/rota-learn
Save the session’s durable lessons to KNOWLEDGE.md.
When to use
Use at the end of a session that surfaced reusable knowledge, after a correction-rich debugging arc, or on “save what we learned”, “capture this learning”, “/rota-learn”.
Reach for another skill when
- the point is a commitment with explicit forbids and permits, not a gotcha:
/rota-decide - you only need to hand off where you stopped, not record lasting knowledge:
/rota-pause
Proposed bullet under ## Alerts: Severity is case-sensitive in the dedup key path. Normalize at the boundary, not at storage time. Written to .rota/KNOWLEDGE.md.
- one new bullet under its topic in .rota/KNOWLEDGE.md, date-stamped
- the rota-knowledge topic index in AGENTS.md refreshed
Flags and modes
--strict- run the Opus verifier over the new bullets
--retro [<session-id|path>]- turn the session’s mistakes into guardrails and standards
--external-review- propose splits for oversized external topics and fold stray files, asking before each split
--term <name>- write a Glossary entry
--promote | --deprecate | --amend <topic> "<title>"- edit one existing bullet’s tier or body without the capture flow
Details
Writes durable knowledge from the current session into .rota/KNOWLEDGE.md, grouped by topic. It captures gotchas, project conventions, constraints, debugging insights and decisions with rationale, and skips anything obvious from reading the code. --strict adds a verifier pass over the new bullets (off by default). --retro [<session-id|path>] reads the newest session transcript (Claude Code or Codex), or the one you name, and cites it as evidence, then classifies the session’s mistakes into guardrails (filed as items), standards, navigation pointers, tool-economy fixes and information-access gaps (filed as items), flags no-op bullets for removal without deleting them, audits steering files (AGENTS.md, CLAUDE.md, global instructions) for restating or dead lines without editing them, and reports a repo whose check command runs in no CI workflow or hook as a no-guardrail finding. In umbrella mode the write (and --term Glossary entries) routes to the scope that the cwd or --repo resolves to: repo-local or the umbrella-shared .rota/KNOWLEDGE.md. Three manual flags edit one existing bullet and skip the capture flow: --promote <topic> "<title>" sets its tier to confirmed without waiting for the hit threshold, --deprecate <topic> "<title>" sets it to deprecated, and --amend <topic> "<title>" appends text to its body (tier and hits stay). --external-review is the maintenance pass over the external knowledge base: it measures the topics, drafts a split plan (3 to 7 subtopics with hooks) for each oversized one, offers to fold stray files an older rota left behind, and always asks you before it runs rota knowledge split, whatever autonomy.level says. See learning, external knowledge and umbrella mode for the full flow.