agent-accelerator 0.3.0 → 0.4.5

package/src/providers/openrouter.ts

@@ -10,7 +10,7 @@
  * `references/testings/openrouter/CAPABILITY-MATRIX.md`.
  *
  * This is the STABLE transport. The beta Responses skin
- * (`src/providers/openrouter-responses.ts`) is discontinued: no video shape,
+ * was removed from core: no video shape,
  * lossy error codes, beta event-vocabulary drift.
  *
  * Notes:
@@ -40,13 +40,26 @@ import { AssistantMessageEventStream } from "../streaming/event-stream.ts";
 import { SSEParser } from "../streaming/sse-parser.ts";
 import { AgentResponse } from "../types/response.ts";
 import { getApiKey, getEnv } from "../utils/env.ts";
-import { buildSessionHeaders } from "../utils/headers.ts";
+import { buildSessionHeaders, setHeaderCaseInsensitive } from "../utils/headers.ts";
 import { normalizeMediaInput } from "../utils/media.ts";
 import { safeStringify } from "../utils/serialization.ts";
 import { toConciseProviderError, assertModalitiesSupported } from "../utils/errors.ts";
 import { withRetries } from "../utils/retry.ts";
+import {
+  parseArguments,
+  normalizeThinkingParts,
+  newToolCallId,
+  redactedHeaders,
+  readErrorPayload,
+  incompleteStreamError,
+  hasTruncatedToolArguments,
+  combinedSignal,
+  timeoutFor,
+  hasAuthHeader,
+  clampToolCallId,
+} from "./shared.ts";
 import { createGenericModelSpec } from "../models/catalog.ts";
-import { getModelFromCatalog, getModelsForProvider } from "../models/catalog.ts";
+import { getModelFromCatalog, getModelsForProvider, modelSupportsStructuredOutput } from "../models/catalog.ts";
 import {
   mapThinkingLevelToOpenRouterChat,
   mapServiceTierToOpenRouter,
@@ -54,6 +67,10 @@ import {
   mapToolChoiceToOpenRouterChat,
   noteProviderTurn,
   parseStreamedToolArguments,
+  mapOutputToOpenRouterChat,
+  normalizeStructuredOutput,
+  augmentSystemPromptWithStructuredOutput,
+  emitProviderWarning,
 } from "../providers.ts";
 
 // ---------------------------------------------------------------------------
@@ -92,6 +109,7 @@ interface ChatRequestBody {
   service_tier?: string;
   session_id?: string;
   plugins?: Array<{ id: string }>;
+  response_format?: { type: "json_schema"; json_schema: { name: string; schema: Record<string, unknown>; strict: boolean; description?: string } };
   stream?: boolean;
   [k: string]: unknown;
 }
@@ -156,6 +174,12 @@ interface ChatResponse {
 
 const DEFAULT_BASE_URL = "https://openrouter.ai/api/v1";
 
+/** Q-18: Chat transport never echoes thought signatures (no wire field).
+ * Local helper documents the namespacing decision per provider file. */
+function withProviderSignature(_sig: string | undefined): { thoughtSignature?: undefined } {
+  return {};
+}
+
 /** Internal headers that must never leak onto native REST requests. */
 const INTERNAL_HEADERS = new Set([
   "x-thought-signature-map",
@@ -177,7 +201,10 @@ function resolveBaseUrl(options?: ProviderRequestOptions): string {
 }
 
 function resolveApiKey(options?: ProviderRequestOptions): string | undefined {
-  return options?.apiKey || getApiKey("openrouter", undefined, options?.env);
+  const key = getApiKey("openrouter", options?.apiKey, options?.env);
+  if (key) return key;
+  if (hasAuthHeader(options?.headers)) return "unused";
+  return undefined;
 }
 
