parallax

parallax

Cross-model verification for Claude Code. Two equal-weight frontier models — the session model and a cross-vendor reviewer (via the OpenAI codex CLI; the current reviewer lane is declared in the skill's model-prompting-notes.md) — verify and refute each other's claims with file:line evidence before a cheaper implementer touches code, and again before the result merges. Neither vendor grades its own homework.

Companion to superpowers, not a replacement: it fills the cross-model review gap superpowers rules out of scope.

How it works

flowchart LR
    A[superpowers<br/>brainstorm] --> B{{"mode plan debate<br/>session ⇄ reviewer"}}
    B -->|converged| C["frozen plan<br/>+ debate record"]
    B -->|escalated| U[user decides]
    C --> D["implementer subagent<br/>(zero judgment calls)"]
    D --> E{{"mode diff debate<br/>spec + port fidelity"}}
    E -->|PASS| F([merge])
    E -->|FIX| D
    E -->|ESCALATE| U

• Mode plan — after brainstorming, before the implementation plan is written. The two models debate the approach, port-fidelity claims, and the API/behavior risk register until convergence or the round cap, then the converged plan is frozen with a full debate record (participants, rounds, resolved/struck/escalated points, verification status).
• Mode diff — after implementation, alongside superpowers code review. A PostToolUse hook fingerprints the superpowers code-reviewer dispatch and injects the diff-mode reminder with the same base/head SHAs, so both reviews always look at the same range. Verdicts are PASS / FIX / ESCALATE from each side.

The debate rules that keep this honest (skills/multi-model-verify/references/debate-protocol.md):

• Strike rule — every externally checkable claim carries a citation the other side can read (References/<addon>/<file>:<line>, API docs, a dated probe). Uncited claims are struck, not debated.
• Equal weight — disagreements resolve by evidence, never by which model said it. Unresolvable points escalate to the user with both positions.
• Round cap — 4 exchanges by default; convergence in one round on a sound plan is the system working, and manufactured objections are a protocol violation, not diligence.
• Session continuity — the reviewer keeps debate state across rounds via codex exec … resume <SESSION_ID>; the full context is sent once.
• Final adjudication — the session always has the last step: it verifies the reviewer's final round against the repo and emits the terminal verdict itself. A reviewer PASS/FIX is input, never the decision; genuine deadlocks escalate to the user. The rule attaches to the role, not the model — it holds whoever fills the session seat.

What's in the box

Piece	What it does
skills/multi-model-verify/	The debate skill: both modes, debate protocol, frozen-plan format, model prompting notes, fallbacks/consent gate
hooks/	PostToolUse + PostToolUseFailure hook (matcher Task|Agent): fingerprints the superpowers code-reviewer dispatch, injects the mode-diff reminder with matching SHAs; inert everywhere else
agents/implementer.md	Zero-judgment executor for frozen-plan tasks, pinned to the cheap lane (currently model: sonnet)
commands/drift-triage.md	/parallax:drift-triage — reads the newest drift report, verifies each finding against the live contract surfaces, repairs on a branch
commands/doctor.md	/parallax:doctor — operational health check: checkout-vs-installed version, hook registration, superpowers fingerprint, codex transport round-trip, drift task + pending entries. Reports, never fixes
evals/	Four gate tiers for the skill itself — see Verify
tools/check-drift.ps1	Weekly drift watch over the three upstreams the contract depends on — see Drift protection
tools/write-attestation.ps1 · tools/verify-attestation.ps1	SHA-bound review attestations — see Attestation lane
.githooks/pre-push	Non-blocking attestation check on main pushes (git config core.hooksPath .githooks to enable)

Fails loud, never silent

The governing rule (skills/multi-model-verify/references/fallbacks.md): no transition that reduces vendor diversity, evidence quality, or conversation continuity happens without explicit user consent.

flowchart TD
    X[codex failure] --> R{one same-parameters retry}
    R -->|recovered| OK[debate continues]
    R -->|still failing| G["CONSENT GATE banner:<br/>what failed · what degraded mode<br/>would and would NOT verify"]
    X -. "quota-exhausted:<br/>skip the retry" .-> G
    G -->|fix codex| OK
    G -->|run degraded| D["single-vendor, visibly flagged;<br/>DEGRADED plan poisons any diff PASS"]
    G -->|abort or unattended| B2[BLOCKED / DEGRADED-NOT-AUTHORIZED]