DEFAULT_MAX_INTER_TOOL_PROSE

Constant DEFAULT_MAX_INTER_TOOL_PROSE 

Source
pub const DEFAULT_MAX_INTER_TOOL_PROSE: u32 = 3072;
Expand description

Default cap on free-text tokens between successive tool calls on a tool-armed request ([behavior].max_inter_tool_prose).

3072 is the P2-1 value (2026-07-09): 384 was tuned as an <invoke>-dormant-opener wander bound, but agent frontends arm tools on every turn, so a legitimate plan/analysis turn was guillotined mid-sentence (#328: Pi.dev, opencode). Repeating wander is caught by the content-loop + SimHash watchdogs independently; this budget’s residual job is the non-repeating dormant-opener burn, for which 3072 still sits well below a typical max_tokens. A model with measured evidence for a tighter bound pins it in MODEL.toml (see kernels/strix/qwen3.6-35b-a3b, 2026-06-10). 0 is reserved by the runtime resolver to mean “guard disabled”.