 /**
@@ -336,11 +363,17 @@ async function fullHistoryMessages(context: ProviderContext): Promise<{
     messages.push({ role: "system", content: context.systemPrompt });
   }
   let hasFiles = false;
+  let lastRole: string | null = "system";
   for (const m of context.messages) {
     if (m.role === "system") continue;
     if (m.role === "user") {
+      // Strict routes reject `user` directly after `tool`: bridge with a
+      // synthetic assistant turn (pi api/openai-completions.ts:1233).
+      if (lastRole === "tool") {
+        messages.push({ role: "assistant", content: "I have processed the tool results." });
+      }
       if (typeof m.content === "string") {
-        if (m.content) messages.push({ role: "user", content: m.content });
+        if (m.content) { messages.push({ role: "user", content: m.content }); lastRole = "user"; }
       } else {
         const { blocks, fileUrls, hasFiles: hf } = await contentPartsToBlocks(m.content);
         if (hf) hasFiles = true;
@@ -348,7 +381,7 @@ async function fullHistoryMessages(context: ProviderContext): Promise<{
         for (const url of fileUrls) {
           blocks.push({ type: "text", text: url });
         }
-        if (blocks.length > 0) messages.push({ role: "user", content: blocks });
+        if (blocks.length > 0) { messages.push({ role: "user", content: blocks }); lastRole = "user"; }
       }
     } else if (m.role === "assistant") {
       if (typeof m.content === "string") {
@@ -362,7 +395,7 @@ async function fullHistoryMessages(context: ProviderContext): Promise<{
       for (const part of m.content) {
         if (part.type === "tool_call") {
           calls.push({
-            id: part.callId || part.id,
+            id: clampToolCallId(part.callId || part.id),
             type: "function",
             function: { name: part.name, arguments: JSON.stringify(part.arguments || {}) },
           });
@@ -372,20 +405,25 @@ async function fullHistoryMessages(context: ProviderContext): Promise<{
         // Prior-turn reasoning is NOT resent: chat history is messages +
         // tool calls only (documented decision; avoids strict-route 400s).
       }
+      // Skip empty assistant turns (no text + no tool calls): strict
+      // routes reject them, and they carry no signal (pi drop rule).
+      if (texts.length === 0 && calls.length === 0) continue;
       messages.push({
         role: "assistant",
         content: texts.length > 0 ? texts.join("\n") : null,
         ...(calls.length > 0 ? { tool_calls: calls } : {}),
       });
+      lastRole = "assistant";
     } else if (m.role === "tool") {
       if (typeof m.content === "string") {
         const queue = (m.name && idsByName.get(m.name)) || [];
         messages.push({
           role: "tool",
-          tool_call_id: queue.shift() || "call_0",
+          tool_call_id: clampToolCallId(queue.shift() || newToolCallId()),
           content: m.content,
           name: m.name,
         });
+        lastRole = "tool";
         continue;
       }
       if (!Array.isArray(m.content)) continue;
@@ -393,10 +431,11 @@ async function fullHistoryMessages(context: ProviderContext): Promise<{
         if (part.type === "tool_result") {
           messages.push({
             role: "tool",
-            tool_call_id: pairing.get(part.id) || part.id,
+            tool_call_id: clampToolCallId(pairing.get(part.id) || part.id),
             content: resultToOutput(part.result),
             name: part.name,
           });
+          lastRole = "tool";
         }
       }
     }
@@ -429,27 +468,6 @@ function mapUsage(raw?: ChatResponse["usage"]): TokenUsage {
   return usage;
 }
 
-function parseArguments(raw: string | undefined): Record<string, unknown> {
-  if (!raw) return {};
-  try {
-    const parsed: unknown = JSON.parse(raw);
-    if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
-      return parsed as Record<string, unknown>;
-    }
-    return { raw };
-  } catch {
-    return { raw };
-  }
-}
-
-/** Trims/collapses assembled thinking parts (no edge-tripling). */
-function normalizeThinkingParts(parts: string[]): string | undefined {
-  const cleaned = parts
-    .map((p) => p.replace(/\n{3,}/g, "\n\n").trim())
-    .filter((p) => p.length > 0);
-  return cleaned.length > 0 ? cleaned.join("\n") : undefined;
-}
-
 function throwResponseError(response: ChatResponse, modelId: string): void {
   const message =
     response.error && typeof response.error.message === "string" && response.error.message
@@ -476,7 +494,9 @@ function parseResponse(
 
