C1 spike: review methodology comparison + X1 benchmark + headless probe

[SSFSKIM]

Problem & intent

Before argus-review becomes the default review methodology, we must (a) learn what Claude's native review methods know that argus doesn't, and (b) build the measuring stick every later "no regression" claim stands on. This is child C1 of docs/doperpowers/specs/2026-07-26-claude-review-stack-roadmap.md — a SPIKE: deliverable is findings + a benchmark, never a merge to the stack.

Constraints

• Compare argus-review v0.3.0 against: the official code-review plugin (~/.claude/plugins/marketplaces/claude-plugins-official/plugins/code-review/) and the built-in /review + /code-review sources (~/Developer/GitHub/codex_somersault/Claude Code Src/src/commands/ — review.ts, review/ incl. ultrareview, security-review.ts). Mechanism AND content (rubric wording, lens texts, false-positive guidance, confidence rubrics).
• The codex-app detached path is settled by argus references/provenance.md — do NOT re-examine (human ruling; guardian review is a tool-use classifier, cloud-tasks is not the app review path).
• The bench's argus runs go through the headless invocation path (G3's harness) — the context C4 will deploy.

Success criteria

• G1: findings document — per methodology: mechanism map, content-level comparison, adoption candidates for argus each marked adopt/reject with reasons.
• G2: the X1 benchmark exists (seeded-bug diffs + a few real previously-reviewed PRs; scoring = seeded-bug recall + false-positive count; codex baseline stable across two runs) and has produced baseline numbers for BOTH the codex engine and argus v0.3.0.
• G3: headless-invocation probe — a daemon-style non-interactive argus invocation (pinned level) produces the Findings/Verdict contract captured to a file within a bounded wait. Failure flows back to the roadmap before C2 dispatches.

Open questions

• Bench set composition/size — C1 fixes it (grow until the codex baseline is stable across two runs).
• What "meaningful false-positive growth" means — C1 fixes it in X1; later children re-litigate none of it.

Decision log

• Bench defined inside C1, not at C4 time: the codex baseline must be measured while codex is still the default (roadmap Decision Log).
• Headless probe front-loaded here from C4 (external review F2): it depends on nothing downstream and de-risks the whole chain.

[SSFSKIM]

[board] ready-for-agent: roadmap C1 — dispatchable now; worked in-session

[SSFSKIM]

[gate] pass — well-defined: roadmap C1 section + ticket gates G1–G3 leave no unowned fork (bench composition is itself a C1 deliverable, X1). well-scoped: one spike ownable by one session. Worked in the roadmap session (registrar=worker); no daemon bind.

[SSFSKIM]

[board] in-progress: C1 spike started in roadmap session — order: G3 probe → G1 comparison → G2 bench

[SSFSKIM]

[progress] G3 headless-invocation probe: PASS. claude -p "/argus-review:argus-review …" --permission-mode auto from a fixture repo triggered the skill, dispatched the plain argus-reviewer subagent, and captured the full ## Findings / ## Verdict contract to a file (rc=0, ~10 min wall on an 11-line diff). Both seeded bugs found (empty-shipment regression P1, >= boundary P2); both intentional-change baits correctly not flagged (0 FP). Caveats recorded for C4/X2: (1) wrapper prose surrounds the verbatim block — the engine consumer must tolerantly extract ## Findings…## Verdict (native codex parses tolerantly too); (2) findings cite absolute paths in this fixture run; (3) ~10 min wall on a tiny diff — timing budget data for C3 auto-routing and C4's engine wait bound. X2 feasibility is proven; C2/C3 may proceed on this mechanism.

[SSFSKIM]

[findings] C1 spike complete — all three gates met. Full record: docs/doperpowers/specs/2026-07-26-c1-review-methodology-findings.md (commit f915176) + tests/review-bench/results/2026-07-26-baseline-r1/scores.json.

G1 (comparison): four native pipelines mapped (built-in /review, /code-review, /ultrareview, official plugin — /code-review extracted from the 2.1.220 binary; the somersault checkout predates it). 9 adoption decisions for C2: ADOPT lint/CI-catchable FP exclusion, language-pitfall + wrapper/proxy lens pressure, Conventions/CLAUDE.md lens (engine-purity flag); REJECT 0-100 scoring, cleanup angles, finding floors, Haiku verification; ROUTE diff-size-scaled routing → C3 (ceil(lines/150) prior art), sweep enrichment → C2. argus's structural keeps: full isolation (built-in runs in a conversation-inheriting fork), P0-P3 severity (load-bearing for the loop verdict), adversarial max pass, injection defense.

G2 (X1 baselines, 17 seeded bugs): codex gpt-5.6-sol xhigh 17/17, FP 1 [P1], identical across two runs (stability met). argus plain (Opus 5, effort high, headless) 16/17, FP 0 — sole miss is the L4 ssh injection (plain has no security lens). argus high 17/17, FP 2 [both P3] — injection found. Bar verdict: argus PASSES at high (recall tie, FP +1 allowed, and its FPs are verdict-neutral where codex's [P1] FP is verdict-flipping). C3/C4 consequence: engine context deploys high; plain stays the fast interactive default (3-5x faster, zero FP).

G3 (headless probe): PASS — claude -p slash invocation under --permission-mode auto captures the Findings/Verdict contract; caveats for C4: tolerant block extraction, absolute paths, and one live risk — the reviewer EXECUTED repo code under auto permissions (conduct text lacks the verifier's never-execute sentence) → C2 prompt fix + C4 sandbox posture decision.

Follow-ups registered: #36 (fixture maintenance before C2's re-run). Recommended close: done (spike, non-PR).