Files
OmniRoute/open-sse/executors/codex-app-server.ts
Praveen K Palaniswamy 65e81158ab fix(ollama): route models by advertised capability (#11088)
Landed with the design call resolved per the owner's pick — **option 1**: the synced store is now endpoint-agnostic (persistDiscoveredModels and managedModelImport no longer drop non-chat models at write time), and chat selectability moved to read time (auto-pool expansion in autoStrategy applies filterChatSelectableModels; the models-route projection already had its chatOnly filter). Your discovery test now passes end-to-end (3/3): /api/show capabilities persist per connection and image/embedding requests route through the advertising host.

Reconciliation notes: conflicted areas merged onto the current tip (adobe discovery import, requestedModel preflight signature, resolvedProvider fast-path coexists with the synced-route override — explicit resolution wins); carried base-red drains (#10055 memoization, #11071 test variants) dropped as already-landed; the managed-model-import exclusion test was propagated to the new contract (image/video models persist; the read filter still hides them from chat pickers — pinned by a new assertion). Full battery: 205/206 focused (the one red is a confirmed periodic-timer timing flake on the loaded devbox — 20/20 isolated), autoCombo vitest 30/30, combo suites 46/46, gates + typecheck clean.

Thank you @yourspraveen — the capability probe + routing design was right; it just needed the store contract opened up. Fixes #11087.
2026-08-23 11:45:01 -03:00

449 lines
18 KiB
TypeScript

import {
bridgeToResponsesSSE,
buildResponseJSON,
} from "../vendor/codex-chatgpt-web/bridge.ts";
import { AsyncEventQueue } from "../vendor/codex-chatgpt-web/event-queue.ts";
import type { AdapterEvent } from "../vendor/codex-chatgpt-web/types.ts";
import { sanitizeErrorMessage } from "../utils/error.ts";
import { PROVIDERS } from "../config/constants.ts";
import { BaseExecutor, type ExecuteInput, type ExecutorExecuteResult } from "./base.ts";
import {
CodexAppServerClient,
type CodexAppServerClientOptions,
} from "./codex/appServerClient.ts";
import { resolveAppServerConfig, type CodexAppServerConfig } from "./codex/appServerConfig.ts";
import {
translateNotification,
translateToolCall,
type DynamicToolCallLike,
} from "./codex/appServerEvents.ts";
const JSON_HEADERS = { "Content-Type": "application/json" };
const SSE_HEADERS = {
"Cache-Control": "no-cache, no-transform",
Connection: "keep-alive",
"Content-Type": "text/event-stream; charset=utf-8",
};
/** A single text UserInput as accepted by turn/start (text_elements is required). */
interface CodexTextUserInput {
type: "text";
text: string;
text_elements: [];
}
/**
* Flatten an OpenAI Responses request body into the plain prompt text the
* app-server turn expects. The body's `input` is a string, a single message item,
* or an array of message items with `content` parts; we concatenate the user-facing
* text. This is intentionally lossless-enough for a text turn (images/tool parts are
* out of scope for the initial app-server transport).
*/
export function extractPromptText(body: unknown): string {
if (!body || typeof body !== "object") return "";
const input = (body as Record<string, unknown>).input;
if (typeof input === "string") return input;
if (input == null) return "";
const items = Array.isArray(input) ? input : [input];
const chunks: string[] = [];
for (const item of items) {
collectText(item, chunks);
}
return chunks.join("\n").trim();
}
function collectText(item: unknown, out: string[]): void {
if (typeof item === "string") {
if (item.length > 0) out.push(item);
return;
}
if (!item || typeof item !== "object") return;
const rec = item as Record<string, unknown>;
if (typeof rec.text === "string" && rec.text.length > 0) {
out.push(rec.text);
return;
}
const content = rec.content;
if (typeof content === "string") {
if (content.length > 0) out.push(content);
return;
}
if (Array.isArray(content)) {
for (const part of content) {
if (part && typeof part === "object") {
const text = (part as Record<string, unknown>).text;
if (typeof text === "string" && text.length > 0) out.push(text);
} else if (typeof part === "string" && part.length > 0) {
out.push(part);
}
}
}
}
/** Optional reasoning effort carried on the Responses body (`reasoning.effort`). */
