Drift

/up-docs:drift [wiki-subtree-or-tag]

Run drift analysis via the up-docs-audit-drift sub-agent (Sonnet). Read-only by design: the auditor surfaces findings; the user decides whether to re-invoke propagators to fix them.

If a wiki subtree or tag is provided, scope the analysis to that llm-wiki wiki/ subtree. Otherwise, analyze the whole wiki/.

Architecture

The auditor sub-agent runs the full four-phase drift flow in its own isolated context:

1. Infrastructure → Wiki sync (SSH/pct/curl against live state)
2. Wiki internal consistency (cross-reference map)
3. Link integrity & enrichment
4. Notion-relevance review

The phase mechanics (scripts/convergence-tracker.sh, scripts/server-inspect.sh, scripts/link-audit.sh) still live in this plugin — the sub-agent invokes them directly from its Bash tool.

Workflow

1. Gather Session Context

First, verify Python 3 is available — all helper scripts depend on it:

command -v python3 >/dev/null 2>&1 || { echo "ERROR: python3 not found in PATH — install python3 and retry."; exit 1; }
bash ${CLAUDE_PLUGIN_ROOT}/scripts/context-gather.sh
bash ${CLAUDE_PLUGIN_ROOT}/scripts/convergence-tracker.sh init

Combine with conversation history.

2. Build the Session-Change Summary

Read ${CLAUDE_PLUGIN_ROOT}/templates/session-change-summary.md for the canonical format. This grounds the auditor's scan — it starts from the summary items and expands to adjacent infrastructure that might be transitively affected.

3. Dispatch up-docs-audit-drift

Invoke via the Agent tool with subagent_type: "up-docs:up-docs-audit-drift" (the up-docs: prefix is required — plugin-defined agents are only addressable through their plugin namespace). The prompt:

• Session-change summary at the stable front
• Wiki subtree/tag scope argument (if provided) at the end
• The reference docs skills/drift/references/convergence-tracking.md and skills/drift/references/server-inspection.md are read by the sub-agent itself; do not duplicate their content into the prompt.

4. Pass Findings Through

The sub-agent returns both a JSON findings block (canonical artifact) and a markdown findings table (for human reading). Emit both in the skill's final output.

If the sub-agent includes an ⚠ ESCALATION RECOMMENDED block, include it verbatim.

5. Offer Next Step (bounded choice)

After findings land, use AskUserQuestion to offer:

• Re-invoke propagators with findings as a new session-change summary (fixes them at propagator cost — all propagators run on Sonnet)
• Re-run the audit with Opus (if escalation was recommended)
• Accept findings as advisory and exit

Do not auto-invoke any of the above.

Notes

• This skill no longer runs SSH/pct/curl directly — the sub-agent does.
• Convergence + oscillation detection live in scripts/convergence-tracker.sh. The default state-file path is ${TMPDIR:-/tmp}/up-docs-tracker-${CLAUDE_CODE_SESSION_ID:-default}.json so that the 6+ separate invocations in one drift session share state. Override with UP_DOCS_TRACKER_STATE for tests or for non-session usage.
• Findings are advisory: the auditor has no write tools for llm-wiki or Notion. Fixes go through the propagators on a follow-up pass with the user's explicit consent.