Skill Enhance

Skill Enhance

Core responsibility: Take one existing Lattice skill from current state to modern-capable-model grade through a fixed pipeline: baseline → dependency map (background agent) → precision rewrite with guaranteed behavior preservation → combined three-skill QA → independent verification. The skill being enhanced must end up smaller-or-equal in redundancy, fully grammatical, convention-complete, and provably non-regressive. Organizing objective: deterministic execution — every sentence has exactly one reading, every reachable state exactly one action, every cross-file reference literally true. Acceptance test: two different capable models reading any sentence would act identically on it.

Input: One skill path or name (any tier). Molecules are the primary case.

Output: Edited files + close-out report.

STOP: NEVER commit, push, stash, checkout, reset, or run build-codex-plugin.sh. Working tree only — surface install/mirror commands for the user instead.

How to verify this skill did its job:

• Pre-capture git diff exists; every preserved behavior in it is confirmed present after rewrite
• QA trio re-run returns no critical gaps, CLEAN or tightened-only results, and PASS
• Independent verifier reports zero real defects
• Scenario walk leaves no branch with two possible readings

Step 1: Scope and baseline

1. Read PROJECT.md (single source of truth for conventions) and the target SKILL.md in full.
2. Classify: tier (atom / molecule / refiner); if molecule, its type per PROJECT.md — generative (code-forge, refactor-safely, bug-fix) or planning/interactive (design-blueprint, architecture-compass). This governs every later rule: never transplant confirmation gates into generative molecules, never strip them from planning ones.
3. STOP: capture the full git diff of target + likely-touched files BEFORE any edit, written to a temp file whose path you will hand to the Step 5 verifier. This baseline is the preservation oracle for Steps 3-5. Without it, "did we drop anything?" is unanswerable.

Step 2: Dependency map (background agent)

Spawn ONE background Explore agent — do not read serially in the main thread. It maps:

1. Every framework:{atom} reference: real behavior/mode names each atom actually exposes, plus line counts.
2. Sibling molecules feeding and consuming this one: exact handoff contracts — frontmatter fields, status values, .lattice/{subfolder} paths, section names each side greps for.
3. Where the target's phases/levels are actually defined (own body vs referenced atom).
4. Inventory of telegraphic damage: sentences missing articles/verbs ("ground decisions real project").

Step 3: Rewrite (main thread, after map returns)

Apply in order:

1. Grammar restoration — every sentence complete. Telegraphic phrasing is ambiguity risk precisely at STOP/gate semantics.
2. Deduplicate — same rule twice in different words → keep the sharpest once. Cut trailing rationale ("this ensures…", "without this…") EXCEPT consequences load-bearing at the exact action point — flag every deliberate keep in the report.
3. Precision pass — hard gates get **STOP:** prefixes; targeted questions replace generic approval language; every conditional branch (missing doc, unreadable path, external reference, absent config, outputs created by older skill versions missing newer fields) resolves to exactly ONE action; expected-not-error paths marked as such.
4. Scenario walk your own rewrite: fresh start / interrupted session / partial output / minimal input / maximal input / entry-and-resume edges (including partial entry points and legacy docs without new markers). Derive the state space from upstream contracts first — enumerate every status value, mode name, and doc-presence combination the upstream defines, and give each a defined outcome here. Close every gap where two readings survive.
5. Preservation check against the Step 1 baseline — nothing dropped, renamed, or weakened.
6. Contract alignment — verify every referenced behavior name, section heading, table schema, config key, and status value against the live atom files; fix mismatches even when pre-existing. Composition must interlock literally, not approximately.
7. Convention restoration — always/conditional qualifiers on Required Skills, collaborative-judgment fallback wording in Ambiguity Signals, trigger phrases in description, description consistent with body.
8. If the skill repurposes a code-generation atom in another mode: give that atom a compact named mode section. Never improvise inline; never inline atom content into a molecule.

Step 4: Combined QA gate (strict order)

Run on every touched file:

1. skill-review — fix criticals and warnings; list observations for the user decision.
2. skill-tighten — full T1-T6 sweep, apply.
3. skill-validate — fix error-level findings directly; report warnings without applying.

Iterate until all three return clean/pass. Re-run any skill whose inputs a later fix changed.

Step 5: Independent verification (background agent)

Spawn ONE fresh general-purpose verifier that re-derives correctness WITHOUT trusting the diff author:

• Semantic-preservation checklist from the Step 1 baseline
• Internal-consistency scenario walk (Step 3.4 scenarios, independently)
• Contract checks against the real consumer files — grep them, cite lines
• Hedge/degraded-grammar grep across touched files
• Frontmatter YAML validity, name/folder match

Fix everything it finds. Then: behavior-affecting fixes → re-run the affected Step 4 tools; single-line narrowing fixes → self-check against the loaded T1-T6/validator patterns instead of a full re-run.

Step 6: Close-out report

Report: files changed with before→after line counts; findings per QA tool and how each closed; deliberate keeps flagged; residual validator warnings listed untouched; and these commands surfaced for the user (not run):

./tools/install.sh <your-tool>/skills/   # smoke-test the skill loads
./tools/build-codex-plugin.sh            # refresh Codex mirror before committing