| name | kb-fix |
| description | Small bug-fix lane with bounded diagnosis, agent-owned verification, and escalation. Use when the user reports a bug, failing test, broken behavior, or asks for a narrow fix that should not require a full brainstorm or plan unless it grows. |
| argument-hint | [bug description, failing command, or issue path] |
KB Fix
Fix narrowly, verify yourself, and escalate when the bug is not actually small.
Preflight
- Read
todo.md, docs/context/PROJECT.md, and relevant handoff/subsystem docs.
- Check
docs/context/research/ and docs/solutions/ for related failures.
- If the request is older than 72 hours or comes from a parked handoff, run a stale-work refresh first.
Loop
- Orient - classify as regression, known bug, flaky behavior, environment issue, or missing requirement.
- Reproduce - build an agent-runnable signal: failing test, command, browser flow, curl, fixture, or logs.
- Micro-plan - before editing, write a compact plan with the reproduced signal, likely cause, intended file/behavior target, protected test/oracle files, and exact verification command/probe.
- Hypothesize - write 1-3 falsifiable causes before editing.
- Fix narrowly - edit the smallest responsible area.
- Verify - rerun the reproduction and relevant tests. For UI bugs, use browser verification when available.
- Record - update
todo.md, handoff files if needed, and learning docs when the fix teaches something reusable.
- Refresh memory if durable - run
kb-map refresh when the fix changes behavior, architecture, run/test commands, integrations, sharp edges, or "do not repeat" knowledge. Skip with a note when the fix is cosmetic or isolated.
Do not tell the user to test normal app behavior if the agent can test it.
The micro-plan is not a kb-plan manifest. Escalate to kb-plan only when the
bug becomes multi-slice, crosses several owning surfaces, or needs dependency
ordering. For a narrow fix, the micro-plan exists to freeze the failing signal
and verification target before the code changes.
Ceilings
Default ceiling:
- 3 reproduction strategies.
- 5 fix/verify iterations.
- 2 root-cause hypothesis resets after confirmed reproduction.
- 1 escalation before asking for human help.
Progress can extend the ceiling. Progress means the failure narrows, reproduction improves, evidence points to a subsystem, or one verification layer now passes.
No progress means the same failure persists after unrelated edits, hypotheses change without evidence, or the next attempt cannot say what it proves.
Escalation
Escalate to:
kb-research when external/framework behavior is unclear.kb-brainstorm when product behavior or scope is unclear.kb-plan when the fix becomes multi-slice.kb-map refresh when architecture docs changed.
When stuck, produce:
## Bug Escalation Report
### Symptom
### Expected Behavior
### Reproduction
### Attempts
| Attempt | Hypothesis | Change/Test | Result |
|---|---|---|---|
### Evidence
### Current Best Theory
### Why I Am Stuck
### Recommended Route
Park only the blocked slice or handoff; continue unrelated runnable work when possible.
