cross-model-convergence

Cross-Model Convergence

The lead — whichever agent and harness is running this skill — and an external reviewer alternate find → verify → fix rounds until a full round is clean. Agreement is not the terminator; evidence is. A clean round, not a compliment, ends the loop.

The reviewer must be a different model from the lead. Calling your own model's CLI is self-review; a single-model clean round never counts as double-clean.

Announce at start: "I'm using the cross-model-convergence skill with [reviewer] on [problem class]."

REQUIRED SUB-SKILLS: razorback:architecture-quality (Audit Mode), one reviewer channel (razorback:codex-cli, razorback:gemini-cli, razorback:claude-cli, or razorback:cursor-agent in read-only mode) pointing at a model that is not the lead, razorback:receiving-code-review, razorback:test-driven-development, razorback:verification-before-completion.

Setup (Round 0)

Write these down in the conversation before the first sweep. They cannot be changed mid-loop.

• Problem class: default is architectural cleanup via architecture-quality Audit Mode. The user can name another class (security, error handling, dead code); the sweep method changes, the loop does not.
• Reviewer: any external CLI skill whose model differs from the lead — codex-cli, gemini-cli, claude-cli, or cursor-agent. A Claude lead defaults to codex-cli; a Codex lead picks claude-cli, gemini-cli, or cursor-agent. cursor-agent reviews read-only (--mode ask with --trust, never --force).
• Severity floor (the nitpick line). Defaults — a finding does NOT count when it is: pure style or naming with no maintainability impact; an unmeasured micro-optimization; speculative strength; blocked by a heuristic's "do not act when" line; or a churn-heavy rewrite with marginal benefit. Disputed-but-real items are not nitpicks; they go in the report.
• Round cap: default 4 (initial audit + 3 fix rounds). The cap bounds discovery, not fixing of already-verified defects.

The Loop

1. Sweep. Run the problem-class audit (Audit Mode for architecture). Verify your own findings with Miller inspect/trace evidence before sending — rigor is symmetric, not just for the reviewer's claims.
2. Dispatch to reviewer: "Here are N verified findings: [list with file:line and evidence]. (a) Verify or refute each, naming what you checked. (b) Independently hunt for problems in the same class that this list misses." Reviewer dispatches are read-only — state "verify and report only; do not modify any file" and diff-check the worktree after each reviewer call — revert any reviewer edit and note it; it does not consume the cap. If the channel is stateless, re-supply the merged list and prior fixes inline each round; use session resume where the channel offers it (e.g. cursor-agent create-chat/--resume).
3. Triage the response with receiving-code-review rigor: confirm each reviewer finding in code before accepting it. Unconfirmable claims get pushed back with evidence, not implemented. Reviewer evidence need not be Miller-based — file:line plus reasoning counts; the lead re-verifies with Miller either way.
4. User gate — once, after round 1: present the merged, deduplicated list with strength ratings; get approval of the fix scope. Under an unattended goal-driven run, this gate is satisfied by the pre-approved Setup plus recording the merged list (Goldfish checkpoint) instead of prompting mid-run.
5. Fix approved items with TDD; verify with verification-before-completion.
6. Re-review round: your own re-sweep, plus the reviewer re-prompted adversarially: "For each fix, name the file and how you verified it. Then name the weakest remaining areas in this class regardless of the fixes."
7. Termination check (below). Checkpoint with Goldfish at the end of every round — round number, merged list, approved scope, fixed/disputed/parked items, and the reviewer session handle (e.g. cursor-agent chat id) — so the loop survives compaction and goal-driver re-runs; every re-run starts by recalling that state.

New findings inside the approved problem class and above the floor are fixed in the next round without re-asking. Reviewer proposals beyond the approved scope (new modules, redesigns) are evaluated, ranked, and parked as candidates for the final report — they do not count as open issues and cannot hold convergence hostage.

Termination

DONE only when one full round is double-clean:

• every approved finding is fixed and verified (tests green, original finding no longer reproducible), and
• your own re-sweep finds nothing new above the floor, and
• the reviewer's substantiated re-check surfaces nothing new above the floor —

all in the same round.

