reflect

What I do

• •Run a lightweight reflection after meaningful task work is complete.
• •Capture short, reusable lessons in docs/lessons/*.lessons.md files.
• •Prefer practical guidance over exhaustive retrospectives.
• •Support quick reflection directly and deep reflection via @reflector.

When to use me

Use this near the end of tasks that involved troubleshooting, trial-and-error, new workflows, or notable user feedback.

Skip heavy analysis. It is fine to add nothing when no durable lesson emerged.

Use deep mode when you want one agent to judge another agent's session quality.

Reflection checklist

1. •Scan what just happened in the task.
2. •Extract at most 1-3 generic lessons worth reusing.
3. •Focus on:
   • •successful command sequences or debugging flows
   • •recurring pitfalls and reliable fixes
   • •rule changes from .opencode/commands/rule.md
   • •maintenance updates needed in AGENTS.md or .opencode/* context files
4. •Write concise bullet points only (no long sections).
5. •Keep each lessons file short (target 10-50 lines, always under 100).

Deep review option (@reflector)

When context quality needs extra review, hand off to @reflector with either:

• •a task summary
• •transcript snippets
• •exported session context from opencode export <sessionID>

Then store only Lessons bullets returned by @reflector.

When useful, keep a timestamped deep-review artifact in docs/reflections/.

Reflection is also allowed to refine existing lessons and agent context docs (AGENTS.md, .opencode/skills/*, .opencode/agents/*) when updates are maintenance-oriented.

Guardrails

• •Never write secrets, tokens, credentials, or private one-off data.
• •Prefer durable, generic phrasing over task-specific details.
• •De-duplicate near-identical bullets instead of growing noise.
• •Avoid design-direction edits and one-off human preference/style changes unless explicitly requested.