qmistake

name	Qmistake
description	Record mistakes so they are never repeated. When the user points out a mistake, confirms it and saves to .qe/MISTAKE.md. Loaded at every session start.
invocation_trigger	When the user says something was wrong, points out a mistake, corrects behavior, or says "don't do that again". Also triggered by /Qmistake directly.
recommendedModel	haiku

Qmistake — Mistake Registry

Role

Records mistakes that Claude made so they are never repeated. Each mistake is stored in .qe/MISTAKE.md which is loaded at every session start via the session-start hook.

When to Trigger

• User explicitly says: "that's wrong", "mistake", "don't do that", "you keep doing X"
• User corrects Claude's behavior or output
• User calls /Qmistake directly
• After a bug is traced back to Claude's incorrect approach

Workflow

Step 1: Identify the Mistake

Parse the user's complaint/correction. Extract:

• What happened: What Claude did wrong
• What should have happened: The correct behavior
• Context: Which skill/agent/file was involved (if applicable)

Step 2: Confirm with User

Use the QE interaction adapter to confirm before recording:

Mistake identified:

  What went wrong: {description}
  Correct behavior: {what should happen instead}
  Context: {skill/file/situation}

Record this mistake?
  (a) Yes, record it
  (b) Edit description first
  (c) Cancel

If user picks (b), ask for the corrected description, then confirm again.

Step 3: Write to MISTAKE.md

Append the entry to .qe/MISTAKE.md. Create the file if it doesn't exist.

File format:

# Mistake Registry

> These mistakes were identified by the user. Do NOT repeat them.
> This file is loaded at every session start.

---

### M001: {short title}
- **Date**: {YYYY-MM-DD}
- **Wrong**: {what Claude did wrong}
- **Correct**: {what should happen instead}
- **Context**: {skill/agent/file involved}
- **Severity**: {critical / important / minor}

---

Entry numbering: M001, M002, M003... Auto-increment by reading the last entry number.

Severity levels:

• critical — Causes data loss, breaks builds, sends wrong output to user
• important — Wrong approach, wasted time, incorrect assumption
• minor — Style issue, suboptimal choice, cosmetic

Step 4: Confirm

Report: "Recorded as M{NNN}. This will be loaded at every session start."

Subcommands

/Qmistake (no args)

Show current mistake count and last 5 entries.

/Qmistake add

Interactive: ask what went wrong, confirm, save.

/Qmistake list

Show all entries from .qe/MISTAKE.md.

/Qmistake resolve M{NNN}

Mark a mistake as resolved (add [RESOLVED] prefix). Resolved mistakes stay in the file for reference but are de-prioritized in session injection.

/Qmistake learn

Promote auto-extracted recurring failure patterns into the registry (Phase 4, D021).

1. Run the extractor (deterministic, zero-dep, no model calls):

   npm run learn:failures   # → node scripts/learn-from-failures.mjs
   

   It scans .qe/learning/failures/** (auto-captured on Stop by failure-capture.mjs), counts each pattern by distinct task_uuid, and writes recurring ones (≥ 2 distinct tasks) to .qe/learning/mistake-candidates.md.
2. Read .qe/learning/mistake-candidates.md. For each candidate, present it via the interaction adapter exactly as in Step 2 above — the user decides whether it is a genuine systemic mistake worth recording.
3. On confirmation, write it through the normal Step 3 path (M{NNN} entry in .qe/MISTAKE.md). Skip candidates the user rejects.

This is candidate-suggestion only: the extractor never writes to .qe/MISTAKE.md, and promotion always passes through user confirmation (D011 user-in-the-loop).

Session Integration

The session-start.mjs hook reads .qe/MISTAKE.md and injects its content into the session context. This ensures every new conversation starts with awareness of past mistakes.

Injection format (in hook output):

[MISTAKES] {N} recorded mistakes. Critical: {X}, Important: {Y}.
Top entries:
- M001: {title} — {wrong behavior}
- M002: {title} — {wrong behavior}
Full list: .qe/MISTAKE.md

Only unresolved entries are injected. If there are more than 10, show only critical + important.

Will

• Record user-identified mistakes with confirmation
• Auto-number entries
• Integrate with session-start hook for cross-session persistence
• Surface auto-extracted recurring-failure candidates (/Qmistake learn) for user-confirmed promotion

Will Not

• Auto-record mistakes without user confirmation (auto-extraction produces candidates only; promotion is always user-confirmed)
• Delete entries (only resolve)
• Modify project code or framework files

See also: Qlearn (superset)

Qmistake records mistakes to .qe/MISTAKE.md. /Qlearn is the broader cross-session learning memory (mistakes + conventions + gotchas + decisions) in .qe/learnings.md, ranked by time-decayed relevance and injected top-N at session start. Prefer Qlearn add --type mistake for new lessons that benefit from decay ranking; Qmistake stays valid and the two registries coexist.