**: `.devtool/epic/<epic_dir>/review_results.md` serves as the persistent Shared Context SSOT across rounds (`Earlier Rounds Shared Context Ledger`). `delta_round_context()` injects the **Universal Anti-Ping-Pong Guardrail** (both single-model and mixed-model), the **Previously Fixed Locations Registry (`Verify-Then-Close`)**, and **Diff-Hunk Isolation** on fixed files. See [review-triage.md](references/review-triage.md) for the complete 4-strategy specification.
   - CLI invocations run in read-only sandboxes under isolated process groups (`--timeout` default 1800s); any working-tree mutation discards the round immediately.

## The Triage & Fix Loop

Every review round is triaged and verified against active code and the approved design before taking action. Follow [review-triage.md](references/review-triage.md) in full:

1. **Check Decisions & Print Triage Skeleton**:
   ```bash
   python3 skills/dev-implementation/resources/scripts/run_external_review.py check --epic-dir <epic_dir>
   python3 skills/dev-implementation/resources/scripts/run_external_review.py triage --epic-dir <epic_dir>
   ```
2. **Verify Against Code & Approved Design (`d3nexus:receiving-code-review`)**:
   - Inspect exact files/lines and anchor strictly to the approved HLD, `bdd_scenarios.en.md`, and `Decisions Already Made`. Fill `## Top 3 Highest-Risk Findings (Executive Report)` and `## Findings` (`Valid`, `Debatable`, `False positive`, `Not shown`).
3. **Execute Auto-Triage Gate or Present Triage**:
   - **`AUTO_PASS_CLEAN` / `AUTO_PASS_FINAL_MINOR_SWEEP` / `AUTO_PASS_FINAL_SWEEP` (Round 3 with `0 Critical`)**: Fix any valid non-critical items in a single sweep, run Gate 4 green, and finish without calling another review round.
   - **`AUTO_FIX_ELIGIBLE` (Round 1..2 with `Critical`/`Important`, no `shared-contracts/` conflict, no Fix Loop)**: In autonomous mode (`lifecycle_runner.py`), the AI Programmer drafts `fix_plan_round<N>.md`. Role 2.2 (AI Assistant) evaluates the fix plan against approved HLD specifications emitting `APPROVED`, `ADJUSTED`, or `CRITICAL_CONFLICT`. Upon approval or adjustment, Role 2.3 applies fixes, commits them, records `fixed in <sha>`, runs Gate 4 in full, and advances to Round N+1. If `CRITICAL_CONFLICT` or `SHARED_CONTRACT_CONFLICT` is detected, the runner halts immediately for the Human Architect. In interactive mode, present the 4-bullet header block and wait for developer decisions:
     ```markdown
     Cross-Model Review Triage: Round <N> (<CLI>)

     • Candidate SHA: <commit_sha>
     • Reviewer: <CLI> ( <model> )
     • Thinking Level: <Light | Low | Medium | High | Extra High | None>
     • Verdict: <VERDICT> (<count> <Severity> findings)
     ```
   - **`CIRCUIT_BREAKER_HALT` (`SHARED_CONTRACT_CONFLICT` in any round, `FIX_LOOP`, or `MAX_ROUNDS_CRITICAL_REMAINING` when `>= 1 Critical` survives Round 3)**: **STOP IMMEDIATELY**, never auto-fix or revert user-approved HLD decisions, and report `## Top 3 Highest-Risk Findings (Executive Report)` to the Human Architect.

5. **Commit Fixes, Record Decisions & Automatically Re-Run Gate 4**:
   - When applying fixes (via `AUTO_FIX_ELIGIBLE` / `AUTO_PASS_FINAL_MINOR_SWEEP` or after developer confirmation), make each fix in its own commit: `[<EPIC_NAME>] Fix <what>`.
   - Update each Decision cell in `review_results.md`: `fixed in <sha>` or `dismissed: <reason>` in developer's exact words.
   - Commit decisions: `[<EPIC_NAME>] Record review decisions for round <n>`.
   - **Automatically re-run Gate 4 machine acceptance in full** on the new candidate SHA without asking for a separate confirmation before running Gate 4.
   - Record state via `python3 scripts/resume_context.py --state --epic <epic_slug> --status "Stage 4.1: Review Choice"` and emit next prompt via `python3 scripts/resume_context.py --emit-prompt finish --epic <epic_slug>` (or `--emit-prompt cross-review --epic <epic_slug>`).
   - In autonomous lifecycle mode, once Cross-Model Review converges (`AUTO_PASS_CLEAN` or `AUTO_PASS_FINAL_MINOR_SWEEP`) and Gate 4 is green, proceed to Finish (`finishing-a-development-branch`); in interactive mode, stop calling tools and present the applied fix summary, the new Gate 4 PASS verdict, and Review Choice again.

## References

- [Review Triage Protocol](references/review-triage.md): Verification discipline and decision cell standards.
- [Local Review Setup](references/local-review-setup.md): Guidance on setting up Ollama, Modelfiles, context budgets, and live logs.