Skill Sync

$1 $ARGUMENTS

Argument Contract

Pipeline

flowchart TD
    S1["Stage 1 — Token Profile<br>uvx skilllint@latest check SKILL.md<br>wc -w on references/ for size map"] --> S1Q
    S1Q{"Existing file<br>already over SK007?"}
    S1Q -->|"Yes"| S1R["Apply /refactor-skill<br>progressive-disclosure extraction<br>Restart pipeline after"]
    S1Q -->|"No"| S2
    S2["Stage 2 — Parallel Read Agents<br>3 Agent() calls in ONE turn"] --> S3
    S3["Stage 3 — Synthesis<br>Fetch docs index, validate NEW URLs<br>Write change plan to .tmp/scratch/plans/"] --> S3Q
    S3Q{"Change plan has<br>changes?"}
    S3Q -->|"No — all VERIFIED"| Done1(["No changes needed — exit"])
    S3Q -->|"Yes"| S4
    S4["Stage 4 — Clean-Tree Gate<br>git status --porcelain"] --> S4Q
    S4Q{"Output empty?"}
    S4Q -->|"Not empty"| S4Stop(["STOP — dirty working tree<br>List uncommitted files"])
    S4Q -->|"Empty"| S45
    S45["Stage 4.5 — Orchestrator Pre-Fetch<br>Fetch all SOURCE URLs from change plan<br>Write verified content to source-material file<br>URLs that 404 → UNVERIFIABLE, skip"] --> S5
    S5["Stage 5 — Write Agent<br>Agent(skill-content-updater)<br>Receives change plan + source-material path"] --> S5Q
    S5Q{"Post-write skilllint<br>exit code?"}
    S5Q -->|"0"| S6
    S5Q -->|"non-zero"| S5R["Extract over-budget sections<br>to references/<br>Update change plan"]
    S5R --> S5
    S6["Stage 6 — Completion Gate<br>skilllint on ALL modified files"] --> S6Q
    S6Q{"All pass?"}
    S6Q -->|"Yes"| Done2(["Pipeline complete"])
    S6Q -->|"No"| S5

Stage Definitions

Stage 1 — Token Profile

Run uvx skilllint@latest check <skill-path>/SKILL.md (target the file, not the directory — passing a directory silently reports "Total files: 0" and exits 0 without validating anything). Note whether it passes or fails SK007.

To identify the largest reference files before adding content, run wc -w <skill-path>/references/*.md | sort -rn | head -5 and surface the top 3 by word count. This indicates where budget pressure is highest.

Pre-write SK007 branch: if the existing SKILL.md is already over SK007, apply /plugin-creator:refactor-skill (progressive-disclosure extraction) before making content changes. Restart the pipeline after the refactor completes.

Stage 2 — Parallel Read Agents

Dispatch exactly 3 Agent() calls in ONE turn (not sequential, not TeamCreate):

1. Agent(subagent_type="plugin-creator:skill-auditor") — input: <skill-path>; output: .tmp/scratch/reports/skill-sync-{slug}-completeness-YYYYMMDD.md (read-only)
2. plugin-creator:skill-content-updater (read role) — upstream drift scan; output: drift report with NEW/STALE/VERIFIED/UNVERIFIABLE verdicts per claim
3. general-purpose — structure validation; checks progressive disclosure, frontmatter schema, broken reference links; output: structure report

Each agent writes its report to .tmp/scratch/reports/. Report formats: ./references/report-formats.md

Note on drift scanner URL fetching: The skill-content-updater agent fetches the SOURCE: citation URLs already present in the skill, as defined in ./references/url-fetch-spec.md. Do not instruct it to fetch an external docs index — that is the skill-sync-source-validator agent's responsibility in Stage 4.5.

Stage 3 — Synthesis

Orchestrator reads all three reports and writes the change plan to .tmp/scratch/plans/skill-sync-{slug}-YYYYMMDD.md.

If all verdicts are VERIFIED and no structural issues found: write a "no changes needed" change plan and skip Stages 4–6.

Change plan format and synthesis precedence rules: ./references/change-plan-format.md

Stage 4 — Clean-Tree Gate

git status --porcelain

If output is not empty: STOP. List the uncommitted files. Report to the user. Do not proceed until the tree is clean.

Stage 4.5 — Source Validation Agent

Dispatch Agent(subagent_type="plugin-creator:skill-sync-source-validator"). Pass the change plan path.

The agent validates and pre-fetches all source content before the write agent runs, keeping URL fetching out of the write agent's context and providing a structured fallback chain when primary fetch methods fail. It handles:

• Validating all NEW URLs against the docs index (llms.txt / sitemap) — downgrades fabricated URLs to UNVERIFIABLE directly in the change plan file
• Pre-fetching content for every SOURCE URL using the priority fetch chain defined in ./references/url-fetch-spec.md
• Writing .tmp/scratch/fetched/source-material-YYYYMMDD.md with verified content per change plan entry
• Updating the change plan header with the source-material file path

Stage 5 — Schema-Aware Write Agent

Dispatch one Agent(subagent_type="plugin-creator:skill-content-updater") in write role. Pass the change plan path. The change plan header now contains the source-material file path — the write agent reads from that file and does not need to fetch any URLs.

After the agent returns, run uvx skilllint@latest check <modified-skill-path>/SKILL.md. If non-zero: extract the over-budget sections to references/, update the change plan with the extraction directive, and re-dispatch the write agent.

Stage 6 — Completion Gate

Run uvx skilllint@latest check on every file modified in Stage 5. All must exit 0. If any fail: return to Stage 5 with a targeted remediation change plan.