Skip to main content

​
PHASE-2 — automation from the VM (defined, to wire after phase 1 runs are validated)

Owner: the Claude Code agent on the Azure VM manifest-agent (the one launched via the agent script, repos under ~/repos).

​
Schedules

  • Every 2 hours — light watch: step 1 of the skill only. If no new merged PR on mnfst/manifest since lastAuditedPr, exit silently (near-zero cost). If there are, run the full loop.
  • 00:00 nightly — full audit regardless, as a safety net (catches anything the watch missed, plus docs drift not tied to a PR).

​
Memory: Supabase (same local instance already on the VM, new schema docs_harness)

Tables:
  • runs (id, started_at, pr_from, pr_to, findings_count, issue_url)
  • findings (id, run_id, category, doc_page, doc_quote, evidence, status: raised | accepted | rejected | fixed, issue_url, pr_url, updated_at)
  • learnings (id, created_at, lesson)
The flat files references/decisions.md and references/learnings.md are the phase-1 equivalents; migration = import then keep Supabase as the source of truth, with the flat files regenerated as read-only mirrors so the repo stays self-explanatory.

​
VM prerequisites (to install at wiring time — warn Seb before

touching the VM, he may be using it live)
  1. Clones of mnfst/docs and mnfst/manifest under ~/repos, pulled at each run.
  2. gh token: issues:write on mnfst/docs, read on mnfst/manifest.
  3. Supabase schema docs_harness created on the existing instance.
  4. Cron entries for the two schedules, launching the Claude Code agent with the /docs-sync skill.
  5. A ping to Seb (WhatsApp/Discord via Hermes) when an issue is filed.

​
Guards carried over from phase 1

No PRs ever. One digest issue max per run. Rejected findings never re-raised. Silence when there is nothing to say.