Module effort_routing 

Per-turn effort routing (#1148, opt-in dynamic thinking budget).

Unlike the static effort.rs which applies a constant reasoning level across all turns (for cache stability), this module classifies each turn and adjusts thinking effort dynamically.

Opt-in only (proxy.effort_routing = true). When disabled, the static effort.rs path remains the sole controller. When enabled, this module overrides the static level with a per-turn classification.

Cache stability tradeoff

Provider prompt caches (Anthropic cache_control, OpenAI prefix caching) break when reasoning parameters change. This module accepts that tradeoff because:

1. Output tokens on Opus-class models cost 5x input tokens — savings from reduced thinking often exceed the cache-miss penalty.
2. Routine turns (file reads, passing tests) generate disproportionate thinking waste for trivial tool-result acknowledgements.
3. The module uses a two-level strategy (not N levels) to minimize cache key diversity: routine or full — only two cache prefixes to warm.

Classification

A turn is classified as routine when the last assistant message was a tool call and the tool result indicates success on a non-complex operation:

• File read (tool_use with ctx_read, read_file, Read)
• Successful shell command (exit_code == 0, no error indicators)
• Search results (grep/glob/find)
• Status checks (git status, test passing)

A turn is classified as full (keep maximum thinking) when:

• The user sent a new message (requires understanding intent)
• The tool result contains errors/failures
• Multiple tool results arrived (complex multi-step)
• The content is architecturally complex (refactoring, debugging)

Structs

RoutingStats

Statistics for monitoring effort routing effectiveness.

Enums

TurnClass

Turn classification result.

Functions

classify_turn

Classify the current turn based on the message array. Returns Routine if the latest context is a simple tool-result acknowledgement, Full otherwise.

classify_turn_anthropic

Classify based on Anthropic messages structure.

classify_turn_responses

Classify based on OpenAI Responses API input array (different structure).

effort_for_turn

Map a turn classification to the effort level to apply. base is the operator’s configured static effort level.

intent_aware_effort

Adjust effort from a classifier intent while preserving the configured effort for intents that are not clearly simple reads or coding work.

stats

Snapshot routing statistics.