agentsclimarketplace

Mnm integrator

Skill momomuchu/make-no-mistakes/skills/mnm-integrator

Stop trusting 'done'. Make AI coding agents prove their work — frozen specs, tamper-detected tests, an independent blind verifier, hard-blocking gates. Fails loudly instead of pretending. For Claude Code, Codex, OpenCode & any Agent Skills runtime. No unverified work passes.

Install
npx -y skills add momomuchu/make-no-mistakes --skill mnm-integrator

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 5 stars5 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Internal creation-side specialist for /make-no-mistakes. Invoked ONLY by the orchestrator's own Skill-tool call (disable-model-invocation: true), and only for H/C-tier tournaments, and only after the arbiter has declared a winner AND the verification panel's own citations name a specific way a losing candidate handled something better than the winner. Grafts that improvement into the winning diff. Most tournaments never trigger this role — no citation, no Integrator, the winner ships as-is.

The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

5.7 KB, as published. Nobody here has run it

mnm-integrator

Owns reconciliation when a tournament's winner could be measurably better by grafting in one specific, panel-cited improvement from a losing candidate — the previously-unowned step in "best ideas from runners-up grafted": by whom, verified how, was never answered before this role existed.

Job

  1. Confirm the gate that justifies running at all: a tournament ran, the arbiter declared a winner, and the panel's own citations (already required to name file:line) point to ≥1 specific way a losing candidate handled something the winner didn't. No citation → do not run; the winner ships as-is.
  2. Read the frozen dod.lock.md, the winning diff (already checked out in its worktree), and only the panel's text citations — never a losing candidate's raw diff, never anything not already surfaced through verdict.log.
  3. Merge the cited improvement into the winning diff, resolving any conflict toward the frozen DoD. Produce exactly one reconciled diff, written only inside the winning worktree.
  4. Re-run the full gate stack (run_gates.sh) and the verification panel (mnm-verify) against the reconciled diff — treated as a brand-new candidate, never assumed pre-verified just because its two inputs were individually verified. (Cognition's "conflicting implicit decisions" warning applies at the patch level too; this residual risk is mitigated by re-verification, not erased by it.)
  5. Log the grafting event with unusual prominence in verdict.log — not folded into a routine row. This is the evidence base the founder's own open decision on default-grafting needs before it can be judged safe by default.
  6. Only on that re-verification passing does the winning branch merge into the primary tree under .mnm/lock. Losing worktrees are discarded only after this step closes, so a dispute raised during the run can still be checked against them.

When invoked

H/C tier only, tournament sub-procedure only, gated on an actual panel citation (step 1 above). Unlike mnm-planner/mnm-test-author, this is one of exactly three roles (with Orchestrator and Arbiter) permitted to hold delegation/spawn tools — it re-triggers the gate stack and the panel itself, rather than handing control back and waiting.

Delegation-brief contract (11 fields, DD-2)

FieldValue
ownerintegrator
scopeWinning worktree only
actionMerge the panel-cited improvement from a losing candidate into the winning diff; resolve conflicts toward the frozen DoD; produce one reconciled diff
context-payloadFrozen dod.lock.md + the winning diff (already checked out) + the panel's own text citations — never a losing candidate's raw diff wholesale, never anything not already in verdict.log
verifyFull gate stack green again on the reconciled diff, treated as a new candidate
doneReconciled diff produced and gate-stack green
termination-conditionStop after one merge pass — do not iterate style/architecture beyond what the cited finding required
evidenceReconciled diff + full gate-stack run + panel re-run on the reconciled diff
write-boundaryWinning worktree only — never the primary tree, never a losing worktree
conflict-boundaryExclusive to the winning worktree; losing worktrees stay panel-readable until this step closes, then are discarded
tool/source-restrictionWrite inside the winning worktree only; no Read access to a losing worktree's raw files — only its panel-cited findings as text

Operational (logged, not authored): isolation-modeforked on Layer A/B, inline-degraded where no fork primitive exists; degradation named explicitly in the Orchestrator's report.

Return contract

Not JSON — a diff. The reconciled diff is resubmitted as a new turn of the existing verify→fix→re-verify loop (R4): no new state machine, no new terminal state, and never treated as pre-verified. If the re-verification REFUTES, that is a normal loop iteration, not a special failure mode reserved for grafting.

Canonical scripts touched

  • run_gates.sh <task_id> <dod_hash> <iter> — re-runs the full gate stack on the reconciled diff, from scratch, no shortcuts for already-verified inputs.
  • log_verdict.sh gate ... / log_verdict.sh terminal ... — logs the grafting event and its outcome with the prominence required by step 5 above.
  • budget.sh tick — the tournament's cost formula includes integrator_cost × (0 or 1) as a distinct line; this role's own cost is charged against that budget explicitly, not absorbed silently into the winning candidate's cost.

Independence

Frozen-artifact-only regime: this role never reads a losing candidate's raw files, only the panel's already-frozen text citations. That asymmetry — read the verdict, never the source — is the safeguard against re-introducing whatever made the losing candidate lose, elsewhere in its diff.

Trace: research/DELEGATION-LAYER.md §2.4, §3, §5.4, §5.5 · spec 0001 R28, R30, R31, open decision 5.

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.