| name | sync-docs |
| model | sonnet |
| disable-model-invocation | true |
| description | On-demand currency check: diffs meta-dev's standards and lifecycles against current official Claude Code docs and routes drift through the update lifecycles. |
sync-docs
meta-dev's anti-rot mechanism: the standards encode claims about Claude Code that the product keeps changing underneath. Nothing else in this plugin re-checks those claims against reality, so without this skill they silently age into fiction. Runs only on explicit /meta-dev:sync-docs invocation — never automatically, since a docs check mid-unrelated-task would be a surprise, not a service.
Step 1 — Verify each module's claims against current docs
- Read every standards module in
${CLAUDE_SKILL_DIR}/../../standards/ (core.md plus each per-artifact module) and extract its factual claims about Claude Code behavior — the things a docs change could falsify (event names, frontmatter fields, loading mechanics, CLI flags, file locations).
- Group claims by domain: skills, subagents, hooks, plugins, memory/instruction files.
- Dispatch the
claude-code-docs agent in parallel, one dispatch per domain (a user-global agent at ~/.claude/agents/ with a docs MCP binding — deliberately not shipped inside meta-dev, since plugin agents cannot carry mcpServers; if it is missing, fall back to fetching code.claude.com/docs pages directly and say so in the report), each carrying that domain's extracted claims to verify. Require each claim come back tagged confirmed / changed / new-capability / deprecated, with the doc URL as evidence — an unsourced verdict isn't a verdict, it's a guess.
Completion criterion: every module's claims have a verdict and a source URL from a domain dispatch.
Step 2 — Build the drift report
- For each module, list what changed and which lifecycle skills (create/update/improve/restructure/retire-skill and their agent/hook/plugin counterparts) depend on the now-stale claim.
- Read the register at
${CLAUDE_PLUGIN_DATA}/declined.md (persistent across plugin updates); if it doesn't exist yet, seed it by copying the shipped register ${CLAUDE_SKILL_DIR}/../../docs/declined.md. ~/.claude/skill-workspaces/ is scratch space, not a register; nothing durable is tracked there.
- Drop any drift item that matches an already-declined entry, unless the doc evidence changed since the decline (re-surface those as new, citing the earlier decline).
Completion criterion: drift report lists only items not already declined, or explicitly re-surfaced with new evidence.
Step 3 — Present and route
- Present the drift report to the user before changing anything — this is a scheduled currency check, not a hands-free improve; the user decides what's worth acting on now.
- For each item the user accepts, route by target:
- Drift in a lifecycle skill's behavior or instructions → dispatch
meta-dev:update-skill on that skill (it owns the diff-confirmation and edit; don't duplicate that consultation here).
- Drift in a standards module itself → edit
standards/<module>.md directly, with the user's approval on the specific wording (standards have no separate lifecycle skill of their own).
- For each item the user declines, capture the reason in the same turn — it is the only chance before the report scrolls out of context.
Completion criterion: every accepted item has been routed to its update path or edited; every declined item has a captured reason.
Step 4 — Record declines
Append each newly-declined item to ${CLAUDE_PLUGIN_DATA}/declined.md as one entry: what was found, the doc URL, the reason declined, today's date. Skip items already present from Step 2.
Completion criterion: the register contains every decline from this run and no duplicates.
Step 5 — Finish and report
- If any routed update touched a skill workspace, run the mandatory cleanup per
${CLAUDE_SKILL_DIR}/../skill-creation/references/eval-pipeline.md §9: archive snapshot + log to ~/.claude/skill-archive/, then rm -rf the workspace.
- Report three lists to the user: applied (with what changed), declined (with reasons), deferred (accepted but not yet actioned, with why).
Completion criterion: workspace cleanup done for every touched skill; the applied/declined/deferred report delivered.