function extractEffort(body: unknown): string | undefined {
if (!body || typeof body !== "object") return undefined;
const reasoning = (body as Record<string, unknown>).reasoning;
if (reasoning && typeof reasoning === "object") {
const effort = (reasoning as Record<string, unknown>).effort;
if (typeof effort === "string" && effort.length > 0) return effort;
}
return undefined;
}
/** A codex app-server DynamicToolSpec (experimental-api) advertised on thread/start. */
interface DynamicToolFunctionSpec {
type: "function";
name: string;
description: string;
inputSchema: Record<string, unknown>;
}
interface AppServerToolMaps {
/** wireName -> {namespace, name} for restoring MCP namespaced calls in the bridge. */
namespace: Map<string, { namespace: string; name: string }>;
/** wireNames the bridge must relay as custom_tool_call (freeform, e.g. apply_patch). */
freeform: Set<string>;
/** wireNames the bridge must relay as tool_search_call. */
toolSearch: Set<string>;
/** DynamicToolSpecs to advertise to codex on thread/start (experimental-api). */
specs: DynamicToolFunctionSpec[];
}
const EMPTY_OBJECT_SCHEMA: Record<string, unknown> = { type: "object", properties: {} };
const FREEFORM_INPUT_SCHEMA: Record<string, unknown> = {
type: "object",
properties: { input: { type: "string", description: "Raw tool input." } },
required: ["input"],
};
function asRecord(v: unknown): Record<string, unknown> | null {
return v && typeof v === "object" && !Array.isArray(v) ? (v as Record<string, unknown>) : null;
}
/**
* Build the bridge tool maps + the codex dynamicTools specs from the harness's
* Responses `tools` array. This mirrors chatgpt-web-codex.ts:toolMaps() /
* parser.ts:buildTools(): every harness tool is exposed to codex FLAT under its
* wire name ("<namespace>__<name>" for MCP tools) so the round-trip is
* namespace-preserving (codex echoes the call via item/tool/call; the bridge
* restores {namespace, name} from `toolNsMap`). Custom (freeform) and tool_search
* tools are tracked so the bridge relays them as custom_tool_call / tool_search_call.
*/
function buildAppServerToolMaps(body: unknown): AppServerToolMaps {
const namespace = new Map<string, { namespace: string; name: string }>();
const freeform = new Set<string>();
const toolSearch = new Set<string>();
const specs: DynamicToolFunctionSpec[] = [];
const rec = asRecord(body);
const tools = rec && Array.isArray(rec.tools) ? (rec.tools as unknown[]) : [];
const pushFn = (name: string, description: string, inputSchema: Record<string, unknown>) => {
specs.push({ type: "function", name, description, inputSchema });
};
for (const raw of tools) {
const t = asRecord(raw);
if (!t) continue;
const type = t.type;
const desc = typeof t.description === "string" ? t.description : "";
if (type === "function" && typeof t.name === "string") {
const wireName = t.name;
pushFn(wireName, desc, asRecord(t.parameters) ?? EMPTY_OBJECT_SCHEMA);
} else if (type === "namespace" && Array.isArray(t.tools) && typeof t.name === "string") {
const ns = t.name;
for (const innerRaw of t.tools as unknown[]) {
const inner = asRecord(innerRaw);
if (inner && inner.type === "function" && typeof inner.name === "string") {
const wireName = `${ns}__${inner.name}`;
namespace.set(wireName, { namespace: ns, name: inner.name });
const innerDesc = typeof inner.description === "string" ? inner.description : "";
pushFn(wireName, innerDesc, asRecord(inner.parameters) ?? EMPTY_OBJECT_SCHEMA);
}
}
} else if (type === "custom" && typeof t.name === "string") {
const wireName = t.name;
freeform.add(wireName);
pushFn(wireName, desc, FREEFORM_INPUT_SCHEMA);
} else if (type === "tool_search") {
const wireName = "tool_search";
toolSearch.add(wireName);
pushFn(
wireName,
desc || "Search for additional tools to load for the next turn.",
asRecord(t.parameters) ?? {
type: "object",
properties: { query: { type: "string" }, limit: { type: "number" } },
required: ["query"],
}
);
} else if (
typeof t.name === "string" &&
type !== "web_search" &&
type !== "image_generation" &&
type !== "web_search_preview"
) {
// Any other named, client-executed tool → pass through as a function so the
// routed model can call it; the bridge relays its call as a function_call.
pushFn(t.name, desc, asRecord(t.parameters) ?? EMPTY_OBJECT_SCHEMA);
}
// web_search / image_generation are OpenAI-hosted server-side tools — not relayable.
}
return { namespace, freeform, toolSearch, specs };
}
/**
* Executor for the Codex app-server WS transport. Drives one turn against a local
* `codex app-server` over JSON-RPC and re-emits its notifications as OpenAI
* Responses SSE via the shared bridge.
*
* Errors are delivered IN-BAND (an `error` AdapterEvent → `response.failed` SSE
* frame for streaming, or an error field in the JSON body for non-streaming),
* never thrown out of execute().
*/
export class CodexAppServerExecutor extends BaseExecutor {
private readonly clientOptions: CodexAppServerClientOptions;
/**
* @param clientOptions transport options (websocketFn, timeouts).
* @param providerId which provider identity this executor reports as. Defaults
* to "codex" so the existing per-connection `codexTransport==="app-server"`
* flag path (routed through CodexExecutor for the `codex` provider) keeps its
* original identity. The first-class `codex-app-server` sibling passes
* "codex-app-server" so logs/quota scoping and the golden executor map reflect
* the real provider. Falls back to PROVIDERS.codex when the sibling registry
* entry is not present (defensive; both share the codex backend).
*/
constructor(clientOptions: CodexAppServerClientOptions = {}, providerId = "codex") {
super(providerId, PROVIDERS[providerId] ?? PROVIDERS.codex);
this.clientOptions = clientOptions;
}
override async execute(input: ExecuteInput): Promise<ExecutorExecuteResult> {
const psd = input.credentials?.providerSpecificData;
const config = resolveAppServerConfig(psd);
if (!config) {
return errorResponse(
503,
"Codex app-server transport is not configured (missing url or token)",
"codex_app_server_unconfigured"
);
}
const promptText = extractPromptText(input.body);
const effort = extractEffort(input.body);
const toolMaps = buildAppServerToolMaps(input.body);
const hasTools = toolMaps.specs.length > 0;
const events = new AsyncEventQueue<AdapterEvent>();
const client = new CodexAppServerClient(this.clientOptions);
const run = async () => {
let terminated = false;
// Resolves when the turn reaches a terminal state (turn/completed, error,
// or an item/tool/call passthrough). `turn/start` resolving only means the
// turn was ACCEPTED (status: inProgress) — the model's output arrives later
// as notifications. run() MUST await this before the finally-block closes
// the client, otherwise the socket is torn down mid-turn and the event
// queue never receives its terminal event (the request then hangs until the
// caller's timeout). See translateNotification: it returns true on the
// terminal notification, which is where we settle this.
let settleTurn!: () => void;
const turnDone = new Promise<void>((resolve) => {
settleTurn = resolve;
});
const markTerminated = () => {
if (terminated) return;
terminated = true;
settleTurn();
};
const finishTurn = () => {
if (terminated) return;
events.push({ type: "done", endTurn: true });
events.close();
markTerminated();
};
try {
await client.connect(config.url, config.token);
await client.request("initialize", {
clientInfo: {
name: "omniroute-codex-app-server",
title: null,
version: "1.0",
},
// Harness function tools are advertised via thread/start's `dynamicTools`,
// which is an EXPERIMENTAL app-server field: opt into experimental API so
// codex accepts it (and can emit the item/tool/call ServerRequest).
capabilities: hasTools
? { experimentalApi: true, requestAttestation: false }
: null,
});
const threadResult = (await client.request("thread/start", {
cwd: config.cwd,
// OmniRoute is a router: the HARNESS that consumes OmniRoute owns tool
// execution and policy. codex must therefore NEVER block a turn waiting
// on its own interactive approval, and its own sandbox must not gate the
// model — the harness decides what actually runs. So we pair
// approvalPolicy:"never" (non-interactive; codex never prompts) with
// sandbox:"danger-full-access" (codex's own sandbox imposes no
// restriction), mirroring codexInstructions.ts:50 ("never +
// danger-full-access = take advantage of it"). Any server→client
// approval request that still arrives is auto-APPROVED by the client
// (see CodexAppServerClient), never denied — denial would sabotage the
// harness's tool calls. Callers can override both via providerSpecificData.
approvalPolicy: config.approvalPolicy ?? "never",
sandbox: config.sandbox ?? "danger-full-access",
// INBOUND harness tools → codex. The client tells the app-server which
// function tools are available for the thread via the `dynamicTools`
// field on thread/start (a DynamicToolSpec[] under the experimental API,
// verified from the real codex binary; see appServerEvents.ts). codex
// then invokes them by sending the `item/tool/call` ServerRequest back
// to the client (DynamicToolCallParams), which we PASS THROUGH.