   const choice = response.choices?.[0];
   const message = choice?.message;
-  const text = typeof message?.content === "string" ? message.content : "";
+  // Q-23: refusal content surfaces as text with a `refusal` finish reason.
+  const refusal = typeof message?.refusal === "string" && message.refusal ? message.refusal : "";
+  const text = typeof message?.content === "string" && message.content ? message.content : refusal;
   // Thinking: `reasoning_details[].text` duplicates `message.reasoning` when
   // both ride along (observed live) — prefer details, else the plain field.
   const thinkingParts: string[] = [];
@@ -493,7 +513,7 @@ function parseResponse(
   for (const tc of message?.tool_calls ?? []) {
     const args = parseArguments(tc.function?.arguments);
     toolCalls.push({
-      id: tc.id || `call_${Math.random().toString(36).slice(2, 9)}`,
+      id: tc.id && !tc.id.includes("|") && tc.id.length <= 40 ? tc.id : clampToolCallId(tc.id || newToolCallId()),
       name: tc.function?.name || "unknown",
       arguments: args,
       rawArguments:
@@ -504,15 +524,17 @@ function parseResponse(
   }
 
   // Normalized finish reasons: tool_calls|stop|length|content_filter|error.
+  // Q-23: refusal content forces `refusal` (even when the wire says stop).
   const wireReason = choice?.finish_reason ?? undefined;
+  const finishReason =
+    toolCalls.length > 0 ? "tool_calls" : refusal ? "refusal" : wireReason || "stop";
   return {
     text,
     thinking: normalizeThinkingParts(thinkingParts),
     thoughtSignature: undefined,
     toolCalls: toolCalls.length > 0 ? toolCalls : undefined,
     usage: mapUsage(response.usage),
-    finishReason:
-      toolCalls.length > 0 ? "tool_calls" : wireReason || "stop",
+    finishReason,
     responseId: response.id,
     model: modelId,
     provider: "openrouter",
@@ -521,33 +543,6 @@ function parseResponse(
   };
 }
 
-function redactedHeaders(headers: Record<string, string>): Record<string, string> {
-  const out: Record<string, string> = {};
-  for (const [k, v] of Object.entries(headers)) {
-    out[k] = k.toLowerCase() === "authorization" ? "[REDACTED]" : v;
-  }
-  return out;
-}
-
-function readErrorPayload(bodyText: string): { message: string; code?: string | number; errorType?: string } {
-  try {
-    const parsed: unknown = JSON.parse(bodyText);
-    const first = Array.isArray(parsed) ? parsed[0] : parsed;
-    const err = (first as { error?: { message?: string; code?: string | number; metadata?: { error_type?: string } } })?.error;
-    if (err && typeof err.message === "string") {
-      const errorType = (err as { metadata?: { error_type?: string } })?.metadata?.error_type;
-      return {
-        message: err.message,
-        code: err.code,
-        ...(typeof errorType === "string" ? { errorType } : {}),
-      };
-    }
-    return { message: bodyText.slice(0, 300) };
-  } catch {
-    return { message: bodyText.slice(0, 300) };
-  }
-}
-
 // ---------------------------------------------------------------------------
 // Provider
 // ---------------------------------------------------------------------------
@@ -587,7 +582,33 @@ export class OpenRouterChatCompletionsProvider implements Provider {
     const serviceTier = mapServiceTierToOpenRouter(options?.serviceTier);
     applyCacheForOpenRouter(options?.cache, `openrouter/${modelId}`);
 
