UNPKG

@bitkyc08/opencodex

Version:

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

460 lines (429 loc) • 22.2 kB
import { expandRetainedCompactionInput } from "../../responses/retained-compaction"; import { createHash } from "node:crypto"; import { COMPACT_PROMPT, appendCheckpointContinuationNote, compactionItemToText, decodeCompactionSummary, isCompactionItemType } from "../../responses/compaction"; import { debugProviderDiagnostic } from "../../lib/debug"; import { modelInList } from "../../types"; import { isPlainObject } from "./internal"; import { activateDeferredTool } from "./tool-schema"; import { stripOpenAiOnlyWebSearchFields } from "./web-search"; /** * Codex requests output.encrypted_content, which xAI, Meta, Azure and Kimi Code reject with HTTP 400. * OpenCode Go uses the same compatibility policy because some of its Responses backends reject it. * The caller gates this on the destination. Preserve reasoning.encrypted_content * for stateless replay and leave unknown or malformed values for upstream validation. */ export function stripUnsupportedOutputInclude(body: unknown): unknown { if (!isPlainObject(body) || !Array.isArray(body.include)) return body; const include = body.include.filter(item => item !== "output.encrypted_content"); if (include.length === body.include.length) return body; if (include.length === 0) { const { include: _dropped, ...rest } = body; return rest; } return { ...body, include }; } /** 仅移除已验证孤立边界中无摘要、未被引用的推理项密文,保留整条历史项。 */ export function stripArkKimiOrphanCipher(body: unknown): unknown { if (!isPlainObject(body) || !Array.isArray(body.input)) return body; const input = body.input; const isUser = (item: unknown) => isPlainObject(item) && item.role === "user" && (item.type === undefined || item.type === "message"); const referenced = new Set(input.filter(item => isPlainObject(item) && item.type === "item_reference").map(item => item.id)); let changed = false; const next = input.map((item, i) => { // A matching call/result pair is the other confirmed boundary before an orphan. const call = input[i - 2], result = input[i - 1]; const afterTool = isPlainObject(call) && call.type === "function_call" && isPlainObject(result) && result.type === "function_call_output" && typeof call.call_id === "string" && call.call_id.length > 0 && result.call_id === call.call_id; // 仅放宽已配对工具结果之后的 developer 边界,其他角色仍保留密文。 const following = input[i + 1]; const developerAfterTool = afterTool && isPlainObject(following) && following.role === "developer" && (following.type === undefined || following.type === "message"); if ((!isUser(result) && !afterTool) || (!isUser(following) && !developerAfterTool) || !isPlainObject(item) || item.type !== "reasoning" || !Array.isArray(item.summary) || item.summary.length || typeof item.encrypted_content !== "string" || !item.encrypted_content || (item.id !== undefined && referenced.has(item.id)) // Codex can explicitly serialize the absent content channel as null. || !Object.keys(item).every(key => ["type", "id", "status", "summary", "encrypted_content"].includes(key) || (key === "content" && item.content === null))) return item; const { encrypted_content: _cipher, ...rest } = item; changed = true; return rest; }); return changed ? { ...body, input: next } : body; } /** * Provider-opaque replay metadata is client history, not an OpenAI input-item field (#7111). * The caller gates this on OpenAI-operated destinations; arbitrary gateways may accept it. * Strip only the known extension, never an item-field allowlist that can erase newer wire fields * or nested tool data. Copy on write so a later Google route can still replay its signature. */ export function stripProviderInputItemMetadata(body: unknown): unknown { if (!isPlainObject(body) || !Array.isArray(body.input)) return body; let input: unknown[] | undefined; for (let index = 0; index < body.input.length; index += 1) { const item = body.input[index]; if (!isPlainObject(item) || !Object.hasOwn(item, "extra_content")) continue; const next = { ...item }; delete next.extra_content; input ??= [...body.input]; input[index] = next; } return input ? { ...body, input } : body; } function isUrlCitationIndex(value: unknown): value is number { return typeof value === "number" && Number.isSafeInteger(value) && value >= 0; } /** Coerce a replayed citation index to a non-negative safe integer, falling back to 0. */ function repairUrlCitationIndex(value: unknown): number { if (isUrlCitationIndex(value)) return value; if (typeof value === "string" && /^\d+$/.test(value)) { const index = Number(value); if (isUrlCitationIndex(index)) return index; } return 0; } /** * OpenAI Responses requires `title` (string) and integer `start_index`/`end_index` on every * `url_citation` annotation of a replayed `output_text` part. History written by an older bridge * (title omitted when the search source had none) or a lossy client otherwise fails every later * request carrying it, most visibly remote compaction, with a deterministic HTTP 400 * (`Missing required parameter: input[N].content[M].annotations[K].title`). * Repair only those fields on `url_citation` annotations of assistant message content; other * annotation types, unknown keys and every other item are untouched. Copy on write: an already * valid body is returned by reference. */ export function repairUrlCitationAnnotations(body: unknown): unknown { if (!isPlainObject(body) || !Array.isArray(body.input)) return body; let input: unknown[] | undefined; for (let itemIndex = 0; itemIndex < body.input.length; itemIndex += 1) { const item = body.input[itemIndex]; if (!isPlainObject(item) || !Array.isArray(item.content)) continue; let content: unknown[] | undefined; for (let partIndex = 0; partIndex < item.content.length; partIndex += 1) { const part = item.content[partIndex]; if (!isPlainObject(part) || part.type !== "output_text" || !Array.isArray(part.annotations)) continue; let annotations: unknown[] | undefined; for (let annIndex = 0; annIndex < part.annotations.length; annIndex += 1) { const ann = part.annotations[annIndex]; if (!isPlainObject(ann) || ann.type !== "url_citation") continue; const titleOk = typeof ann.title === "string"; const startOk = isUrlCitationIndex(ann.start_index); const endOk = isUrlCitationIndex(ann.end_index); if (titleOk && startOk && endOk) continue; const next: Record<string, unknown> = { ...ann }; if (!titleOk) next.title = typeof ann.url === "string" ? ann.url : ""; if (!startOk) next.start_index = repairUrlCitationIndex(ann.start_index); if (!endOk) next.end_index = repairUrlCitationIndex(ann.end_index); annotations ??= [...part.annotations]; annotations[annIndex] = next; } if (!annotations) continue; content ??= [...item.content]; content[partIndex] = { ...part, annotations }; } if (!content) continue; input ??= [...body.input]; input[itemIndex] = { ...item, content }; } return input ? { ...body, input } : body; } export function stripInvalidItemIds(body: unknown): unknown { if (!isPlainObject(body) || !Array.isArray(body.input)) return body; const validPrefixes: Record<string, string> = { message: "msg_", agent_message: "amsg_", reasoning: "rs_", function_call: "fc_", custom_tool_call: "ctc_", tool_search_call: "tsc_", web_search_call: "ws_", }; let changed = false; const input = body.input.map(item => { if (!isPlainObject(item) || typeof item.type !== "string") return item; const validPrefix = validPrefixes[item.type]; if (!validPrefix) return item; if (typeof item.id === "string" && item.id.startsWith(validPrefix)) return item; if (!("id" in item)) return item; changed = true; const next = { ...item }; delete next.id; return next; }); return changed ? { ...body, input } : body; } /** * Codex-private tool fields that only the ChatGPT backend understands. * * A third-party Responses gateway validates its schema and rejects the whole request before * inference — xAI answers `Argument not supported: external_web_access` — so these are removed at * the noncanonical boundary while the tool and every public option stay. * * Keep this a table. Each private bit Codex attaches has so far arrived as its own bespoke strip * with its own traversal, and the traversals disagreed about which containers they covered; a new * one should be a row here instead. `toolTypes` omitted means the field is private on any tool. */ const CANONICAL_ONLY_TOOL_FIELDS: readonly { field: string; toolTypes?: ReadonlySet<string>; capabilityGated?: boolean }[] = [ // ChatGPT's browsing policy bit. The public hosted tool is enabled by its presence alone. // OWNERSHIP: official OpenAI API-key traffic and unclassified gateways ACCEPT this field, so // it is only stripped when the provider capability denies it (supportsOpenAiWebSearchToolFields // === false), matching stripOpenAiOnlyWebSearchFields; see // tests/responses/responses-routed-web-search-fields.test.ts. { field: "external_web_access", toolTypes: new Set(["web_search", "web_search_preview"]), capabilityGated: true }, // Deferred-discovery marker. `activateDeferredTool` clears it only for tools a `tool_search_output` // already loaded, so a still-deferred declaration — including one promoted out of a namespace // group — otherwise reaches the wire carrying it. { field: "defer_loading" }, ]; export function stripCanonicalOnlyToolFields(body: unknown, includeCapabilityGated: boolean): unknown { if (!isPlainObject(body)) return body; const rewriteTools = (tools: unknown[]): unknown[] => { let changed = false; const rewritten = tools.map(tool => { if (!isPlainObject(tool)) return tool; let next = tool; for (const { field, toolTypes, capabilityGated } of CANONICAL_ONLY_TOOL_FIELDS) { if (capabilityGated && !includeCapabilityGated) continue; if (!Object.hasOwn(next, field)) continue; if (toolTypes && (typeof next.type !== "string" || !toolTypes.has(next.type))) continue; const { [field]: _private, ...rest } = next; next = rest; } if (next === tool) return tool; changed = true; return next; }); return changed ? rewritten : tools; }; let rewrittenBody = body; if (Array.isArray(body.tools)) { const tools = rewriteTools(body.tools); if (tools !== body.tools) rewrittenBody = { ...rewrittenBody, tools }; } if (!Array.isArray(body.input)) return rewrittenBody; let input: unknown[] | undefined; for (let index = 0; index < body.input.length; index += 1) { const item = body.input[index]; if (!isPlainObject(item) || item.type !== "additional_tools" || !Array.isArray(item.tools)) continue; const tools = rewriteTools(item.tools); if (tools === item.tools) continue; input ??= [...body.input]; input[index] = { ...item, tools }; } return input ? { ...rewrittenBody, input } : rewrittenBody; } /** * Codex keeps this ChatGPT-internal item metadata when its configured provider name is `openai`. * Loopback OpenCodex injection intentionally retains that provider identity for history continuity, * even when the proxy ultimately routes the request to a public Responses destination. Those * destinations reject the private field as an unknown `input[*]` parameter, so remove it at the * noncanonical boundary without mutating the caller-owned raw body. */ export function stripInternalChatMessageMetadataPassthrough(body: unknown): unknown { if (!isPlainObject(body) || !Array.isArray(body.input)) return body; let changed = false; const input = body.input.map(item => { if (!isPlainObject(item) || !Object.hasOwn(item, "internal_chat_message_metadata_passthrough")) { return item; } changed = true; const next = { ...item }; delete next.internal_chat_message_metadata_passthrough; return next; }); return changed ? { ...body, input } : body; } /** * OpenAI-private TOP-LEVEL request keys, the sibling of `CANONICAL_ONLY_TOOL_FIELDS` one level up. * * Codex attaches these on the request body itself rather than on a tool or an input item, and gates * them on its own auth rather than on the destination URL. Loopback injection keeps Codex pointed at * its built-in `openai` provider, so the client still believes it is addressing the canonical * ChatGPT backend and keeps the key no matter where this proxy routes the turn. A Responses gateway * that validates its top-level schema then rejects the whole request before inference. * * Keep this a table, and keep it to keys a client is OBSERVED to send. It is not an unknown-field * sanitizer: a top-level key nobody has traced to a client is forwarded untouched, because deleting * it would silently drop a parameter some other caller means. */ const CANONICAL_ONLY_TOP_LEVEL_FIELDS: readonly string[] = [ // Cyber access program selector, new in Codex 0.155. codex-rs mints it from // `cyber_access_program::for_auth`, which filters on ChatGPT auth alone and never on the // destination base URL, and serializes it on the Responses request, the compaction input and the // WebSocket `response.create` envelope. No public specification defines it, so a strict // third-party gateway answers with an unknown-parameter error naming it, and every turn of that // thread fails (#4853). // // `codex_output_schema` is deliberately NOT here. In codex-rs it is the `name` of the JSON-schema // `text.format` object, not a top-level key, so listing it would delete a field this client never // sends and discard it for any client that does send it meaningfully. "access_programs", ]; /** * Remove the OpenAI-private top-level keys. * * The caller decides the boundary; see the call site in `passthrough.ts`, which applies this only * to a destination OpenCodex does not operate. Returns the input unchanged when no listed key is * present, so the common path allocates nothing and the caller-owned raw body is never mutated. */ export function stripCanonicalOnlyTopLevelFields(body: unknown): unknown { if (!isPlainObject(body)) return body; if (!CANONICAL_ONLY_TOP_LEVEL_FIELDS.some(field => Object.hasOwn(body, field))) return body; const next = { ...body }; for (const field of CANONICAL_ONLY_TOP_LEVEL_FIELDS) delete next[field]; return next; } /** * OpenAI-private input item TYPES, the item-level sibling of the two tables above. * * Each one is a positional control the OpenAI Responses backend interprets where it sits in * history, never a turn of the conversation, so dropping it leaves every message, call and output * in place. Codex gates sending one on its own `is_openai` provider check, which loopback * injection always satisfies, so a gateway this proxy does not operate can still receive one and * reject the whole body as an unknown item type. * * Same rule as the top-level table: only types a client is OBSERVED to send. */ const CANONICAL_ONLY_INPUT_ITEM_TYPES: ReadonlySet<string> = new Set([ // Codex 0.164 `ResponseItem::ConfigurationUpdate` ({type, reasoning: {effort}}): a durable // reasoning-effort change recorded at its position in history. codex-rs sends it only for a // catalog row with `supports_reasoning_effort_updates`, and keeps it in saved history after a // model switch. Mapping the effort onto another provider is out of scope (#6895). "configuration_update", ]); /** * Remove OpenAI-private input items. The caller decides the boundary, exactly as for * {@link stripCanonicalOnlyTopLevelFields}; the canonical and OpenAI-operated destinations keep * the items verbatim and in order. Returns the input unchanged when none is present. */ export function isCanonicalOnlyInputItem(item: unknown): boolean { return isPlainObject(item) && typeof item.type === "string" && CANONICAL_ONLY_INPUT_ITEM_TYPES.has(item.type); } export function stripCanonicalOnlyInputItems(body: unknown): unknown { if (!isPlainObject(body) || !Array.isArray(body.input)) return body; if (!body.input.some(isCanonicalOnlyInputItem)) return body; return { ...body, input: body.input.filter(item => !isCanonicalOnlyInputItem(item)) }; } /** * Sampling fields a model on the provider's `noStopModels` / `noPenaltyModels` list rejects on * every wire. Claude inbound translates `stop_sequences` into a Responses `stop` * (`src/claude/inbound.ts`), and a direct Responses caller can send penalties. Some listed models * only have the Responses wire (xAI grok-4.20-multi-agent-0309 answers Chat Completions with 400), * so the Chat adapter's omission cannot cover them. Returns the input unchanged when nothing is * removed, so the caller-owned raw body is never mutated. */ export function stripRejectedSamplingParams( body: unknown, provider: { noStopModels?: string[]; noPenaltyModels?: string[] }, modelId: string, ): unknown { if (!isPlainObject(body)) return body; const dropStop = Object.hasOwn(body, "stop") && modelInList(provider.noStopModels, modelId); const penalties = ["presence_penalty", "frequency_penalty"].filter(field => Object.hasOwn(body, field)); const dropPenalties = penalties.length > 0 && modelInList(provider.noPenaltyModels, modelId); if (!dropStop && !dropPenalties) return body; const next = { ...body }; if (dropStop) delete next.stop; if (dropPenalties) for (const field of penalties) delete next[field]; return next; } /** * When `store: false`, the upstream cannot retrieve stored turns by ID. An ID * forwarded in `input` is then interpreted as a reference to a stored item that does not * exist, producing a 404. Strip all item IDs in this case unless `preserveIds` is true * — `call_id` pairing is unaffected. Also repairs missing or broken xAI custom_tool_call IDs * when `requireCustomCallIds` is true, regardless of `preserveIds`. * Matches codex-rs behavior (core/src/client.rs:918-925). */ export function stripItemIdsWhenUnstored( body: unknown, requireCustomCallIds = false, preserveIds = false, ): unknown { const repairCustomCallIds = requireCustomCallIds === true; if (!isPlainObject(body) || (body.store !== false && !repairCustomCallIds)) return body; if (!Array.isArray(body.input)) return body; let changed = false; const input = body.input.map(item => { if (!isPlainObject(item)) return item; if (repairCustomCallIds && item.type === "custom_tool_call") { try { if (typeof item.id === "string" && item.id.startsWith("ctc_")) return item; if ( typeof item.call_id !== "string" || typeof item.name !== "string" || typeof item.input !== "string" ) return item; const digest = createHash("sha256") .update(JSON.stringify([item.call_id, item.name, item.input])) .digest("hex") .slice(0, 40); changed = true; debugProviderDiagnostic("openai-responses", "xai-custom-tool-call-id-repaired", { hadId: typeof item.id === "string", }); return { ...item, id: `ctc_${digest}` }; } catch { debugProviderDiagnostic("openai-responses", "xai-custom-tool-call-id-unrepaired", {}); return item; } } if (preserveIds || body.store !== false || !("id" in item)) return item; changed = true; const next = { ...item }; delete next.id; return next; }); return changed ? { ...body, input } : body; } /** * Normalize replayed compaction items for the destination backend. * * A compaction item carries an `encrypted_content` blob the client replays verbatim on every later * turn, and only the backend that minted it can decode it. Proxy-minted `ocx1:` and structured `ocx2:` envelopes are * transparent base64 rather than encryption, so no upstream can read them and they always become * plain user messages. Native blobs have multiple possible minters, so a destination's ability to * decode its own blobs does not make a blob from a previous serving identity portable. On a known * identity mismatch the blob degrades to the same note the bridged parser uses, even when the * destination normally accepts native blobs. Without a known mismatch, the destination capability * keeps the existing behavior. * * A bare `context_compaction` marker carries no blob and is forwarded untouched. */ export function scrubOcxCompactionItems( body: unknown, destinationDecodesNativeBlob: boolean, threadServingIdentityChanged: boolean, addContinuationNote = false, ): unknown { if (!isPlainObject(body) || !Array.isArray(body.input)) return body; const expanded = expandRetainedCompactionInput(body.input) as unknown[]; let changed = expanded !== body.input; let decodedCheckpoint = false; const input = expanded.map(item => { if (!isPlainObject(item) || !isCompactionItemType(item.type)) return item; const encrypted = typeof item.encrypted_content === "string" ? item.encrypted_content : undefined; if (encrypted === undefined) return item; if ( decodeCompactionSummary(encrypted) === null && destinationDecodesNativeBlob && !threadServingIdentityChanged ) return item; changed = true; if (decodeCompactionSummary(encrypted) !== null) decodedCheckpoint = true; return { type: "message", role: "user", content: [{ type: "input_text", text: compactionItemToText(encrypted) }], }; }); if (!changed) return body; // #6764: a decoded checkpoint is plain user text from here on, so name its role in the // instructions. Compaction requests skip this; the summarizer must not be told to continue. return addContinuationNote && decodedCheckpoint ? { ...body, input, instructions: appendCheckpointContinuationNote(body.instructions) } : { ...body, input }; }