implement for Codex

Left alone, an agent will build for an hour and hand you one giant batch of changes to untangle. implement is the building discipline: thin slices that each work end to end, a test written before the code it proves, a working checkpoint after every slice. Things discovered along the way get filed as new work instead of chased on the spot. You see progress you can check, not a long silence.

Read the Codex desktop binding when this workflow needs harness mechanics, model routing, or recovery.

How a working lane builds an item. The item's body is the spec, the brief's verification set defines done, and this skill is the discipline between those two points. It is not a lane type of its own, and it deliberately has no binding slots: everything repo-specific (verify commands, constraints, the tracker) already arrived in the lane's brief.

Tracer first

The first verified checkpoint proves the thinnest end-to-end slice: the path that goes all the way through, one field wide, before anything widens. Widening follows, slice by slice, each responding to what the last one taught. This pairs with to-tickets' slicing: tickets sliced tracer-style are built tracer-style. An item with no thin end-to-end slice in it is a slicing problem: report it back through the orchestrator rather than building horizontal layers and calling the stack a tracer.

Seam-scoped test-first

A seam is a public boundary where behavior is observed without reaching inside, and the seams under test are the ones the item's verification set names, agreed upstream when the ticket was filed, not invented mid-lane. Two bars, by whether a seam is named:

At every named seam: red before code. The failing test is written first, watched failing, and stays in the suite after it goes green. One seam, one test, one minimal implementation per cycle.
Outside named seams, validate the actual risk. Use relevant existing checks for small reversible edits. Add behavioral tests when they establish new behavior or catch a reported regression; do not write tests that merely mirror implementation.

Proportional validation never waives required project tests or a regression check for a reported bug. Required tests accompany the delivered change and pass for its current revision, whether committed or left as a diff under a no-commit instruction.

Tests verify behavior through the seam, never internal structure: a test that breaks under refactor while behavior held was testing the wrong thing.

Green checkpoints

Save green checkpoints within the user's commit instructions. Commit coherent slices when commits are authorized. A no-commit instruction instead preserves the working diff and its verification in the checkpoint. An interrupted Codex turn does not erase uncommitted files; inspect the workspace before recovery.

File, don't fix

Building surfaces adjacent work: a bug next door, dead code, a stale doc, a refactor begging to happen. All of it goes to the tracker, a comment on the driving ticket or a suggested new ticket for disposition, and the lane's diff stays inside its brief. The lane never silently expands its own scope; the tracker entry is how the discovery survives without the brief growing.

What the close-out audit checks

Enforcement is artifacts, not forensics: the orchestrator's close-out audit checks three things the finished lane either shows or does not:

Tests exist and pass at every seam the verification set names.
The tracer slice is identifiable: the first verified checkpoint demonstrably proves a thin end-to-end path.
The diff contains no out-of-scope files.

No commit-by-commit reconstruction of red-before-green: the kept tests, the visible tracer, and the clean diff are the evidence.

Interlocks

Four situations route out of this skill, each to a named place:

A bug surfaces mid-build → diagnose. The named-cause test binds before the lane resumes building: no fix ships without a cause stated in one plain sentence, with evidence.
The structure fights the lane: a change that should have been local sprawls, a test only writable past an interface → fix nothing structural; report it as friction in the close-out summary, codebase-review's third entry gate.
The spec has a material unresolved decision → ask the decider through the coordinator and pause only dependent work. Resolve routine implementation choices from evidence and recorded decisions; do not re-ask settled questions.
A step only a human can take → generate a wizard and put the blocked label on the driving item until the human clears it.

Done when (checkable: verify each line before reporting complete)

The first verified checkpoint proves the thinnest end-to-end slice, and later checkpoints widen it; no step stacks a horizontal layer that goes nowhere alone.
Every seam the item's verification set names has a test that was written red before its code, passes now, and stays in the suite; other changes have validation proportional to their behavior and risk.
Coherent work is checkpointed with its verification; commits follow the user's authorization and explicit no-commit restrictions.
The diff contains only in-scope files, and every adjacent discovery is on the tracker rather than in the diff.
Every mid-build bug carries diagnose's closing artifacts; structural friction is reported in the close-out summary; every material unresolved decision was raised while routine choices used available evidence; every human-only step rides a wizard with the blocked label on the driving item, or none of these arose.
The close-out summary points at the three audit artifacts: the seam tests, the tracer checkpoint, and the diff's scope.

Attribution

Adapted from Matt Pocock's implement and tdd (MIT). The implement-session shape is his: a session that builds from the spec or tickets, tests as it goes, and ends in review. So is the red–green loop by vertical slice: red before green, one seam, one test, one minimal implementation per cycle, each test a tracer bullet responding to what the last cycle taught (his terms), testing only at pre-agreed seams rather than everywhere, and behavior-through-public-interfaces as what a good test is.

What this pack changes: the seam agreement moves upstream (the item's verification set names the seams at filing time, instead of a mid-session ask) and validation outside named seams follows actual risk while preserving required tests and regression checks; the tracer-first mandate on the first verified checkpoint; verified checkpoints support same-task recovery within the user's commit instructions; the file-don't-fix scope rule; enforcement by close-out artifacts rather than commit forensics; and the four named interlocks into diagnose, codebase-review, the decider, and wizard.