@juspay/neurolink 12.36.1 → 12.37.0

package/dist/proxy/codexResponsesFormat.js

@@ -0,0 +1,591 @@
+/**
+ * Wire-format translation: Anthropic `ClaudeResponse` shape -> the Codex
+ * Responses shape a native Codex client expects.
+ *
+ * Pure wire-format only — no account selection, retry policy, or transport.
+ * Mirrors `claudeFormat.ts`'s role for the opposite direction: that module
+ * serializes NeuroLink's internal result into Claude SSE, this one serializes
+ * an already-Claude-shaped result into Codex SSE (or one JSON envelope).
+ */
+import { randomBytes } from "node:crypto";
+import { formatSSE } from "./claudeFormat.js";
+/** Empty by default so every existing caller (no `toolKindByName` argument)
+ *  keeps treating every tool_use block as a `function` call, unchanged. */
+const NO_TOOL_KINDS = new Map();
+/**
+ * A `custom` tool was wrapped, on the request side, into a single-field
+ * `{ input: "..." }` schema (§3.4 of `codexOutboundFallback.ts`), so the
+ * client's grammar text is that one field. Undefined when the record does not
+ * carry it as a string.
+ */
+function unwrapCustomToolInput(value) {
+    return typeof value === "object" &&
+        value !== null &&
+        !Array.isArray(value) &&
+        "input" in value &&
+        typeof value.input === "string"
+        ? value.input
+        : undefined;
+}
+/** The accumulated `input_json_delta` text of a custom call, unwrapped, or
+ *  undefined when it is not (yet) a complete `{ input: "..." }` object. */
+function unwrapCustomToolInputJson(text) {
+    try {
+        return unwrapCustomToolInput(JSON.parse(text));
+    }
+    catch {
+        return undefined;
+    }
+}
+/** Same idiom as `generateToolUseId` (`claudeFormat.ts`), Codex's own prefix. */
+export function generateCodexResponseId() {
+    return `resp_${randomBytes(24).toString("base64url")}`;
+}
+const CODEX_ITEM_ID_PREFIX = {
+    message: "msg_",
+    function_call: "fc_",
+    custom_tool_call: "ctc_",
+};
+/** Item ids are prefixed by kind so a client can tell message/tool items apart at a glance. */
+export function generateCodexItemId(kind) {
+    return `${CODEX_ITEM_ID_PREFIX[kind]}${randomBytes(18).toString("base64url").slice(0, 24)}`;
+}
+/**
+ * Reuse the Anthropic `tool_use.id` (`toolu_...`) verbatim as the Codex `call_id`.
+ * No id-mapping table exists anywhere in this codebase's translation modules, and
+ * nothing on the way in validates `call_id` format — see the design doc's §5 for
+ * the full reasoning and its one unresolved edge (a real, non-fallback Codex
+ * backend rejecting a non-`call_`-prefixed id on a later turn).
+ */
+export function codexCallIdFromToolUseId(toolUseId) {
+    return toolUseId;
+}
+/**
+ * `ClaudeResponse.stop_reason` is `string | null` with no enum in this repo
+ * (`proxy.ts`), so any value this repo cannot name — including a real
+ * Anthropic "refusal" or "pause_turn" — falls into "completed" rather than
+ * surfacing as unusual. Deliberate simplification, stated explicitly.
+ */
+function mapClaudeStopReasonToCodex(stopReason) {
+    if (stopReason === "max_tokens") {
+        return {
+            status: "incomplete",
+            incomplete_details: { reason: "max_output_tokens" },
+        };
+    }
+    return { status: "completed", incomplete_details: null };
+}
+/**
+ * Absent cache fields are omitted rather than zeroed (never conflate "not
+ * observed" with "observed as zero" — the field's own doc comment on
+ * `CodexResponseUsage` states this contract).
