Gap Trap Agent Skill

gap-trap

Sets up a quality framework for repositories where agents write most of

the code. Every rule the agents follow has a gate that fails a commit, a

push, or a CI run when the rule breaks. A rule a script could check but

that names no gate is a defect. The framework is the one described in

reference/framework.md; read it before either mode.

Two modes: setup builds the framework in a repo that has none.

refine audits a repo that has it and proposes the next rules and gates

from what broke. Bare gap-trap with no keyword means setup when the

repo has no AGENTS.project.md, otherwise refine. A hand-written

AGENTS.md or CLAUDE.md on its own is input to setup, not a sign the

framework exists.

Model gate

This skill runs on Opus or a more capable model (Opus 4.8 or newer, Opus

5, Fable, Mythos). Discovery decides which parts of a codebase get a

contract, and a wrong contract costs every later session. Your system

prompt names your model.

If it is Sonnet, Haiku, or anything below Opus, stop. Read nothing,

write nothing. Say exactly this and end the turn:

> gap-trap runs on Opus or a more capable model. You are on

> <model name>. Switch with /model, then run gap-trap again.

Do not dispatch Opus subagents from a smaller orchestrator, start "the

easy parts", or offer a lighter version.

setup

• Discover. Follow reference/discovery.md. It produces a plan

file in your scratch directory: stack, commands, CI provider,

contract candidates with real symbol names, the first domain facts

from history (the revert probe runs at any commit count), and one

row per rule found in an existing instruction file with where it

goes and what gates it.

• Confirm once. Show the plan as a short list: the contracts you

will write (at most six), the gates you will add, the test and lint

commands you found, the migrated-rules table, and where slop-mop

goes. A build, type check, or

lint that is already red is the first item: setup fixes only what

its own gate files need (a missing type package, a lockfile) and

records the rest as a ratchet count or an out-of-scope entry. Ask one

round of questions, only about what discovery could not settle. Then

proceed.

• Install slop-mop. git clone https://github.com/pliablepixels/slop-mop.git

into your scratch directory and copy slop-mop/slop-mop to

~/.claude/skills/slop-mop (skip if present). It is not vendored into

the repo; the project rule in AGENTS.project.md requires it.

• Write the instruction files from templates/. AGENTS.md and

agents/generic/claude-workflows.md are copied unchanged, with one

exception: a rule whose gate cannot exist in this repo (I3's

accessibility lint and C3's locales in a repo with no UI) gets its

Gate: clause replaced by Gate: n/a, no <thing> in this repo.

Nothing else in AGENTS.md changes. AGENTS.project.md,

CLAUDE.md, and agents/project/ are filled from the plan; drop the

UI and end-to-end lines and {{E2E_CMD}} when the repo has no UI, and

delete every placeholder you cannot fill with a verified value.

Contracts name real symbols only. Existing instruction files

(AGENTS.md, CLAUDE.md, .cursorrules,

.github/copilot-instructions.md, CONTRIBUTING.md) are migrated

from the plan's migrated-rules table, one row at a time: a contract

row becomes a contract block, a rule row becomes a project rule with

its gate name or Gate: review, a playbook row goes verbatim into

the named agents/project/ file, a command row fills the

verification block, and a duplicate or dropped row is left out. No

rule is dropped silently; the confirm step showed the table. The

old file is replaced by the template that owns its name, and

CONTRIBUTING.md stays where it is. The user sees the diff.

• Write the gates from reference/gates.md, in the repo's own test

runner and CI provider: the instruction gate, proven red, the PR body

check, the ratchet (always, seeded with the counters the repo

supports: files over the C2 length, lint problems per rule when a

linter runs, each contract Never clause that has violations today),

and the mutation smoke when discovery named a risky module (auth, a

parser, the API client); otherwise refine adds it after the first

fix chain. The stack table there picks the reference: native ports

for Node and Python, shell references that need only git, grep, and

awk for every other language. Set the instruction gate's word budget

at the current count plus half, rounded up to the next 500.

• Prove each gate red with a scratch violation, remove the

violation, run the repo's full test command, and commit one logical

change per commit. Do not push unless asked.

• Report what was written, the instruction word count against the

budget, which contracts say Gate: review and why, and the one

next step: run gap-trap refine after the first incident or in a

month.

• Offer a check. End the report with one question: check the repo

against the new framework now? On no, stop. On yes, change no files

and list, grouped by contract or rule ID:

• each ratchet counter, its count, and the file and line behind each

unit of it;

• each Gate: review contract: read the code its Path: and

Never: lines cover and list suspected violations with file and

line, marked unverified because no gate checks them;

• the red build, type check, or lint items recorded in step 2.

Fix nothing in this pass. A fix the user picks is its own commit,

and a fix that lowers a count reruns the ratchet with --update.

Instruction files are read by agents. Write them as terse statements

with IDs and gate names. Human-voice prose belongs in the repo's docs,

not in AGENTS*.md or agents/.

refine

Follow reference/refine.md. It audits rules without gates, contract

names that no longer exist, and the commit history since the last

refine, and it proposes edits through the self-improvement protocol as

one PR. It removes rules with no incident behind them as readily as it

adds gates.

Common mistakes

| Mistake | Fix |

|---|---|

| A contract with a guessed symbol name | Every name in a Path: line is greppable in the tree; the instruction gate fails otherwise |

| Writing ten contracts on day one | Start with the places that have one sanctioned path and a violation today; add the rest through refine |

| A gate that passed on first run | Prove it red with a scratch violation first; a gate that never failed proves nothing |

| Vendoring slop-mop into the repo | It installs per agent from its repo; the project rule requires it |

| Instruction prose polished for people | Agents read it. Statement, one-clause why, gate name |

| Skipping the model gate because the repo is small | Discovery quality is the deliverable; the gate is first for a reason |