-    const { messages, hasFiles } = await fullHistoryMessages(context);
+    // Structured outputs: native `response_format` when supported, else
+    // emulated via system-prompt instruction + client validation (catalog
+    // `structured_output: false` models would 400 on the wire shape).
+    let effectiveContext = context;
+    let outputEmulated = false;
+    let outputSpec: import("../types/model.ts").StructuredOutputSpec | undefined;
+    try {
+      outputSpec = normalizeStructuredOutput(options?.output ?? (options as { outputSchema?: unknown })?.outputSchema);
+    } catch {
+      // why: invalid output shapes fail fast at Agent construction/loop validation.
+      outputSpec = undefined;
+    }
+    if (outputSpec && !modelSupportsStructuredOutput("openrouter", modelId)) {
+      outputEmulated = true;
+      effectiveContext = {
+        ...context,
+        systemPrompt: augmentSystemPromptWithStructuredOutput(context.systemPrompt, outputSpec),
+      };
+      emitProviderWarning({
+        provider: "openrouter",
+        capability: "structured output",
+        requested: `${outputSpec.name} (openrouter/${modelId})`,
+        reason: "the catalog reports structured_output: false for this model, so native response_format is unsupported.",
+        fallback: "emulated JSON instruction + client validation (no response_format is sent)",
+      });
+    }
+    const { messages, hasFiles } = await fullHistoryMessages(effectiveContext);
     const body: ChatRequestBody = {
       model: modelId,
       messages,
@@ -601,12 +622,23 @@ export class OpenRouterChatCompletionsProvider implements Provider {
     if (toolChoice !== undefined) body.tool_choice = toolChoice;
     if (effort) body.reasoning = { effort };
     if (serviceTier) body.service_tier = serviceTier;
+    // Structured outputs: Chat Completions `response_format{type:json_schema,json_schema}`.
+    // Skipped when emulated (unsupported model): instruction already in messages.
+    if (outputSpec && !outputEmulated) {
+      try {
+        const mapped = mapOutputToOpenRouterChat(outputSpec);
+        if (mapped) body.response_format = mapped.response_format;
+      } catch {
+        // why: invalid output shapes fail fast at Agent construction/loop validation.
+      }
+    }
     // PDF ingestion is a documented plugin, not a message shape: enable it
     // only when file parts are present (never by default).
     if (hasFiles) body.plugins = [{ id: "file-parser" }];
-    // No temperature/top_p/max_tokens/stop/response_format/user/seed knobs
+    // No temperature/top_p/max_tokens/stop/user/seed knobs
     // (no canonical options exist; omitting also keeps provider cache keys
-    // stable). No store/previous_response_id/conversation: they do not exist
+    // stable). `response_format` above is the sole structured-output exception.
+    // No store/previous_response_id/conversation: they do not exist
     // on this endpoint.
     // Session affinity: top-level `session_id` is the documented sticky-routing
     // key on Chat Completions (prompt-caching guide) and takes precedence over
@@ -635,8 +667,8 @@ export class OpenRouterChatCompletionsProvider implements Provider {
     for (const [k, v] of Object.entries(headers)) {
       if (INTERNAL_HEADERS.has(k.toLowerCase())) delete headers[k];
     }
-    headers["Content-Type"] = "application/json";
-    headers["Authorization"] = `Bearer ${apiKey}`;
+    setHeaderCaseInsensitive(headers, "Content-Type", "application/json");
+    if (apiKey !== "unused") setHeaderCaseInsensitive(headers, "Authorization", `Bearer ${apiKey}`);
     return { method: "POST", headers, body: JSON.stringify(body), signal };
   }
 
@@ -648,13 +680,18 @@ export class OpenRouterChatCompletionsProvider implements Provider {
     sessionId: string | undefined,
     signal?: AbortSignal
   ): Promise<{ status: number; statusText: string; headers: Record<string, string>; text: string }> {
-    const res = await fetch(url, this.requestInit(body, options, apiKey, sessionId, signal));
-    const text = await res.text();
-    const headers: Record<string, string> = {};
-    res.headers.forEach((v, k) => {
-      headers[k] = v;
-    });
-    return { status: res.status, statusText: res.statusText, headers, text };
+    const { signal: effective, cleanup } = combinedSignal(signal, timeoutFor(options as { timeoutMs?: number } | undefined));
+    try {
+      const res = await fetch(url, this.requestInit(body, options, apiKey, sessionId, effective));
+      const text = await res.text();
+      const headers: Record<string, string> = {};
+      res.headers.forEach((v, k) => {
+        headers[k] = v;
+      });
+      return { status: res.status, statusText: res.statusText, headers, text };
+    } finally {
+      cleanup();
+    }
   }
 
