Cheffy

| # | Gate | Level | Check required |
|---|---|---|---|
| 1 | Proof | MUST | yes |
| 2 | Tests | MUST | yes |
| 3 | Repo gates | MUST | yes |
| 4 | Zero redundancy | MUST | no |
| 5 | Scope | MUST | no |
| 6 | Taste | SHOULD | no |
| 7 | Critic | MUST when the Critic trigger applies | no |
| 8 | Comments | MUST | no |
| 9 | Prose | MUST when the run changes prose files | yes |
| 10 | Trace | MUST | no |
| 11 | Security | MUST | yes |
| 12 | Non-functional | SHOULD | no |

What it does

A coding agent often tells me that a task is done before anything proves it. Cheffy adds that proof. It is the head agent of Hidkit, my quality-first delivery harness for coding agents, which runs as a Claude Code plugin.

I start it with /hidkit:cheffy <task>. Its SKILL.md sets disable-model-invocation: true, so it loads only when I call it. It stays on until I say “stop Cheffy”.

Recipes and lanes

Cheffy matches each task to one recipe: bug-fix for a reported defect, feature for new or changed behavior, and eval for a small blind comparison of variants. Then it picks the lightest lane whose conditions all hold:

• Quick: one file, 20 lines or fewer.
• Light: 3 files or fewer, 80 lines or fewer, and no trust boundary. Cheffy follows the recipe inline, with no delegation.
• Full: every other task. The recipe runs with its roles.

The lane is a one-way ratchet. Cheffy never moves a task to a lighter lane.

Roles, the pass, and the ledger

Cheffy delegates only for isolated judgment, for bulk reading, or for slices that run in parallel. The five roles are the investigator, the implementer, the critic, the verifier, and the judge.

Before it reports done, Cheffy runs the pass in the snippet. A gate that requires a check cites the ids of passing check events. One FAIL gate makes the verdict FAIL.

Gate 11 runs a secret scan, a dependency scan, and a SAST scan on every change. Only a person can approve a waiver.

Every run writes a ledger: an append-only file of the delegations, checks, and decisions. Cheffy runs each verification command through trace check, which records the exit code, an output hash, and the output path.

Where it stands

Hidkit is in Phase 1, and it is not a release. The exit criterion is quality above plain Claude Code at 2 times the baseline cost or less. That criterion is not met yet.

In the last measurement, Cheffy had no regression on any task. Its quality gain (0.10) was below the noise (0.20), and its cost was 3.6 to 4.5 times the baseline. I publish these numbers because the measurement is the point.

Hidkit derives the principles, recipes, and pass rules of Cheffy from pstack, under the MIT License.