| name | reflect |
| description | On-demand mistake/learning logger. Records errors, corrections, and insights to `.agents-stack/learnings/` for cross-session recall. Also queries past learnings on request. |
| trigger | On demand. User says "remember this mistake," "log this error," "show me past learnings," "recall," or similar. |
| inputs | ["user's natural language description of the mistake or learning"] |
| outputs | [".agents-stack/learnings/<date>--<slug>.md (record), or printed recall results (query)"] |
| boundaries | Read and write `.agents-stack/learnings/` only. No code changes. No pipeline integration. |
Reflect Worker
Record a learning or recall past learnings — on demand, when the user says so.
Record a Learning
When the user says something like "remember this mistake" or similar:
- Extract the learning from context (what went wrong, why, what to do instead)
- Write to
.agents-stack/learnings/<YYYY-MM-DD>--<slug>.md
Format:
---
title: Short, specific title
type: error | correction | insight
source: <workstream-id or context>
trigger: When does this apply? Be specific enough for future recall.
---
What went wrong or what was learned.
Why it happened — not the symptom.
What to do instead — actionable, not abstract.
Optional: workstream IDs, phase, or file paths.
Type Guide
| Type | When |
|---|
error | A mistake was made (wrong approach, bug introduced) |
correction | User corrected the agent's behavior or preference |
insight | Non-obvious solution or pattern discovered |
Recall Past Learnings
When the user says "show me past learnings," "recall," or similar:
- Read all files from
.agents-stack/learnings/
- Filter by any keywords or type the user specifies
- Print a concise summary grouped by type
If the user asks about a specific area (e.g., "what did we learn about auth"), filter by keyword match on title, trigger, and content.
Rules
- One learning per file. If the same issue recurs, create a new entry with a later date.
- Prevention rules must be actionable ("always do X") — not vague ("be more careful").
- Do not edit or delete existing learnings. Create new ones.
- Recall reads only — never modify during recall.
