liquidware-ninjaSecond brain & toolkit
Personal project. Research and learning only. Not for production. No warranty.Not for production
  • Built on my own time for research and learning. Not affiliated with, endorsed by or supported by my employer, Liquidware or Microsoft.
  • Bring your own Liquidware licenses. Don't run any of it in production or against tenants and subscriptions that matter. Test in a dedicated lab first.
  • Provided "as is", without warranty of any kind, under each repository's license. You use it at your own risk and own the Azure charges it incurs.
  • Contributions and forks are welcome and encouraged on public repos. See Contribute.

0014. Pilot the family's enforced knowledge capture, in warn mode

Context and Problem Statement

The family accepted a rule that knowledge capture is enforced, not just asked for: a thread closeout on the Claude side, a Claude Code Stop hook in each repo, and a kb-capture PR check with a no-kb opt-out. This KB is one of the pilot repos, next to the AVD repos and the Nerdio HCL kit. How does this repo run the pilot without slowing work that has a fixed date?

Considered Options

  • Join the pilot in warn mode: the hook only reminds, the PR check runs for real
  • Join the pilot in block mode
  • Stay out of the pilot

Decision Outcome

Chosen option: "Join the pilot in warn mode", because the family runs every KB that already has CI in warn mode first, and the PR check still makes "nothing to capture" an explicit choice.

Also recorded in: 2ndbrain, the source of truth for how the closeout, the hook and the check work.

For this repo:

  • Installed from the family's template export, not by hand: scripts/kb/capture_check.py, the kb-capture workflow and the Stop hook in .claude/settings.json. The family scripts stay identical to the export's; only kb.toml differs.
  • Warn mode: capture_hook = "warn" in kb.toml. The Stop hook posts a reminder and never blocks.
  • The PR check runs for real. A PR that changes code without changing docs/knowledge/ needs the no-kb label and a No KB entry: <reason> line in its body. The family's own automation PRs are exempt.
  • Fixed-date work doesn't wait. A demo PR that goes red takes no-kb and a reason instead of waiting on a KB entry.
  • The project's private repo is exempt from the hook and the check for the pilot. Its threads still run the closeout, and a lesson from it reaches this KB only after a guardrail check. It is named only in the project's own instructions, never here.
  • Thread closeout. Every thread in this repo's Claude project runs the 2ndbrain closeout before its final reply, and the project's coordinator doesn't resolve a thread until the closeout is recorded or the owner waives it. Private-layer writes still wait for the owner's yes.
  • Pilot dates: through 2026-10-31, counted from when the hook and the check go live, with a review on 2026-11-01. At each checkpoint the project adds one field report to the AI KB: PRs merged, how many carried a KB change, how many used no-kb, and how often the hook warned.

Consequences

  • Good, because lessons from code changes are asked for every time, by the harness and by CI.
  • Good, because a no-kb reason makes a skipped entry visible and reviewable.
  • Bad, because the hook runs only in Claude Code sessions, and a person can still merge past a red check.
  • Neutral, because warn mode can move to block after the review, by changing one setting.

More Information

Asked for in a handoff issue from the AI project (avd-liquidware#12), approved by the owner on 2026-10-19. The review on 2026-11-01 decides whether this repo stays in warn mode or moves to block.

Linked from 2 pages