+ */
+export function synthesizeCodexUsage(usage) {
+    const cacheRead = usage.cache_read_input_tokens;
+    const cacheCreation = usage.cache_creation_input_tokens;
+    const inputTokens = usage.input_tokens + (cacheRead ?? 0) + (cacheCreation ?? 0);
+    const details = cacheRead === undefined && cacheCreation === undefined
+        ? undefined
+        : {
+            ...(cacheRead === undefined ? {} : { cached_tokens: cacheRead }),
+            ...(cacheCreation === undefined
+                ? {}
+                : { cache_write_tokens: cacheCreation }),
+        };
+    return {
+        input_tokens: inputTokens,
+        output_tokens: usage.output_tokens,
+        total_tokens: inputTokens + usage.output_tokens,
+        ...(details ? { input_tokens_details: details } : {}),
+    };
+}
+/**
+ * Shared by the streaming serializer's non-failure item close and the
+ * non-streaming path (`serializeCodexResponse`), so item shape is defined
+ * once. Always produces a "completed" item — the streaming failure path
+ * (an item closed "incomplete" mid-turn) builds its item directly, since it
+ * must preserve the raw accumulated text/argument bytes rather than a value
+ * reconstructed through this function.
+ */
+export function buildCodexResponseItem(block, itemId, toolKindByName = NO_TOOL_KINDS) {
+    if (block.type === "text") {
+        // The streaming path opens a message item only on a block's first delta,
+        // so an empty text block produces no item there; match it here.
+        if (block.text === "") {
+            return null;
+        }
+        return {
+            id: itemId,
+            type: "message",
+            role: "assistant",
+            status: "completed",
+            content: [{ type: "output_text", text: block.text, annotations: [] }],
+        };
+    }
+    if (block.type === "tool_use" &&
+        toolKindByName.get(block.name) === "custom") {
+        return {
+            id: itemId,
+            type: "custom_tool_call",
+            status: "completed",
+            call_id: codexCallIdFromToolUseId(block.id),
+            name: block.name,
+            input: unwrapCustomToolInput(block.input) ?? JSON.stringify(block.input),
+        };
+    }
+    if (block.type === "tool_use") {
+        return {
+            id: itemId,
+            type: "function_call",
+            status: "completed",
+            call_id: codexCallIdFromToolUseId(block.id),
+            name: block.name,
+            arguments: JSON.stringify(block.input),
+        };
+    }
+    // thinking / image / tool_result — not emitted; see the design doc's §5.
+    return null;
+}
+/** Non-streaming path: `§7` of the design — the document root IS a `CodexResponseEnvelope`. */
+export function serializeCodexResponse(result, requestModel, toolKindByName = NO_TOOL_KINDS) {
+    const items = [];
+    for (const block of result.content) {
+        const item = buildCodexResponseItem(block, generateCodexItemId(block.type !== "tool_use"
+            ? "message"
+            : toolKindByName.get(block.name) === "custom"
+                ? "custom_tool_call"
+                : "function_call"), toolKindByName);
+        if (item) {
+            items.push(item);
+        }
+    }
+    const { status, incomplete_details } = mapClaudeStopReasonToCodex(result.stop_reason);
+    return {
+        id: generateCodexResponseId(),
+        object: "response",
+        created_at: Math.floor(Date.now() / 1000),
+        status,
+        model: requestModel,
+        output: items,
+        usage: synthesizeCodexUsage(result.usage),
+        incomplete_details,
+        error: null,
+    };
+}
+/**
+ * Push-based streaming serializer. A genuinely incremental tool-call-argument
+ * path (`openToolCall`/`pushToolCallArgsDelta`/`closeToolCall`) forwards each
+ * raw `input_json_delta` chunk as it arrives rather than buffering the whole
+ * call before emitting anything — `partial_json` is not individually valid
+ * JSON, so a serializer that validated per-chunk could not do this.
+ */
+export class CodexResponsesStreamSerializer {
+    model;
+    toolKindByName;
+    state = "idle";
+    responseId = generateCodexResponseId();
+    outputIndex = 0;
+    openKind = null;
+    openItemId = "";
+    textAccum = "";
+    argsAccum = "";
+    toolCallId = "";
+    toolName = "";
+    closedItems = [];
+    terminalEnvelope = null;
+    constructor(model, toolKindByName = NO_TOOL_KINDS) {
+        this.model = model;
+        this.toolKindByName = toolKindByName;
+    }
+    *start() {
+        this.assertNotTerminal();
+        this.state = "streaming";
+        const response = {
+            id: this.responseId,
+            object: "response",
+            status: "in_progress",
+            created_at: Math.floor(Date.now() / 1000),
+            model: this.model,
+            output: [],
+            usage: null,
+            incomplete_details: null,
+            error: null,
+        };
+        yield formatSSE("response.created", { type: "response.created", response });
+        // Whether the real CLI tolerates a stream without in_progress was never
+        // bisected (test/fixtures/sse-bisection-findings.md), so the order the
+        // real backend uses is kept: created, then in_progress, then items.
+        yield formatSSE("response.in_progress", {
+            type: "response.in_progress",
+            response,
+        });
+    }
+    *pushDelta(text) {
+        this.assertNotTerminal();
+        if (this.openKind !== "message") {
+            yield* this.closeOpenItem(false);
+            yield* this.openMessage();
+        }
+        this.textAccum += text;
+        yield formatSSE("response.output_text.delta", {
+            type: "response.output_text.delta",
+            output_index: this.outputIndex,
+            item_id: this.openItemId,
+            content_index: 0,
+            delta: text,
+        });
+    }
+    /** Opens the tool-call item header at content_block_start, before any argument text
+     *  is known — mirrors what content_block_start already gives us (id, name). A tool
+     *  the request declared `custom` opens as a `custom_tool_call` item. */
+    *openToolCall(toolUseId, name) {
+        this.assertNotTerminal();
+        yield* this.closeOpenItem(false);
+        const kind = this.toolKindByName.get(name) === "custom"
+            ? "custom_tool_call"
+            : "function_call";
+        this.openKind = kind;
+        this.openItemId = generateCodexItemId(kind);
+        this.toolCallId = codexCallIdFromToolUseId(toolUseId);
+        this.toolName = name;
+        this.argsAccum = "";
+        const item = kind === "custom_tool_call"
+            ? {
+                id: this.openItemId,
+                type: kind,
+                status: "in_progress",
+                call_id: this.toolCallId,
+                name: this.toolName,
+                input: "",
+            }
+            : {
+                id: this.openItemId,
+                type: kind,
+                status: "in_progress",
+                call_id: this.toolCallId,
+                name: this.toolName,
+                arguments: "",
+            };
+        yield formatSSE("response.output_item.added", {
+            type: "response.output_item.added",
+            output_index: this.outputIndex,
+            item,
+        });
+    }
+    /** Forwards one raw partial_json fragment as-is — no JSON.parse here, because a lone
+     *  fragment is not required to be valid JSON on its own. A custom call's fragments
+     *  are the `{"input":"..."}` wrapper, not client text, so they are only buffered;
+     *  `closeToolCall` emits the unwrapped input once the whole wrapper is known. */
+    *pushToolCallArgsDelta(rawChunk) {
+        this.assertNotTerminal();
+        if (this.openKind !== "function_call" &&
+            this.openKind !== "custom_tool_call") {
+            throw new Error("pushToolCallArgsDelta with no open tool-call item");
+        }
+        this.argsAccum += rawChunk;
+        if (this.openKind === "custom_tool_call") {
+            return;
+        }
+        yield formatSSE("response.function_call_arguments.delta", {
+            type: "response.function_call_arguments.delta",
+            output_index: this.outputIndex,
+            item_id: this.openItemId,
+            delta: rawChunk,
+        });
+    }
+    /**
+     * Public close for a tool-use block at `content_block_stop`. A no-op when no
+     * tool-call item is open, so a driver can call it unconditionally.
+     * Validates the accumulated arguments once, here — the one place validity
+     * is required. Throws on malformed JSON; the caller must catch and route to
+     * `emitFailure` (never swallowed, never retried after this point).
+     */
+    *closeToolCall() {
+        this.assertNotTerminal();
+        if (this.openKind === "custom_tool_call") {
+            yield* this.closeCustomToolCall();
+            return;
+        }
+        if (this.openKind !== "function_call") {
+            return;
+        }
+        // A zero-argument call arrives with no input_json_delta at all. The value
+        // that was validated is the value emitted, so it becomes "{}" everywhere,
+        // and one delta carries it so the client sees the same shape as any call.
+        const args = this.argsAccum === "" ? "{}" : this.argsAccum;
+        JSON.parse(args); // throws on malformed JSON
+        if (this.argsAccum === "") {
+            yield formatSSE("response.function_call_arguments.delta", {
+                type: "response.function_call_arguments.delta",
+                output_index: this.outputIndex,
+                item_id: this.openItemId,
+                delta: args,
+            });
+        }
+        yield formatSSE("response.function_call_arguments.done", {
+            type: "response.function_call_arguments.done",
+            output_index: this.outputIndex,
+            item_id: this.openItemId,
+            arguments: args,
+        });
+        const item = {
+            id: this.openItemId,
+            type: "function_call",
+            status: "completed",
+            call_id: this.toolCallId,
+            name: this.toolName,
+            arguments: args,
+        };
+        yield formatSSE("response.output_item.done", {
+            type: "response.output_item.done",
+            output_index: this.outputIndex,
+            item,
+        });
+        this.closedItems.push(item);
+        this.openKind = null;
+        this.outputIndex++;
+    }
+    /**
+     * Public close for a text block at `content_block_stop`, mirroring
+     * `closeToolCall`'s role for the message item. A no-op when no message item
+     * is open, so a driver can call it unconditionally alongside `closeToolCall`
+     * without first inspecting which kind is currently open.
+     */
+    *closeTextBlock() {
+        this.assertNotTerminal();
+        if (this.openKind !== "message") {
+            return;
+        }
+        yield* this.closeMessageItem(false);
+    }
+    *finish(stopReason, usage) {
+        this.assertNotTerminal();
+        yield* this.closeOpenItem(false);
+        const { status, incomplete_details } = mapClaudeStopReasonToCodex(stopReason);
+        const envelope = {
+            id: this.responseId,
+            object: "response",
+            created_at: Math.floor(Date.now() / 1000),
+            status,
+            model: this.model,
+            output: this.closedItems,
+            usage: synthesizeCodexUsage(usage),
+            incomplete_details,
+            error: null,
+        };
+        this.state = "done";
+        this.terminalEnvelope = envelope;
+        yield formatSSE(status === "incomplete" ? "response.incomplete" : "response.completed", {
+            type: status === "incomplete"
+                ? "response.incomplete"
+                : "response.completed",
+            response: envelope,
+        });
+        return envelope;
+    }
+    /**
+     * Post-commit terminal path. Closes any open item as "incomplete" — now
+     * representable via `CodexResponseItemStatus` — then emits a single
+     * `response.failed`. Idempotent: a second call after this one already ran
+     * returns the same envelope without emitting further frames, so a driver
+     * that reaches it defensively from more than one place never double-emits.
+     */
+    // `status` is accepted (not folded into the wire envelope, which has no
+    // HTTP-status field) so a caller's classification — e.g. 502 upstream vs.
+    // 400 malformed-arguments — is available to whatever logs or maps this
+    // failure, mirroring `CodexFallbackStreamError`'s own `status` field on
+    // the reverse direction.
+    *emitFailure(_status, message, code) {
+        if (this.terminalEnvelope) {
+            return this.terminalEnvelope;
+        }
+        yield* this.closeOpenItem(true);
+        this.state = "error";
+        const envelope = {
+            id: this.responseId,
+            object: "response",
+            created_at: Math.floor(Date.now() / 1000),
+            status: "failed",
+            model: this.model,
+            output: this.closedItems,
+            incomplete_details: null,
+            error: { code: code ?? "upstream_error", message },
+        };
+        this.terminalEnvelope = envelope;
+        yield formatSSE("response.failed", {
+            type: "response.failed",
+            response: envelope,
+        });
+        return envelope;
+    }
+    assertNotTerminal() {
+        if (this.state === "done" || this.state === "error") {
+            throw new Error("CodexResponsesStreamSerializer: push after terminal event");
+        }
+    }
+    /**
+     * The custom-kind close: nothing was streamed live, so one input delta
+     * carries the whole unwrapped input, then its done event. Lenient where the
+     * function path throws: the wrapper is Claude's own output for the
+     * single-field schema `mapCodexCustomToolToClaude` declared, so a parse
+     * failure is an internal bug, and the raw accumulated text is emitted rather
+     * than failing the turn over it.
