What doesn't work
Type: review · Status: current
Auto-commits
Hook-driven automatic commits after every note operation created a mess. Commits were noisy, hard to review, and we spent significant effort removing them. Agents should not commit without explicit human approval.
Together with what works, this is the manual observation log that a KB learning loop would eventually feed from — these anti-patterns are ground truth for what mutations the loop should avoid proposing.
Observations needing more evidence
The following areas showed friction, but we haven't tested them enough to draw conclusions.
- 26 skills/commands — 16 local + 10 plugin. Most rarely used, but unclear which are valuable until we use the system more.
- Queue and pipeline machinery — adds significant complexity. Noticeable overhead for a single-contributor project, but may pay off differently at scale.
- Schema validation as a separate ceremony — a dedicated FAIL/WARN/PASS phase adds machinery. The frontmatter fields themselves are useful; the question is whether formal validation justifies its cost.
- Session rhythm protocol — orient → work → persist adds ceremony. Unclear whether it changes behavior beyond what good context already provides.
- Connection requirements outpace connection-making — orphan rate reached ~90%. The gap between connection rules and actual connections was noticeable, but the rules themselves may not be the problem.
Topics: