Reflect

Reflect

Mine the current conversation for durable learnings. Classify them. Show the user. Wait for explicit approval. Only then edit.

Do not auto-edit skills. Do not glob private histories. Do not use Cursor create-skill. Skill edits, once approved, follow writing-for-agents discipline (SKILL.md / descriptions / pointers).

When to invoke

• The user said "reflect" or "/reflect".
• A complex task just landed cleanly and the recipe is worth keeping.
• The agent hit dead ends, found the working path, and the path generalizes.
• The user corrected the agent's approach mid-task.

Skip when the conversation is trivial, off-topic, or already covered by an existing skill the parent followed correctly. One-offs are not learnings.

Process

1. Locate the active transcript

Use only the active transcript path the system prompt names. Do not glob Claude, Cursor, or OpenCode session transcript directories. If no safe path resolves, write a tight digest of this session and use that.

2. Review

Read this session (or the digest). Optionally spawn isolated reviewers from ~/.config/opencode/bestfriend/config/model-pool.json. Missing pool or inherit-parent → review inline and set MODEL_DIVERSITY=false. Do not invent commercial slugs.

Prompt templates live in references/ (judgment, tooling, divergent, synthesizer). Reviewers must not write files.

3. Classify

Every finding is one of:

• STRUCTURAL — a lint rule, script, metadata flag, or runtime check would enforce it better than a skill bullet. File as BACKLOG unless the user asks to implement the gate now.
• SKILL — a durable edit to an existing owned skill or a new skill draft.
• BACKLOG — tracker item, not a skill edit.

4. Show the user. Stop.

Present the full STRUCTURAL / SKILL / BACKLOG list. Wait for explicit approval of which subset to apply. The user may redirect or reject everything.

Do not edit any skill before that approval. Auto-filing a tracker is also off unless the user asked.

5. Apply only the approved SKILL items

• Trivial existing-skill edit (one-line bullet, stale fact): parent does it.
• Substantive skill edit: follow writing-for-agents. Do not invent a Cursor create-skill loop.
• New skill: draft a SKILL.md and show it; do not register it as owned product unless the user asked to vendor it.

If this environment has lib/validate_skills.py, run it on every touched owned skill before declaring done.

6. Summarize

• Edits applied (only approved ones): path + one line.
• Drafts shown, not applied.
• BACKLOG items listed, not filed unless asked.
• Dropped findings + reason.