+     */
+    *closeCustomToolCall() {
+        const input = unwrapCustomToolInputJson(this.argsAccum) ?? this.argsAccum;
+        yield formatSSE("response.custom_tool_call_input.delta", {
+            type: "response.custom_tool_call_input.delta",
+            output_index: this.outputIndex,
+            item_id: this.openItemId,
+            delta: input,
+        });
+        yield formatSSE("response.custom_tool_call_input.done", {
+            type: "response.custom_tool_call_input.done",
+            output_index: this.outputIndex,
+            item_id: this.openItemId,
+            input,
+        });
+        const item = {
+            id: this.openItemId,
+            type: "custom_tool_call",
+            status: "completed",
+            call_id: this.toolCallId,
+            name: this.toolName,
+            input,
+        };
+        yield formatSSE("response.output_item.done", {
+            type: "response.output_item.done",
+            output_index: this.outputIndex,
+            item,
+        });
+        this.closedItems.push(item);
+        this.openKind = null;
+        this.outputIndex++;
+    }
+    *openMessage() {
+        this.openKind = "message";
+        this.openItemId = generateCodexItemId("message");
+        this.textAccum = "";
+        yield formatSSE("response.output_item.added", {
+            type: "response.output_item.added",
+            output_index: this.outputIndex,
+            item: {
+                id: this.openItemId,
+                type: "message",
+                role: "assistant",
+                status: "in_progress",
+                content: [],
+            },
+        });
+        yield formatSSE("response.content_part.added", {
+            type: "response.content_part.added",
+            output_index: this.outputIndex,
+            item_id: this.openItemId,
+            content_index: 0,
+            part: { type: "output_text", text: "", annotations: [] },
+        });
+    }
+    /** Closes whichever item is currently open, as "incomplete" when `failed`. */
+    *closeOpenItem(failed) {
+        if (this.openKind === "function_call") {
+            if (failed) {
+                yield* this.closeFailedFunctionCall();
+                return;
+            }
+            yield* this.closeToolCall();
+            return;
+        }
+        if (this.openKind === "custom_tool_call") {
+            if (failed) {
+                yield* this.closeFailedCustomToolCall();
+                return;
+            }
+            yield* this.closeToolCall();
+            return;
+        }
+        if (this.openKind === "message") {
+            yield* this.closeMessageItem(failed);
+        }
+    }
+    /**
+     * The failure-close counterpart to `closeToolCall`. Arguments are carried
+     * verbatim from `argsAccum` and never re-validated — `JSON.parse` is not
+     * called on the failure path, and no `function_call_arguments.done` is
+     * emitted, since the arguments never reached a validated, complete state.
+     */
+    *closeFailedFunctionCall() {
+        const item = {
+            id: this.openItemId,
+            type: "function_call",
+            status: "incomplete",
+            call_id: this.toolCallId,
+            name: this.toolName,
+            arguments: this.argsAccum,
+        };
+        yield formatSSE("response.output_item.done", {
+            type: "response.output_item.done",
+            output_index: this.outputIndex,
+            item,
+        });
+        this.closedItems.push(item);
+        this.openKind = null;
+        this.outputIndex++;
+    }
+    /**
+     * The failure close for a custom call. Its accumulated bytes are Anthropic's
+     * JSON wrapper, not the client's grammar text, so `input` is the unwrapped
+     * string when the wrapper already parses and "" otherwise — an incomplete
+     * item is never executed. No input done event, as on the function path.
