claude-code - 💡(How to fix) Fix [Feature] Model-initiated dynamic effort allocation for multi-step operations

Problem

Effort level is set per-session via /model and stays constant across whatever Claude does next. But within a single session — especially during agentic multi-step work — reasoning load varies wildly:

• Low load: polling a background process, watching a monitor, simple file edits, summarization
• Medium load: routine code synthesis, well-known refactors, conversational responses
• High load: substrate debugging with symptom far from cause, architectural design with compounding decisions, multi-file refactor with subtle invariants

The current model forces operators to choose one level for the whole session. Pick High → burn tokens on polling loops. Pick Medium → leave correct architectural paths un-found on hard debugging. Existing UX (#65134, #36303) helps operators adjust manually, but the operator has worse information than the model about which step needs which level.

Proposed

Let Claude itself suggest effort changes inline. Two flavors worth considering:

1. Suggestion-only: model emits a structured hint per tool sequence (e.g., [effort-suggested: high] with a brief reason); operator sees it inline and /model adjusts if they agree. Safest — operator stays in control of cost.
2. Bounded auto-adjust: operator configures a range (--effort-min low --effort-max high); model self-adjusts within that range per operation. Surfaced in transcript so cost is visible.

Either flavor needs:

• Visibility (operator sees when/why a suggestion fires; transcript shows actual level per call)
• Bounds (operator can cap max to control cost)
• Quick override (a /effort lock <level> to disable dynamic mode for a turn or two)

Why now

The substrate-style use case (Claude Code driving a multi-agent pilot, monitoring background processes for completions, then doing complex code synthesis when results land) has wildly varying reasoning load within a single session. Today the operator has to either pick the max and pay everywhere, or pick the floor and miss hard reasoning. Model-initiated effort is the right primitive for agentic-loop workflows.

Distinguished from related

• #65134 — operator-UX: can't change /effort while prompt is being typed. Different pain point.
• #36303 (closed/stale) — operator-initiated per-query effort with auto-revert. Still operator-initiated; this proposal is model-initiated.