Select the right agent lane, escalate deliberately, and avoid ad hoc reasoning overrides.
I am the source of truth for OpenCode lane selection, escalation, and reasoning discipline.
Treat agent selection as the primary control for cost, speed, and reasoning quality. Do not try to compensate for the wrong agent choice with longer prompts or ad hoc requests to think harder.
Before substantial work, classify the task into one of these lanes:
explore for file discovery, symbol search, dependency tracing, and read-only narrowingsummary-helper for logs, long command output, long docs, and raw text compressiongeneral for normal implementation and edits after scope is clearfast-helper for low-risk helper work, quick hypotheses, and cheap utility tasks; treat its output as advisory until verifieddeep-review for wide diffs, subtle judgment, repo-level synthesis, and final review on higher-risk changespremium-review only when mistakes are expensive and extra judgment is likely to pay for itselfexperimental-open only for cheap helper work, summaries, or alternative drafts; do not rely on it as the final authority for repo-affecting changesexploresummary-helpergeneralgeneral to deep-review before another broad attemptdeep-reviewpremium-review only for high-consequence judgment, not routine coding