+     */
+    *closeFailedCustomToolCall() {
+        const item = {
+            id: this.openItemId,
+            type: "custom_tool_call",
+            status: "incomplete",
+            call_id: this.toolCallId,
+            name: this.toolName,
+            input: unwrapCustomToolInputJson(this.argsAccum) ?? "",
+        };
+        yield formatSSE("response.output_item.done", {
+            type: "response.output_item.done",
+            output_index: this.outputIndex,
+            item,
+        });
+        this.closedItems.push(item);
+        this.openKind = null;
+        this.outputIndex++;
+    }
+    *closeMessageItem(failed) {
+        const text = this.textAccum;
+        yield formatSSE("response.output_text.done", {
+            type: "response.output_text.done",
+            output_index: this.outputIndex,
+            item_id: this.openItemId,
+            content_index: 0,
+            text,
+        });
+        yield formatSSE("response.content_part.done", {
+            type: "response.content_part.done",
+            output_index: this.outputIndex,
+            item_id: this.openItemId,
+            content_index: 0,
+            part: { type: "output_text", text, annotations: [] },
+        });
+        const item = {
+            id: this.openItemId,
+            type: "message",
+            role: "assistant",
+            status: failed ? "incomplete" : "completed",
+            content: [{ type: "output_text", text, annotations: [] }],
+        };
+        yield formatSSE("response.output_item.done", {
+            type: "response.output_item.done",
+            output_index: this.outputIndex,
+            item,
+        });
+        this.closedItems.push(item);
+        this.openKind = null;
+        this.textAccum = "";
+        this.outputIndex++;
+    }
+}

package/dist/proxy/codexToAnthropicFallback.d.ts

@@ -0,0 +1,54 @@
+/**
+ * Drives the Anthropic -> Codex response/stream codec against a real upstream
+ * `Response`. Mirrors `codexFallback.ts`, inverted: named after its own
+ * upstream, an Anthropic-shape engine, by the same convention `codexFallback.ts`
+ * is named after Codex.
+ *
+ * Scope: translation and the pre-commit/post-commit boundary only. Engine
+ * selection, OAuth/account selection, retry/cooldown policy and analytics
+ * recording are owned by sibling modules and are not this module's concern —
+ * it does not touch `src/lib/server/routes/codexProxyRoutes.ts`.
+ *
+ * SSE parsing note: this direction reads Anthropic-shaped SSE bytes, the same
+ * wire shape `preflightAnthropicStream` already buffers before this module
+ * ever sees the stream. Rather than duplicating `codexFallback.ts`'s
+ * carry/searchFrom scanner (built for the differently-shaped native Codex
+ * event vocabulary), this module reuses `extractSSEEvents` — the incremental
+ * Anthropic-SSE parser `preflightAnthropicStream` itself is built on — so the
+ * preflight-buffered chunks and the live remainder are parsed by the exact
+ * same code path. The carry/searchFrom scanner extraction into a shared
+ * `sseFrameScanner.ts` remains an explicitly-deferred follow-up either way.
+ */
+import type { CodexNativeToolKind, CodexResponseEnvelope, CodexResponseStream } from "../types/index.js";
+/** A non-2xx reply (either path) or a non-JSON reply (non-streaming path) from
+ *  the Anthropic-shape upstream, raised before any Codex frame is sent. */
+export declare class AnthropicFallbackResponseError extends Error {
+    readonly status: number;
+    readonly responseBody: string;
+    constructor(status: number, responseBody: string);
+}
+/** A pre-commit or post-commit failure while translating the Anthropic SSE
+ *  stream, or a well-formed non-streaming reply this module refuses to
+ *  translate (tool calls over `streamLimits.ts`'s count or size ceiling). */
+export declare class AnthropicFallbackStreamError extends Error {
+    readonly status: number;
+    readonly code: string;
+    constructor(status: number, message: string, code: string);
+}
+/**
+ * Consume a fully-buffered, non-streaming Anthropic-shape response and
+ * translate it into one Codex Responses envelope.
+ *
+ * No `message_start`/`message_delta` split here: a single JSON object
+ * carries the whole turn's usage at once (confirmed against this repo's own
+ * `settleFromResponseUsage`, which reads a non-streaming Anthropic body the
+ * same way), so the translation delegates straight to `serializeCodexResponse`.
+ */
+export declare function consumeAnthropicFallbackResponse(response: Response, model: string, toolKindByName?: ReadonlyMap<string, CodexNativeToolKind>): Promise<CodexResponseEnvelope>;
+/**
+ * Translate an Anthropic SSE stream into Codex-shape SSE frames as events
+ * arrive. The caller owns error framing downstream and must never retry
+ * after any frame has been emitted (mirrors `codexFallback.ts`'s own
+ * discipline, `:715-718`).
+ */
+export declare function createAnthropicFallbackStream(response: Response, model: string, toolKindByName?: ReadonlyMap<string, CodexNativeToolKind>): Promise<CodexResponseStream>;