craft:prune — standing harness-prune review

Standalone, session-owned skill, invoked on demand. It is not a pipeline phase — it is not part of the default pipeline and never runs unless a session explicitly invokes it. An uninvoked skill is inert: zero drag, nothing to revert, nothing to maintain against a run that never happens.

This skill is ADVISORY and read-only: it writes no harness file. Its only output is a proposal a human reviews. A run that finds nothing to propose is a recorded no-op, never a blocker.

Input: $ARGUMENTS (zero-argument).

Preamble — read the denylist first, fail closed

Before inspecting anything else, read contracts/core.md. It is the undroppable-core denylist every later candidate is checked against.

• Fail-closed. If contracts/core.md cannot be read, stop here and emit no proposals. Never fall back to an empty denylist and propose against it — an unreadable denylist is a hard stop, not a green light.
• The same refusal covers the non-overridable cross-phase invariants documented alongside the core contract: a candidate mapping to either is refused before it is ever emitted, not merely flagged after the fact.

Inspection scope

Read-only, over exactly these paths:

• contracts/*.md — the invariant floors, read both as the denylist source and as inspection material (a floor can also restate itself redundantly across files).
• The lint set — scripts/*-lint.sh, engine/bin/*-lint.js, and the engine/src modules those bins import.
• Skill and agent prose — skills/**, agents/**.

Resolve the current model identity from the run context; when the run context does not carry it, read docs/guides/model-class-matrix.md. Flag drag the resolved model no longer needs: belt-and-braces guidance that duplicates behaviour the model already carries natively, a lint superseded by a newer mechanism, or prose restating what the model now does without being told.

Rule-vs-fact. A unit that encodes a decision procedure the model can already run, where stating one missing fact would suffice. The candidate's replacement is the fact itself, stated once, at the place the procedure lived.

These two questions are orthogonal: the first asks whether the guidance is needed at all; this one asks whether guidance is the right shape for what it carries. A unit can fail the second and pass the first.

Output — a proposal, never an action

The only output is a structured list. Each candidate carries three fields: unit (the file or section proposed for removal), rationale (why the current model no longer needs it), and what-would-replace-the-safety-it-provided (what covers the gap if it is dropped — "the model's native behaviour," "a newer lint," or "nothing — flag as a genuine regression risk, not a candidate," when nothing would).

Emit each surviving candidate as one greppable line:

PRUNE-CANDIDATE(<unit>): <rationale>

For a rule-vs-fact candidate, the rationale carries a fixed prefix so the two classes stay greppable apart — no new token, the same PRUNE-CANDIDATE(<unit>): line, a distinguishing rationale prefix only:

PRUNE-CANDIDATE(<unit>): rule-vs-fact — <rationale>

The three-field candidate shape is unchanged; for a rule-vs-fact candidate the existing third field (what-would-replace-the-safety-it-provided) carries the missing fact.

This token is defined here only. It does not join the skills/run/SKILL.md token family, and no other skill or script emits it.

Procedure

1. Read contracts/core.md. Unreadable → report the read failure and stop; emit no proposals (fail-closed — see Preamble).
2. Resolve the current model identity (run context, else docs/guides/model-class-matrix.md).
3. Walk the inspection scope. For each unit that looks like drag against the resolved model, or that encodes a decision procedure a stated fact would replace, draft a candidate with the three fields above.
4. Check every drafted candidate against the denylist read in step 1, identically for both classes: a rule-vs-fact candidate mapping to a core invariant is dropped before emission, exactly as a drag candidate is. A candidate mapping to a core invariant, or to a non-overridable cross-phase invariant, is refused — drop it; it is never emitted.
5. Emit the surviving candidates as PRUNE-CANDIDATE(<unit>): <rationale> lines, and as PRUNE-CANDIDATE(<unit>): rule-vs-fact — <rationale> lines for the rule-vs-fact class.
6. Report the list to the user for ratification. Delete or edit nothing.

Enacting an approved prune

This skill proposes only. A candidate the user approves is enacted through a normal craft feature run (design through validation), where the harness's existing gates — contracts-lint, source-hygiene, the DoD lints, and the repo's validation techniques already wired into ci.sh — protect the core exactly as they would for any other change. This skill itself never deletes or edits a harness file.

Before it is enacted, the maintainer should check an approved candidate by hand against the behavioural eval case (or cases) that drive its unit (evals/, procedure in docs/contributing/maintainer-smokes.md). craft never runs claude plugin eval itself. A with-craft score that holds once the unit is removed, or a delta near zero, is evidence the model no longer needs it. A unit no case drives carries no such evidence; say so in the proposal. This is advice, not a gate.

Trigger

Manual and documented here — the ratified floor. There is no auto-detection of a model change, no cadence timer, and no new pipeline phase. A session runs this skill by name when a human decides a review is warranted — for example, after a run deliberately records a new model class in docs/guides/model-class-matrix.md.

Error semantics