Content-free approval is not a clean signal. If the reviewer replies with praise and no specifics ("all fixes look great!"), it didn't look. Re-prompt requiring per-fix verification detail before counting the round clean. One substantiated re-check is required; don't loop the reviewer for compliments.

At the cap: stop discovering, report honestly — what was fixed, what remains with severity, and a recommendation. Hitting the cap is a status report, not a failure. Exception: a verified real defect found at the cap still gets fixed, followed by one bounded confirmation pass (reviewer verifies that fix only — no new discovery sweep).

Failure Handling

• Reviewer unavailable or erroring: swap to another external reviewer skill and say so — the floor and cap are load-bearing, the reviewer identity is not. No reviewer available? Report blocked; a single-model clean round never counts as double-clean.
• Reviewer stays content-free: two re-prompts maximum. Still no specifics? Treat that re-check as failed, report it, and stop — don't loop for compliments.
• Disputed finding: one push-back cycle with evidence each way. Still disputed? Record it as disputed in the report and move on — disputes don't count as open-above-floor and don't block convergence.
• Fix-induced regression: a defect your fix introduced is yours regardless of problem class — fix it; it does not consume the discovery cap.
• User approves nothing at the gate: the run ends as an audit — deliver the merged findings report, no fix loop. Partial approval shrinks scope to the approved subset.
• Failed confirmation pass at the cap: one re-fix and re-confirm. Still failing? Revert or report it as remaining — do not start a new discovery loop.

Doubt Pass (pre-implementation)

A bounded, fix-free variant for challenging a design decision BEFORE anything is built — when course-correction is cheapest. Run it during brainstorming when the architecture-quality gate rates risk medium/high, or whenever the user asks to doubt or challenge a decision. Low-risk designs skip it.

1. CLAIM: state the decision and its load-bearing assumptions in 3-5 sentences — what was chosen, what was rejected, why.
2. DOUBT: send the claim to an external reviewer (same rules as the main loop: different model from the lead, read-only): "Try to refute this decision. Attack the assumptions, name concrete failure modes, and say what you would choose instead and why — specific to this repo's constraints."
3. RECONCILE: triage with receiving-code-review rigor — verify each refutation against the codebase (Miller) and the requirements. A refutation that survives verification changes the design; one that does not gets recorded with the evidence against it.
4. STOP: maximum 3 doubt cycles; stop early when a cycle produces no new surviving refutation.

The Doubt Pass adds no user gate and no stop point: it runs inside brainstorming's existing flow, and its outcome — the decision, surviving objections, and how each was resolved — lands in the design doc that flows to the already-existing spec review. It never implements anything.

Goal-Driver Integration

The goal predicate, for Claude Code's /loop, Codex's goal command, or any until-condition runner:

Terminal state reached: either (a) one full round is double-clean — zero new findings above the written severity floor from both the lead's re-sweep and the reviewer's substantiated re-check, all approved fixes verified — or (b) the round cap is reached and the final status report has been emitted.

The cap must be part of the predicate: a success-only goal loops forever on a codebase that does not converge.

Emit this block at the end of every round — print it in the round output and include it in the round checkpoint — so the driver (and the user) can see state at a glance; a driver's check can key on clean_round: yes or round: <cap>/<cap>:

CONVERGENCE STATUS
round: 2/4
fixed_this_round: 3
open_above_floor: 1
parked_out_of_scope: 2
clean_round: no

For unattended runs, Setup (problem class, reviewer, floor, cap) must be user-approved before the driver starts; the loop then runs to the goal or the cap without prompting, recording instead of asking.

Red Flags — STOP

Rationalization	Reality
"The reviewer said it looks great, we're done"	Praise without specifics means it didn't look. Re-prompt for evidence.
"This finding is just a nitpick" (decided mid-round)	The floor was written in Setup. Judge against it; don't redraw it.
"One more round will finish it" (at the cap)	The cap is the cap. Report remainders.
"The reviewer suggested it, so I should implement it"	Out-of-scope proposals get parked, not implemented.
"We've both gone quiet, that's agreement"	Convergence is a double-clean round, not mutual exhaustion.
"My own findings don't need verification"	Rigor is symmetric. Verify before sending.