UNPKG

@bitkyc08/opencodex

Version:

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

76 lines (70 loc) • 3.6 kB
/** * Recognise an oversized Devin history behind Cognition's opaque refusal. * * Measured live on swe-1-6 (200k context): a 1.2 MB or 3 MB history is refused * after ~11s with `invalid_argument` and no output, the same code a bad tool * schema gets. Passed through as a plain 400, Codex never learns its context is * full and never compacts, so every later turn dead-ends the same way. Only a * request large against the model's window is reclassified; a small request * with the same code stays an ordinary 400. * * Boundary, binary-searched live on swe-1-6 (200k catalog window, 8192 output): * accepted at 199,186 prompt tokens of prose and 200,345 of JSON, refused at * ~203k of either. Raising the output cap to 32000 did not move it (195,571 * input tokens still accepted), so the window is not reduced by the output * reservation and the threshold is the window itself. */ import type { AdapterEvent } from "../../types"; import type { ChatHistoryItem, ToolDef } from "./cloud-direct"; import { prepareToolDescriptionForCognition } from "./cloud-direct/chat"; /** * Share of the window the estimate must reach. Characters per real token ran * from 1.33 (Korean) through 2.31 (JSON) and 4.23 (source code) to 5.53 * (prose), so no character ratio separates "over the window" from "60% of it". * The word-piece count below read 1.00x to 1.54x of the real count on those * samples: at 0.95 every sample at the window is caught and none at 60% is. */ const WINDOW_SHARE = 0.95; /** One token per short letter run, 1-3 digit group, or other visible character. */ const WORD_PIECE = /[A-Z]?[a-z]{1,8}|[A-Z]{1,8}(?![a-z])|\d{1,3}|\S/g; /** Text size that counts as oversized when the model's window is unknown. */ const UNKNOWN_WINDOW_CHARS = 512 * 1024; export const DEVIN_CONTEXT_OVERFLOW_MESSAGE = "Devin rejected this turn and its history is at or past the model's context window. Compact the conversation or start a new session."; /** Text the model would tokenize; image bytes are excluded so a screenshot cannot look like a long history. */ function requestText(messages: ChatHistoryItem[], tools: ToolDef[] | undefined): string { const parts: string[] = []; for (const m of messages) { if (typeof m.content === "string") parts.push(m.content); else for (const p of m.content) if (p.type === "text") parts.push(p.text); for (const call of m.tool_calls ?? []) parts.push(call.arguments); if (m.thinking) parts.push(m.thinking); } for (const tool of tools ?? []) parts.push(prepareToolDescriptionForCognition(tool.description ?? ""), JSON.stringify(tool.parameters ?? {})); return parts.join("\n"); } export function isDevinHistoryOverflow(input: { code: string | undefined; producedOutput: boolean; contextWindow: number | undefined; messages: ChatHistoryItem[]; tools: ToolDef[] | undefined; }): boolean { if (input.code !== "invalid_argument" || input.producedOutput) return false; const text = requestText(input.messages, input.tools); if (!input.contextWindow) return text.length >= UNKNOWN_WINDOW_CHARS; let pieces = 0; for (const _ of text.matchAll(WORD_PIECE)) pieces++; return pieces >= input.contextWindow * WINDOW_SHARE; } /** Same terminal shape the Kiro adapter uses, which Codex reads as "context full, compact". */ export function devinContextOverflowEvent(): Extract<AdapterEvent, { type: "error" }> { return { type: "error", message: DEVIN_CONTEXT_OVERFLOW_MESSAGE, status: 400, errorType: "invalid_request_error", code: "context_length_exceeded", retryable: false, }; }