symforge 8.10.6

SymForge — in-memory code intelligence MCP server for Claude Code, Codex, and Gemini CLI
Documentation
//! Protocol-free token-economics constants shared by the calibration math and
//! the (server-only) L2 controller.
//!
//! D3-ROOT extract-up: these four `u32` floors were the ONLY tie binding the
//! protocol-free `calibration` math to the server-only `controller`. Lifting
//! them into `stel_core` lets the durable ledger + calibration compile under
//! `any(server, embed)`. The server-side `controller` re-exports them
//! (`pub use crate::stel_core::consts::*;`) so every existing caller path
//! (`crate::stel::controller::COMPACT_SCHEMA_TOKENS`, ...) resolves unchanged.

/// Compact-3 worst-case schema tax per call (A-006 conservative path; no amortization credit).
pub const COMPACT_SCHEMA_TOKENS: u32 = 45;
/// Compact `symforge` invoke overhead per call (schema example + Phase 0 doctrine).
pub const COMPACT_INVOKE_TOKENS: u32 = 80;
/// Static per-step predicted-response floor used when a step carries no real
/// byte sizes (plan-only callers / fixtures). Matches the planner's
/// `est_response_tokens` (`planner.rs`). The auto-tune (feature 013, D8-ROOT)
/// does NOT replace this floor — it learns one multiplicative
/// `response_correction_factor` applied to the predictor's response OUTPUT
/// (byte-grounded OR this floor) after grounding-or-floor, so the floor stays the
/// raw sub-model and the correction is layered on top.
pub const STATIC_RESPONSE_FLOOR: u32 = 400;
/// Static per-step manual-baseline floor (the `est_manual_tokens` counterpart to
/// [`STATIC_RESPONSE_FLOOR`]). The auto-tune leaves the manual baseline UNCHANGED
/// (D9): only the predictor's response output is corrected, never the manual
/// figure or the fixed schema/invoke overheads.
pub const STATIC_MANUAL_FLOOR: u32 = 800;