ulw-correctlisted

# Correct a ULW Misclassification Tell the harness "the last turn was misclassified" with a short reason. The classifier's `detect_classifier_misfire` is passive — it only records misfires on the next turn when it sees blocks or bare-affirmation patterns. This skill is the active counterpart: a one-button "this is wrong, here's the correction" verb so you don't have to rephrase three ways or learn the classifier's mental model. ## Usage ```text /ulw-correct this is advisory not execution /ulw-correct intent=advisory domain=writing — was asking what to write, not to write it /ulw-correct domain=mixed — the docs change has a code change attached ``` ## What it does 1. Records a `corrected_by_user=true` row in `classifier_telemetry.jsonl` for the current session and in `~/.claude/quality-pack/classifier_misfires.jsonl` cross-session. The row carries: prior intent/domain (what the classifier inferred), corrected intent/domain (if parseable from the reason), reason text, and the prompt that was misclassified. 2. If the correction names an `intent=X` or `domain=Y` (with X/Y matching a valid value), updates the active session's `task_intent` and/or `task_domain` state so subsequent gates and directives use the corrected routing. 3. Acknowledges with a one-line summary: what changed, what didn't. ## Steps 1. Take the user's correction reason (everything after `/ulw-correct`). 2. Invoke the recording script: ```bash bash ~/.claude/skills/autowork/scripts/ulw-correct-reco