We Ran Godmode on Godmode

Godmode sells one promise: maximum effort, verified. An 8-layer protocol, driven by a small command-line runner, so the model cannot quietly skip the boring steps.

This week we pointed that runner at itself. It did not go well, and that is exactly what made it useful.

📊 The Audit in Numbers

This was a two-round audit: first a full read of the runner's source, then live tests in a throwaway repo to prove every finding actually fires. Numbers below are from the real sessions, not estimates.

🪞 Why Point the Auditor at Itself

The 8-layer protocol works because a deterministic program, not the model's mood, decides when a layer counts as done. Layer 2 is not done until a scanner finds zero leftover stubs; Layer 3 is not done until the test suite actually passes.

But if the enforcement itself is loose, every green tick it prints is decoration. We have audited our own tools before and it hurt in useful ways, so godmode's runner was next in line.

💥 What the Audit Found

Round 1 read every line of the runner. Round 2 built a small scratch repo and made each suspected failure happen for real.

🔬 See the Worst One Yourself

The scratch repo had a committed legacy.js carrying a TODO from 2023, a Python helper, and two files the task actually touched. Click between the two versions to see what each one scans and flags.

On our live site repo the numbers were worse. One landing-page task triggered a scan of 2,325 files, surfaced 119 stale markers from old code, and demanded all of them reach zero before tests could start.

🔧 What v2.4.0 Actually Changes

Every fix follows one rule: the gate must verify the thing it claims to verify. Scope by what changed, refuse to run out of order, and never stamp a layer that was not actually checked.

Here is the ordering guard as a run actually sees it. This envelope is copied verbatim from our regression test:

{"state":"out_of_order","run_id":"gm-1783343082-4e68f9",
 "narrate":"Cannot run harden yet: `test` has not been run
 for this run. The 8-layer protocol is ordered; run
 `node bin/godmode test` first.","next":"test"}

A fixed run now moves through the gates like this:

1. start records the git commit you began from.
2. discover maps the repo and stamps Layer 1.
3. check scans only your changed files for leftovers, and stamps Layer 2 only when they are clean.
4. test runs the real suite and stamps Layer 3 only on a pass.
5. harden security-scans your diff and hands the findings to the model for triage.
6. ripple and polish stay inside the files you touched.
7. end compares the stamps against all 8 layers and tells the truth.

🎮 Try to Skip a Layer

This simulator runs the same guard logic as the real runner. Try clicking harden first, then walk the honest path.

✅ The Receipts: Eight Regression Tests

Fixes without proof would repeat the original sin. So v2.4.0 shipped only after eight regression tests, each targeting one audit finding, passed in the same scratch repo.

🧠 What We Left Broken, On Purpose

Honesty section. The audit found more than this batch fixed, and the leftovers are logged for a second pass.

The ripple gate's caller search is expensive on big repos and matches bare words, the global active-run pointer can be clobbered by two parallel sessions (we watched it happen mid-audit), and the file-ranking in discover once scored a stale JSON report above the CSS file literally named in the task.