tlc-implement

TLC Implement

Freeze the obligations. Free the build. Prove it with someone who did not build it.

┌────────┐   ┌───────┐   ┌────────┐
│ CHECKS │ → │ BUILD │ → │ VERIFY │
└────────┘   └───────┘   └────────┘
 obligations   yours       always

The common failure of a coding agent is a requirement that never became an active obligation, followed by a completion claim. What works against it is a small, frozen, external obligation set and a verifier that is not the author, so the budget goes there and not into choreographing the model. How you build is yours.

The input is an approved .specs/features/<feature>/plan.md from tlc-plan - approved means its **Approved**: <date> - user: "<their words>" line, which validate_checks.py requires - and it is the whole brief: you may have none of the conversation that produced it. A change under roughly three files with no one-way door needs no plan: write checks.md with an ## Intent paragraph instead. Anything larger with no plan, or a plan with an open blocks question, goes back to tlc-plan.

Rules

These hold even if no reference file is read. Each one lives here; the references point back.

1. Every check is one observable claim with a concrete value plus the proof that settles it: a test or command whose exit code decides, or, for a state a capture draws, a picture of it judged against the capture (rule 9).
2. Tests assert what the checks say, not what the code happens to do. Do not write a test by reading the implementation.
3. A test changes only with the claim it proves, and leaves with the code it proves, in the same commit. Changing a test to make a suite pass hides the failure it exists to show: a wrong check is a stop-and-ask.
4. Checks are fixed once validate_checks.py passes; a check may be appended when the walk or the Verifier finds a missing obligation. In the plan, Landing, Surface and the stored shape are additive; the path and Changes underneath are kept true in the commit that changes them.
5. The Verifier is a fresh sub-agent, never the author, and it always runs, once per feature, beside the screens sub-agent. check_proofs.py settles the mechanical part - every proof run and passing at HEAD, the build, the pictures, the walk, the mutations killed - and the Verifier spends its time on judgement: mutations where a wrong rule would stay green, examples nobody named, assertions looser than their criterion. The work is done when every finding is closed by a commit and proofs.md is green at HEAD.
6. The mutation count is declared in checks.md (Mutations:, 3 by default) and decides how many rules the Verifier probes, not how good the work is: code quality, the tests behind every check and a complete Coverage join are owed whatever N is. The checks are the floor; an edge case or guard no check names is still the work.
7. The completion gate is a script: validate_verification.py exits 0. The final report opens with its last line; a red gate is said first, never left for the reader to find.
8. An approved plan authorizes local edits and local commits. git push, deploy and production data changes need an explicit go-ahead for that action.
9. A state a capture draws is one picture. The author reaches that state and saves <capture>.running.png beside the capture from a browser window at the capture's size - its pixel size over its scale: a wider window lays the page out differently, and resizing the image afterwards hides it. Nobody reopens the app to judge it: a screens sub-agent judges the file (verify-screens.md), and a picture whose text reads smaller than the capture's was resized and is retaken. It matches the capture except for what ## Differences from the source lists. A state no capture draws, and an existing screen the feature does not change, are proven by a test or a walk proof, with no picture and no judge.
10. A capture, frame or live site the plan cites as binding is changed only by the user. Removing, renaming or departing from something it draws is asked, recorded as a decision in .specs/STATE.md, and listed in the plan's ## Differences from the source with Confirmed: y - AD-NNN.
11. The files hold the run; the chat does not survive a compaction. What the user settles while you build goes on checks.md's **Settled mid-build:** line in the turn they say it. Each sub-agent goes on the **Dispatched:** line before it is sent. After a compaction or in a new session, re-read this file and the current phase's reference, then resume from the files (memory.md).

Configuration

The project declares it in AGENTS.md or equivalent; checks.md copies the count onto its Mutations: line.

## tlc-implement

mutations: 3

Part	Runs	Cannot catch
check_proofs.py (always)	every proof once per runner at HEAD, each check matched to the tests its filter selects, the Build: commands, pictures and walk checked, mutations injected	a test that passes for the wrong reason
The Verifier (always)	N mutations wherever a rule lives, two examples nobody named, assertions looser than their criterion, checks that contradict a binding capture (verify.md)	the rules no mutation probed
The screens sub-agent (a criterion cites a capture)	an inventory of each capture against the author's picture, every row same or differs, at a named HEAD (verify-screens.md)	spacing, colour and type weight; detail smaller than a screenshot shows; screens no capture draws

Screens take no mutation; spend N where a wrong implementation costs most. Where N looks too low for the feature, say so in one line and let the user raise it.

Artifacts

.specs/
├── STATE.md                    # decisions log (AD-NNN)
├── LESSONS.md                  # rendered by scripts/lessons.py - do not hand-edit
├── lessons.json                # machine-owned
└── features/<feature>/
    ├── plan.md                 # from tlc-plan; read, and kept true
    ├── checks.md               # claims + proofs, the coverage join, swept, handoff
    ├── captures/               # binding source files, and the author's <capture>.running.png
    ├── screens.md              # the screens sub-agent's inventory and verdict, per picture
    ├── proofs.md               # written by check_proofs.py: every proof and mutation at HEAD
    ├── mutations.json          # the Verifier's mutations, re-run after the fix
    └── verification.md         # the Verifier's findings, each closed by a commit

Phases

Phase	What it produces	Reference
Checks	claims with proofs, every set joined member by member, the nine dimensions swept	checks.md
Build	tests from the checks, the code, commits, the screens walked and saved	build.md
Verify	rule 5	verify.md, verify-screens.md

Decisions, resuming and lessons: memory.md.

Scripts

Resolve <skill-dir> as the directory containing this SKILL.md and run python3 <skill-dir>/scripts/<name>.py. Pass --root when cwd is not the project root. A non-zero exit means stop and fix. Run a script rather than read it: its messages say what to change.

When	Command
Before starting to build	validate_checks.py <feature> - it also requires the plan's **Approved**: line
Before each commit	check_commit.py --message "<msg>"
Before dispatching the Verifier, and after the fix	check_proofs.py <feature> [--mutations <mutations.json>]
Before declaring done	validate_verification.py <feature>
Before writing checks, and after the gate	lessons.py ...
After editing a validator or template	selftest.py

selftest.py mutates the fixtures in scripts/fixtures/ once per gate rule and requires every mutant to be killed; the fixtures are also a worked example of a passing artifact set.

Requires

Needs	For	Without it
Python 3.8+ (standard library only)	every script	the gates do not run; the phases still apply
git	mutate.py and check_proofs.py (HEAD, clean tree)	no mutations and no HEAD check: say so in the report
sub-agents	the Verifier and the screens sub-agent	run their file in a new session and write self-verified; the gate warns
a model that reads images	the screens sub-agent	hand the pictures and the captures to the user
a browser the agent can drive	saving <capture>.running.png and the walk	hand the steps to the user

Test runners: check_proofs.py groups bun, vitest, jest, pytest and go test into one run each (scripts/runners.py); any other command runs alone and is judged by its exit code.

Output

Produce the artifact; do not narrate the phase. Lead with the verdict and state decisions plainly. Section headings stay in English; prose follows the project; identifiers are never translated.