   private throwIfError(
@@ -664,14 +701,21 @@ export class OpenRouterChatCompletionsProvider implements Provider {
     modelId: string
   ): void {
     if (status >= 200 && status < 300) return;
-    const { message, code, errorType } = readErrorPayload(bodyText);
-    const err: Record<string, unknown> & Error = new Error(message) as Record<string, unknown> & Error;
+    const { message, code, errorType, raw } = readErrorPayload(bodyText);
+    const hint =
+      status === 401
+        ? " Check OPENROUTER_API_KEY (no quotes/spaces, correct provider) and that the key is loaded (bun --env-file=.env.local)."
+        : "";
+    const fullMessage = `${message}${hint}`;
+    const err: Record<string, unknown> & Error = new Error(fullMessage) as Record<string, unknown> & Error;
     (err as Record<string, unknown>)["statusCode"] = status;
     (err as Record<string, unknown>)["status"] = status;
     (err as Record<string, unknown>)["responseBody"] = bodyText.slice(0, 500);
     (err as Record<string, unknown>)["url"] = url.split("?")[0];
+    (err as Record<string, unknown>)["data"] = { error: { message: fullMessage } };
     if (code !== undefined) (err as Record<string, unknown>)["code"] = code;
     if (errorType) (err as Record<string, unknown>)["errorType"] = errorType;
+    if (raw) (err as Record<string, unknown>)["raw"] = raw;
     throw toConciseProviderError(err, "openrouter", modelId);
   }
 
@@ -814,7 +858,21 @@ export class OpenRouterChatCompletionsProvider implements Provider {
           body,
         };
 
-        const res = await fetch(url, this.requestInit(body, options, apiKey, sessionId, linked.signal));
+        const streamTimeoutMs = timeoutFor(options as { timeoutMs?: number } | undefined);
+        const streamTimeoutTimer =
+          streamTimeoutMs && streamTimeoutMs > 0
+            ? setTimeout(() => {
+                try {
+                  linked.abort();
+                } catch {}
+              }, streamTimeoutMs)
+            : undefined;
+        let res: Response;
+        try {
+          res = await fetch(url, this.requestInit(body, options, apiKey, sessionId, linked.signal));
+        } finally {
+          if (streamTimeoutTimer) clearTimeout(streamTimeoutTimer);
+        }
         if (!res.ok || !res.body) {
           const text = !res.ok ? await res.text().catch(() => "") : "";
           if (!res.ok) this.throwIfError(res.status, url, text, clean);
@@ -824,7 +882,7 @@ export class OpenRouterChatCompletionsProvider implements Provider {
         res.headers.forEach((v, k) => {
           responseHeaders[k] = v;
         });
-        const responseMeta = { status: res.status, statusText: res.statusText, headers: responseHeaders };
+        const responseMeta: Record<string, unknown> & { status: number; statusText: string; headers: Record<string, string> } = { status: res.status, statusText: res.statusText, headers: responseHeaders };
         eventStream.push({ type: "start", raw: { request: rawRequest } } as never);
 
         const parser = new SSEParser();
@@ -839,6 +897,11 @@ export class OpenRouterChatCompletionsProvider implements Provider {
         let lastToolIndex = 0;
         let usage: TokenUsage = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
         let finishReason = "stop";
+        let sawFinishReason = false;
+        let sawUsage = false;
+        let refusalSeen = false;
+        let badFrames = 0;
+        let framesSeen = 0;
         let responseId: string | undefined;
         let completedBody: unknown = undefined;
         let aborted = false;
@@ -854,11 +917,17 @@ export class OpenRouterChatCompletionsProvider implements Provider {
         );
 
         const handleMessage = (data: string): void => {
-          if (!data || data === "[DONE]") return;
+          if (!data) return;
+          if (data === "[DONE]") {
+            sawFinishReason = sawFinishReason || completedBody !== undefined || sawUsage;
+            return;
+          }
           let msg: Record<string, unknown>;
           try {
             msg = JSON.parse(data) as Record<string, unknown>;
+            framesSeen += 1;
           } catch {
+            badFrames += 1;
             return;
           }
           // Mid-stream provider error: top-level `error`, HTTP stays 200.
@@ -881,6 +950,13 @@ export class OpenRouterChatCompletionsProvider implements Provider {
           const choice = (Array.isArray(msg["choices"]) ? (msg["choices"] as Record<string, unknown>[])[0] : undefined) ?? {};
           const delta = (choice["delta"] as Record<string, unknown>) ?? {};
           // Text delta (content-free accounting frames carry "" — ignored).
+          // Q-23: refusal deltas surface as text with a `refusal` finish.
+          const deltaRefusal = delta["refusal"];
+          if (typeof deltaRefusal === "string" && deltaRefusal) {
+            text += deltaRefusal;
+            refusalSeen = true;
+            eventStream.push({ type: "text_delta", delta: deltaRefusal, partialText: text });
+          }
           if (typeof delta["content"] === "string" && delta["content"]) {
             text += delta["content"] as string;
             eventStream.push({ type: "text_delta", delta: delta["content"] as string, partialText: text });
@@ -950,10 +1026,12 @@ export class OpenRouterChatCompletionsProvider implements Provider {
               throw toConciseProviderError(failure, "openrouter", clean);
             }
             finishReason = fr;
+            sawFinishReason = true;
           }
           const chunkUsage = msg["usage"] as ChatResponse["usage"] | undefined;
           if (chunkUsage) {
             usage = mapUsage(chunkUsage);
+            sawUsage = true;
             completedBody = completedBody ?? msg;
             eventStream.push({ type: "usage", usage });
           }
@@ -984,7 +1062,7 @@ export class OpenRouterChatCompletionsProvider implements Provider {
         for (const c of calls.values()) {
           const args = parseStreamedToolArguments(c.startArgs, c.deltaArgs);
           const record: ToolCallRecord = {
-            id: c.id || `call_${Math.random().toString(36).slice(2, 9)}`,
+            id: c.id && !c.id.includes("|") && c.id.length <= 40 ? c.id : clampToolCallId(c.id || newToolCallId()),
             name: c.name,
             arguments: args,
             rawArguments: c.startArgs + c.deltaArgs,
@@ -994,9 +1072,37 @@ export class OpenRouterChatCompletionsProvider implements Provider {
         }
 
         noteProviderTurn(sessionId, "openrouter");
-        if (toolCalls.length > 0) finishReason = "tool_calls";
+        if (toolCalls.length > 0) {
+          finishReason = "tool_calls";
+          sawFinishReason = true;
+        } else if (refusalSeen && finishReason !== "content_filter") {
+          finishReason = "refusal";
+          sawFinishReason = true;
+        }
 
