name	fix-from-feedback
description	Use when asked to fix the bug behind a bet or theme — hand off to the coder subagent and verify the result.

Fix from feedback (bet → coder handoff → verified fix)

The closed loop: evidence picks the bug, a handoff scopes it, the coder fixes it in workspace/, tests prove it, the ledger records it.

Procedure

Pick the bet. Call list_bets and select the target (the one the user named, else highest urgency then impact). It must have evidence_ids — a fix without evidence is out of scope.
Re-read the evidence. Use search_feedback / list_feedback on the bet's evidence_ids so the handoff describes observed behavior, not the bet's paraphrase of it.
Check for an open handoff. Call list_handoffs(status="open") and list_handoffs(status="in_progress") — if one already covers this bet, resume it instead of creating a duplicate.
Create the handoff. Call create_handoff with:
- bet_id and an imperative title ("Fix checkout 500 for business plans over 10 seats"),
- problem: symptom + reproduction in 2–4 sentences, citing fb_ ids,
- evidence_ids: the fb_ ids from step 2,
- acceptance: testable bullets — ALWAYS include "run_workspace_tests passes" plus the specific behavior (e.g. "POST /billing/checkout with plan=business, seats=12 returns 200").
Transfer to coder with the returned hf_ id placed in session state key handoff_id. The coder reads the workspace, makes the SMALLEST fix satisfying acceptance, runs tests until green, and calls update_handoff itself. Do not micro-manage its edits.
Verify on return. Call get_handoff(id) and trust only the stored record:
- status "done" AND result.tests shows passing → proceed.
- status "failed", or tests not green → nothing is fixed; go to step 7 with the failure.
Ledger the outcome. Call record_action with type "code_fix":
- on done: status "executed", target = the hf_ id, payload {files_changed: result.files_changed}, evidence_ids from the handoff, rationale linking bet → fix.
- on failed: status "failed", rationale = result.notes — a failure must be as auditable as a success.
Report: handoff id, files_changed, the test summary line, and the bet it closes. Suggest marking the bet "shipped" only after human review.

Failure handling

Coder reports failed → record it (step 7), surface result.notes verbatim, and propose ONE next step (sharper acceptance criteria, or human takeover). Never retry in a loop.
create_handoff returns an error → fix the shape (acceptance must be a non-empty list) and retry once.
Never edit workspace files yourself — only the coder has workspace tools.

Fix from feedback (bet → coder handoff → verified fix)

The closed loop: evidence picks the bug, a handoff scopes it, the coder fixes it in workspace/, tests prove it, the ledger records it.

Procedure

Pick the bet. Call list_bets and select the target (the one the user named, else highest urgency then impact). It must have evidence_ids — a fix without evidence is out of scope.

Re-read the evidence. Use search_feedback / list_feedback on the bet's evidence_ids so the handoff describes observed behavior, not the bet's paraphrase of it.

Check for an open handoff. Call list_handoffs(status="open") and list_handoffs(status="in_progress") — if one already covers this bet, resume it instead of creating a duplicate.

Create the handoff. Call create_handoff with:

bet_id and an imperative title ("Fix checkout 500 for business plans over 10 seats"),
problem: symptom + reproduction in 2–4 sentences, citing fb_ ids,
evidence_ids: the fb_ ids from step 2,
acceptance: testable bullets — ALWAYS include "run_workspace_tests passes" plus the specific behavior (e.g. "POST /billing/checkout with plan=business, seats=12 returns 200").

Transfer to coder with the returned hf_ id placed in session state key handoff_id. The coder reads the workspace, makes the SMALLEST fix satisfying acceptance, runs tests until green, and calls update_handoff itself. Do not micro-manage its edits.

Verify on return. Call get_handoff(id) and trust only the stored record:

status "done" AND result.tests shows passing → proceed.
status "failed", or tests not green → nothing is fixed; go to step 7 with the failure.

Ledger the outcome. Call record_action with type "code_fix":

on done: status "executed", target = the hf_ id, payload {files_changed: result.files_changed}, evidence_ids from the handoff, rationale linking bet → fix.
on failed: status "failed", rationale = result.notes — a failure must be as auditable as a success.

Report: handoff id, files_changed, the test summary line, and the bet it closes. Suggest marking the bet "shipped" only after human review.

Failure handling

Coder reports failed → record it (step 7), surface result.notes verbatim, and propose ONE next step (sharper acceptance criteria, or human takeover). Never retry in a loop.

create_handoff returns an error → fix the shape (acceptance must be a non-empty list) and retry once.

Never edit workspace files yourself — only the coder has workspace tools.

fix-from-feedback