...(hasTools ? { dynamicTools: toolMaps.specs } : {}),
})) as { thread?: { id?: unknown }; threadId?: unknown };
// The live app-server (codex 0.149.0) returns the thread under
// result.thread.id — NOT a top-level threadId (verified against the real
// binary 2026-08-22). Keep the top-level fallback for forward/back compat.
const threadId =
threadResult && typeof threadResult.thread?.id === "string"
? threadResult.thread.id
: threadResult && typeof threadResult.threadId === "string"
? threadResult.threadId
: "";
client.onNotification((method, params) => {
if (terminated) return;
const isTerminal = translateNotification(method, params, (event) => events.push(event));
if (isTerminal) {
events.close();
markTerminated();
}
});
// OUTBOUND codex tool call → harness. codex asks us to execute a harness
// tool via the `item/tool/call` ServerRequest. OmniRoute is a STATELESS
// ROUTER and CANNOT execute the harness's tool (the tool body lives in the
// harness downstream). So we PASS IT THROUGH: emit tool_call_* AdapterEvents
// (the bridge renders a Responses function_call / custom_tool_call /
// tool_search_call), settle the app-server request with a benign
// DynamicToolCallResponse so codex does not hang, and COMPLETE the turn.
// The harness runs the tool and replays the result in a fresh /v1/responses
// request (the stateless-full-history contract every OmniRoute provider uses).
client.onToolCall((_id, params, api) => {
if (terminated) return;
const toolParams = (params && typeof params === "object" ? params : {}) as DynamicToolCallLike;
translateToolCall(toolParams, (event) => events.push(event));
// Settle the app-server request so the socket does not stall. The router
// does not have the tool output (the harness will produce it next turn),
// so we report the passthrough as an unsuccessful in-line result and end
// the turn — the function_call has already been surfaced to the harness.
api.respond({
contentItems: [
{
type: "inputText",
text: "router: tool executed by harness; call surfaced as function_call",
},
],
success: false,
});
finishTurn();
});
const onAbort = () => {
try {
client.notify("turn/interrupt", { threadId, turnId: "" });
} catch {
/* interrupt best-effort */
}
// Unblock run() so the finally-block can tear down the client. Without
// this, an aborted request would wait on turnDone until the terminal
// notification that will never come.
if (!terminated) {
events.close();
markTerminated();
}
};
input.signal?.addEventListener("abort", onAbort, { once: true });
const turnInput: CodexTextUserInput[] = [
{ type: "text", text: promptText, text_elements: [] },
];
await client.request("turn/start", {
threadId,
input: turnInput,
model: input.model,
...(effort ? { effort } : {}),
});
// `turn/start` resolving only ACCEPTS the turn (status: inProgress). The
// model's output (agentMessage deltas) and the terminal turn/completed
// arrive AFTER, as notifications. Wait for the terminal signal before
// falling through to the finally-block — otherwise client.close() tears
// down the socket mid-turn and the queue never closes (request hangs).
await turnDone;
} catch (err) {
if (!terminated) {
events.push({
type: "error",
message: sanitizeErrorMessage(err instanceof Error ? err.message : err),
status: 502,
errorType: "provider_error",
code: "codex_app_server_turn_failed",
});
events.close();
markTerminated();
}
} finally {
client.close();
}
};
if (!input.stream) {
const running = run();
const collected = await events.collect();
await running;
const response = buildResponseJSON(collected, input.model, {
toolNsMap: toolMaps.namespace,
freeformToolNames: toolMaps.freeform,
toolSearchToolNames: toolMaps.toolSearch,
});
return {
response: new Response(JSON.stringify(response), { status: 200, headers: JSON_HEADERS }),
url: config.url,
};
}
void run();
const stream = bridgeToResponsesSSE(
events,
input.model,
toolMaps.namespace,
toolMaps.freeform,
toolMaps.toolSearch,
() => client.close(),
2_000
);
return {
response: new Response(stream, { status: 200, headers: SSE_HEADERS }),
url: config.url,
};
}
}
function errorResponse(status: number, message: string, code: string): Response {
return new Response(
JSON.stringify({
error: {
code,
message: sanitizeErrorMessage(message),
type: status >= 500 ? "provider_error" : "invalid_request_error",
},
}),
{ status, headers: JSON_HEADERS }
);
}
// re-export config type for consumers/tests
export type { CodexAppServerConfig };