style-evolve

name	style-evolve
type	user-invocable, agent-invocable
description	Single human-facing command for all Citum style work. Use whenever someone asks to fix, improve, convert, create, or tune a citation style. Routes to upgrade, migrate, create, or tune. Always use this rather than calling internal skills directly.
model	sonnet
routes-to	style-maintain, style-migrate-enhance, style-tune, style-qa

Style Evolve

Human UX (Public Entry Point)

/style-evolve upgrade <style-path>
/style-evolve migrate <csl-path>
/style-evolve create
/style-evolve tune <embedded-style-id>

Do not ask users to call internal skills directly.

Setup (read first)

Before doing anything else, read:

• docs/policies/STYLE_WORKFLOW_DECISION_RULES.md — failure classification and stop conditions
• docs/guides/STYLE_WORKFLOW_EXECUTION.md — decision flow, evidence ladder, shared gates
• docs/guides/JJ_AI_CHANGE_STACK.md when .jj is present and local change-stack curation would help isolate a style pass

Autonomous Operation

Run the full pipeline without stopping to ask questions between steps. Only interrupt for the explicit permission gates in CLAUDE.md.

Modes

1. upgrade

Route to ../style-maintain/SKILL.md.

2. migrate

Route to ../style-migrate-enhance/SKILL.md.

3. create

Build a new Citum style from source evidence. Escalate to @dplanner for design.

4. tune

Route to ../style-tune/SKILL.md. Use when the target is an embedded-core style and the goal is 100% fidelity and clean SQI. Both are hard gates for this mode. The migrate output is the seed, not the deliverable.

Co-Evolution Rule

Use docs/policies/STYLE_WORKFLOW_DECISION_RULES.md for failure classification and stop conditions. Keep the router focused on dispatch and host entrypoint behavior.

Shared Gates

• Compatibility fidelity regression is never allowed unless the task explicitly chooses a documented semantic divergence from legacy CSL behavior.
• SQI is a hard gate for embedded-core styles (fidelity AND SQI both required). For dependent styles, SQI is optimization-only after fidelity is stable. See docs/policies/STYLE_WORKFLOW_DECISION_RULES.md for the tier definition and quality bar.
• Before editing a style, classify it by semantic class, implementation form, and portfolio tier using the shared workflow docs and docs/specs/STYLE_TAXONOMY.md.
• Profile-family work may require a create pass for a hidden family root followed by upgrade reduction of the public handles.
• Journal/profile reductions must choose parents from guide-backed authority, not nearest CSL or template similarity.
• Keep waves bounded to one family or one clearly related cohort per PR.
• For styles with configured benchmark_runs, run node scripts/report-core.js --style <name> after the primary oracle pass and treat the rich-input results as official supplemental evidence.
• All modes must pass ../style-qa/SKILL.md before completion.
• If docs or beans are changed: ./scripts/check-docs-beans-hygiene.sh must pass.

Output Contract

Every completed task delivers:

• fidelity metrics
• SQI delta
• authority basis
• rich input evidence summary
• divergences
• code opportunities table
• QA verdict
• research value
• commit SHA and message, or none

Include a routed sub-skill header line in the final output.