Adaptive Reasoning
Overview
Match cognitive effort to problem difficulty. Overthinking wastes tokens; underthinking causes rework.
Difficulty signals
| Signal | Mode |
|---|---|
| Typo, rename, single-file edit, clear instruction | Shallow - act immediately |
| Multi-file change, unclear bug, API design | Standard - inspect, plan briefly, act |
| Security, data loss, architecture, prod incident | Deep - explore alternatives, verify, then act |
| Conflicting requirements or missing facts | Clarify - ask 1-3 precise questions first |
| Chat / Ask mode with vague goal (even if a project folder is linked) | Clarify - questions first; no broad board/tree tour |
| Chat / Ask with no linked project | Clarify - never call tools until the target is concrete |
Workflow
- Classify the request using the table above (do not announce the label unless useful).
- Shallow: apply the change; skip long preambles.
- Standard:
- gather minimal context (
read_file/grep) - state a 2-4 line approach
- execute and verify
- gather minimal context (
- Deep:
- map constraints and failure modes
- compare 2 options when stakes are high
- verify with tests, dry-runs, or
execchecks - document the chosen trade-off in the final answer
- Clarify: call
ask_userwith 2-4 options and one recommended path. Do not ask an open question when a fork exists. If they skip, take the recommended option and say so.
Escalation
If a shallow task reveals surprises (unexpected deps, failing tests), escalate to Standard/Deep mid-turn without restarting from scratch.
Anti-patterns
- Writing a thesis for a one-line fix
- Jumping into code on security-sensitive changes without a threat check
- Asking many open-ended questions instead of a short plan with assumptions