parallel-adversarial-review

parallel-adversarial-review

Two skills for adversarial code review, plus an eval suite.

What's in here

skills/parallel-adversarial-review/

The original PAR pattern, ported from iterative-development. Two same-model reviewer subagents run in parallel under a competitive scoring frame; their findings are aggregated, with the worst severity winning on disagreement.

Use this for routine review.

skills/multi-model-adversarial-review/ (MMAR)

A three-stage pipeline that uses multiple installed coding-agent CLIs as independent reviewers, then runs a cross-critique grid where each reviewer evaluates the others' findings (catching hallucinations and severity inflation), then synthesizes a final deduplicated report.

Stage 1: parallel reviews   (each CLI reviews independently)
Stage 2: cross-critique     (each CLI verifies other CLIs' findings)
Stage 3: synthesis          (one model merges everything, applies rules)

Use this for high-stakes review (security, pre-merge on hot-path code, audits). Costs more.

The driver is scripts/mmar.py. CLI invocations are configured in scripts/adapters.toml so flags can be fixed when CLIs change without touching code.

$ python3 scripts/mmar.py list
amp       DISABLED  installed   amp
claude    ENABLED   installed   claude
codex     ENABLED   installed   codex
droid     DISABLED  installed   droid
gemini    ENABLED   installed   gemini
opencode  ENABLED   installed   opencode
pi        ENABLED   installed   pi

$ python3 scripts/mmar.py review path/to/diff_or_file_or_dir \
    --reviewers claude,codex,gemini \
    --out ./.mmar/run-1

Default-on tier: claude, codex, gemini, pi, opencode — enabled if installed.

Opt-in tier (enabled=false by default): amp and droid (Factory). Flip to enabled=true in adapters.toml after configuring credentials (amp login / Factory account).

For evals/CI, replace live CLI invocations with pre-recorded responses:

$ python3 scripts/mmar.py review evals/fixtures/001-sql-injection/input \
    --reviewers claude,codex,gemini \
    --mock-dir evals/fixtures/001-sql-injection/mocks \
    --out /tmp/mmar-run

Eval suite

Fixture-based eval that scores recall and precision against planted defects.

$ python3 evals/runner.py --mode mock        # cheap, deterministic, CI-safe
$ python3 evals/runner.py --mode live        # real CLIs, costs $$

Current fixtures:

• 001-sql-injection — classic f-string SQLi, with a parameter-bound query nearby that one reviewer hallucinates as also injectable (cross-critique drops it)
• 002-off-by-one — windowed_sum loop overruns by one; mocks include a critic-driven severity downgrade
• 003-clean — negative case, no defects; tests false-positive rate (one reviewer hallucinates a generic "could be passed a large string" worry, critics drop it)
• 004-resource-leak — file handle leaked on exception path; gemini's mock misses it as a serious issue, aggregation still surfaces it

Pass thresholds: recall ≥ 0.8, precision ≥ 0.7. Negative-case fixtures pass iff zero false positives.

$ python3 evals/runner.py --mode mock
fixture                       truth  found   tp   fp   fn   prec    rec     F1  result
-----------------------------------------------------------------------------------------------
001-sql-injection                 1      1    1    0    0   1.00   1.00   1.00  PASS
002-off-by-one                    1      1    1    0    0   1.00   1.00   1.00  PASS
003-clean                         0      0    0    0    0   1.00   1.00   1.00  PASS
004-resource-leak                 1      1    1    0    0   1.00   1.00   1.00  PASS

aggregate (positive cases): precision=1.00  recall=1.00  f1=1.00
passed: 4/4

Tests

$ python3 -m unittest discover -s tests

15 unit tests covering finding parsing, truth matching, and adapter loading/mock invocation.

Adding a CLI

Edit scripts/adapters.toml:

[my-new-cli]
enabled = true
binary = "my-new-cli"
argv = ["--print"]
prompt_via = "argv"      # or "stdin", or "argv-after-flag"
prompt_flag = "--prompt" # only with argv-after-flag
timeout_sec = 300
notes = "..."

The driver picks it up on the next run.

Adding a fixture

evals/fixtures/<id>/
  input/<files>            # code under review
  truth.json               # planted defects (see evals/README.md schema)
  mocks/
    stage1/<reviewer>.txt
    stage2/<critic>__on__<reviewed>.txt
    stage3/synthesizer.txt

For --mode mock you only really need a realistic stage3/synthesizer.txt for scoring; stage1/stage2 just need to exist so the driver runs through.

Layout

.claude-plugin/plugin.json
skills/
  parallel-adversarial-review/SKILL.md, reviewer-wrapper.md
  multi-model-adversarial-review/SKILL.md, reviewer-wrapper.md, critic-wrapper.md, synthesizer-prompt.md
scripts/
  mmar.py            # driver
  adapters.py        # CLI invocation
  adapters.toml      # CLI config
  findings.py        # parsing + scoring
evals/
  runner.py
  fixtures/<id>/...
tests/
  test_findings.py
  test_adapters.py