znlgis Reflect
A lightweight continuous-improvement loop for the opencode configuration. After enough sessions, patterns emerge: the orchestrator routes wrong, a skill is incomplete, a rule is ignored, a workflow is missing. Reflect surfaces these and proposes minimal fixes.
When to use
- The user says "reflect on our sessions" or "review the config"
- You notice the same friction in 3+ sessions (same agent fails, same skill loaded but misused, same rule violated)
- After a major config change — check that nothing broke
Workflow
Step 1: Gather signals
Review what's available from the current session and recent history:
- What agents were dispatched? Any failures or escalations?
- What skills were loaded? Were any requested but missing?
- What AGENTS.md rules were violated or misunderstood?
- What commands were used? Any templating issues?
- Were there unnecessary token burns (large file reads that should've been explored, sequential reads that could've been parallel)?
Step 2: Identify patterns
Look for low-cost, high-impact fixes:
| Signal | Possible Fix |
|---|---|
| Flash agent escalates to pro 3+ times for same category | Adjust agent description or routing rules |
| Skill triggered but agent ignores it | Skill description too vague — add trigger keywords |
| Repeated style violations | Add explicit rule to AGENTS.md or agent prompt |
| Agent keeps trying to delegate when it shouldn't | Add "no delegation" constraint to that agent |
| Sequential reads where parallel would work | Not a config fix — note in report as behavioral |
| Large diffs overwhelm reviewer | Threshold in code-review skill may be too high |
Step 3: Propose changes
Present findings as a checklist:
## Reflect findings
### Patterns found
- [pattern]: [evidence in N sessions]
### Proposed changes
1. [change to file]: [why]
2. [change to file]: [why]
### Token savings estimate
- [N tokens saved per session by each change]
### NOT recommended
- [idea]: [why it's wrong or too expensive]
Only propose changes that are:
- Durable — applies to future sessions, not just this one
- Minimal — smallest config/prompt change that fixes the pattern
- Verifiable — you can check after the next session whether it helped
Rules
- Do not propose changes without evidence (at least 2 sessions showing the pattern).
- Do not add complexity to fix a one-off issue. Reflect finds recurring friction.
- When proposing changes to agent prompts, prefer tightening (more specific triggers) over broadening (looser triggers cause more false routing).
- Follow the opencode-config skill when editing agents/skills/commands/ files.