-        const cleanThinking = thinking.replace(/\n{3,}/g, "\n\n").trim();
+        // Q-23: terminal-state guard — fail `incomplete_stream` (retryable)
+        // instead of resolving empty success. Skipped for non-SSE payloads
+        // (JSON mocks) where no SSE frames were ever seen.
+        const hasTerminal = completedBody !== undefined || sawFinishReason || sawUsage;
+        const cleanThinkingEarly = thinking.replace(/\n{3,}/g, "\n\n").trim();
+        const ctEntry = Object.entries(responseHeaders).find(([k]) => k.toLowerCase() === "content-type");
+        const isSSEPayload =
+          (typeof ctEntry?.[1] === "string" && ((ctEntry[1] as string).includes("event-stream") || (ctEntry[1] as string).startsWith("text/"))) ||
+          framesSeen > 0 ||
+          badFrames > 0;
+        if (!hasTerminal && !text && !cleanThinkingEarly && toolCalls.length === 0 && isSSEPayload) {
+          eventStream.fail(incompleteStreamError("openrouter", clean));
+          return;
+        }
+        // Refuse unfinished tool calls whose args never parsed (output_item.done
+        // never arrived): truncated `{raw}` must not reach the executor (pi
+        // api/openai-responses-shared.ts:766 rule). Non-SSE mocks preserve legacy.
+        if (!hasTerminal && toolCalls.some((c) => hasTruncatedToolArguments(c.arguments)) && isSSEPayload) {
+          eventStream.fail(incompleteStreamError("openrouter", clean));
+          return;
+        }
+
+        const cleanThinking = cleanThinkingEarly;
         const finalResponse = new AgentResponse({
           text,
           thinking: cleanThinking || undefined,
@@ -1007,7 +1113,10 @@ export class OpenRouterChatCompletionsProvider implements Provider {
           responseId,
           model: clean,
           provider: "openrouter",
-          raw: { request: rawRequest, response: { ...responseMeta, body: completedBody } },
+          raw: {
+            request: rawRequest,
+            response: { ...responseMeta, body: completedBody, badFrames } as unknown as ProviderRawData["response"],
+          },
           durationMs: Date.now() - startTime,
         });
         eventStream.push({ type: "done", delta: "", usage, finishReason, responseId });

package/src/providers/registry.ts

@@ -80,12 +80,25 @@ export function ensureCustomProvider(
 ): Provider {
   assertProviderNotRemoved(prefix);
   const id = normalizeProviderPrefix(prefix) as ProviderId;
+  if (FIRST_CLASS.has(id.toLowerCase())) {
+    if (opts?.baseUrl !== undefined || opts?.apiKey !== undefined) {
+      throw new Error(
+        `[Agent Accelerator] Provider "${id}" is first-class and cannot be replaced with custom baseUrl/apiKey. Pass baseUrl/apiKey per-Agent instead.`
+      );
+    }
+    return providerRegistry.get(id)!;
+  }
   const existing = providerRegistry.get(id);
-  if (existing && !FIRST_CLASS.has(id.toLowerCase())) return existing;
-  if (existing && opts?.baseUrl === undefined && opts?.apiKey === undefined) {
+  if (existing) {
+    const prevBase = (existing as { configuredBaseUrl?: string }).configuredBaseUrl;
+    if (opts?.baseUrl !== undefined && prevBase !== undefined && prevBase !== opts.baseUrl) {
+      throw new Error(
+        `[Agent Accelerator] Conflicting re-registration for provider "${id}". Registry is immutable after creation.`
+      );
+    }
     return existing;
   }
-  const created = new OpenAICompatibleChatProvider(id, opts);
+  const created = new OpenAICompatibleChatProvider(id, { baseUrl: opts?.baseUrl, name: opts?.name });
   providerRegistry.set(id, created);
   return created;
 }
@@ -199,11 +212,6 @@ export function resolveModel(model: string | ModelSpec): ResolvedModel {
         parts.length === 2 && modelStr.includes(":") && !hasDedicatedEnv;
 
       if (!looksLikeOpenRouterScope) {
-        if (hasCustomOpenAIBase && !hasDedicatedEnv) {
-          const p = getProvider("openai");
-          const spec = p.getModel(remainingModel) || getModelFromCatalog(remainingModel, remainingModel);
-          return { provider: p, modelId: remainingModel, modelSpec: spec as ModelSpec };
-        }
         const custom = ensureCustomProvider(providerPrefix);
         const spec =
           custom.getModel(remainingModel) ||
@@ -372,10 +380,6 @@ export const ModelProvider = {
     options?: { thinkingLevel?: ThinkingLevel; baseUrl?: string }
   ): ModelProviderInstance {
     const normalized = model.includes("/") ? model : `custom/${model}`;
-    const prefix = normalized.split("/")[0]!;
-    try {
-      ensureCustomProvider(prefix, { baseUrl: options?.baseUrl, apiKey });
-    } catch {}
     return {
       model: normalized,
       apiKey,