UNPKG

@bitkyc08/opencodex

Version:

Universal provider proxy for OpenAI Codex & Claude Code — use any LLM with Codex CLI/App/SDK and Claude Code

37 lines (35 loc) • 1.88 kB
/** * Shadow plan recording (PF-12, `protocols.rollout.shadowPlan`, default off). * * SERVER SIDE (not a leaf): it reads config through `plan-snapshot.ts`. An ingress calls * `recordProtocolShadowPlan` right after its entry mark. With the switch off it returns before * reading anything else. With it on it builds the same side-effect-free snapshot the dashboard * preview uses, with basis `dispatch`, from the selector the client sent and the features the * entry mark already collected, and stores it beside the marks. `trace.ts` runs the pure planner * on it at finalize and compares. * * Nothing here sends, fetches, advances combo state or writes: the snapshot expands combos and * policies from config and uses `routeModel`'s deterministic branches only. The stored input is * fixed vocabulary plus the provider and model names the server already exposes. Any throw is * swallowed: a shadow plan is diagnostics and never fails or changes the request it describes. */ import type { OcxConfig } from "../types"; import type { Protocol } from "./contract"; import { buildProtocolPlanSnapshot } from "./plan-snapshot"; import { resolveProtocolSettings } from "./settings"; import { markProtocolShadowPlanInput, protocolMarkFeatures } from "./trace"; export function recordProtocolShadowPlan( logCtx: object, config: OcxConfig, request: { inbound: Protocol; model: unknown }, ): void { try { if (!resolveProtocolSettings(config).rollout.shadowPlan) return; if (typeof request.model !== "string" || request.model.length === 0) return; const features = protocolMarkFeatures(logCtx) ?? []; const input = buildProtocolPlanSnapshot(config, { model: request.model, inbound: request.inbound, features }, "dispatch"); markProtocolShadowPlanInput(logCtx, input); } catch { /* a shadow plan must never fail the request it describes */ } }