@toddzheng024/dscode-bundle 0.7.6 → 0.7.7

package/plugins/openrouter/wire.mjs

@@ -0,0 +1,413 @@
+import { CONTEXT_WINDOW_EXCEEDED_CODE, EMPTY_RESPONSE_CODE, LlmError, QUOTA_EXCEEDED_CODE, isContextWindowExceededError, isQuotaExceededError, requestImageHandleText, textOnlyImageText } from '@deepseek-ai/dsh-llm';
+import { OPENROUTER_EFFORTS } from '../providers/catalog.mjs';
+import { ultraRequest } from '../ultra/policy.mjs';
+
+// OpenRouter chat completions on the wire: request bodies, the SSE stream, and
+// the translation into harness stream chunks. Pure functions; the adapter owns I/O.
+
+export const PROVIDER = 'openrouter';
+/** `replayState.response.kind` for responses this adapter produced. */
+export const REPLAY_KIND = 'dscode-openrouter';
+const TOOL_RESULT_IMAGE_TEXT = 'Attached image(s) from tool result:';
+// OpenRouter serves DeepSeek V4 thinking as none/high/xhigh; the route keeps the
+// official off/low/high/max detents (and Ultra on max) with the spelling it accepts.
+const DEEPSEEK_WIRE = Object.freeze({ ...OPENROUTER_EFFORTS, ultra: OPENROUTER_EFFORTS.max });
+const DEEPSEEK_V4 = /^deepseek\/deepseek-v4/;
+// Models whose provider rejects an assistant tool-call turn without `reasoning_content`.
+const REASONING_CONTENT_MODELS = /^(?:deepseek\/deepseek-v4|moonshotai\/kimi-k2\.6)/;
+// Alibaba caches only at explicit breakpoints.
+const EXPLICIT_CACHE_MODELS = /^qwen\//;
+const DEFAULT_LEVELS = ['low', 'medium', 'high'];
+// OpenRouter picks the upstream endpoint, but only among endpoints that honour every
+// request parameter (by default it silently drops unsupported ones) and that do not
+// serve fp4-quantized weights. INT4 stays: Kimi ships native INT4 weights.
+const ROUTING = Object.freeze({ require_parameters: true, quantizations: Object.freeze(['int4', 'int8', 'fp6', 'fp8', 'mxfp8', 'fp16', 'bf16', 'fp32', 'unknown']) });
+const NAMES = { off: 'Off', minimal: 'Minimal', low: 'Low', medium: 'Medium', high: 'High', xhigh: 'XHigh', max: 'Max', ultra: 'Ultra' };
+export const ULTRA_DESCRIPTION = 'DSCODE: max reasoning plus deliberate subagent collaboration; higher total token use.';
+
+/**
+ * Reasoning controls a model offers through this route.
+ * @param model - OpenRouter model id.
+ * @param entry - the model's listing entry, if any.
+ * @returns `{ levels, defaultEffort?, wire }` with harness effort ids and their wire
+ *   spelling, or undefined when the model has no reasoning control.
+ */
+export function modelReasoning(model, entry) {
+  if (DEEPSEEK_V4.test(model)) return { levels: ['off', 'low', 'high', 'max', 'ultra'], defaultEffort: 'high', wire: DEEPSEEK_WIRE };
+  const reasoning = entry?.reasoning;
+  if (!reasoning) return undefined;
+  // No allowlist means OpenRouter maps any standard level; offer the common three.
+  const levels = reasoning.efforts ?? DEFAULT_LEVELS;
+  const ultra = levels.includes('max');
+  const wire = Object.fromEntries([...(reasoning.mandatory ? [] : [['off', 'none']]), ...levels.map(level => [level, level]), ...(ultra ? [['ultra', 'max']] : [])]);
+  // Like the official route, requests default to high; a model without high keeps its own default.
+  const defaultEffort = levels.includes('high') ? 'high' : levels.includes(reasoning.defaultEffort) ? reasoning.defaultEffort : undefined;
+  return { levels: Object.keys(wire), ...(defaultEffort ? { defaultEffort } : {}), wire };
+}
+
+/** Display metadata for one harness effort id. */
+export function effortInfo(id) {
+  return { id, name: NAMES[id] ?? id, ...(id === 'ultra' ? { description: ULTRA_DESCRIPTION } : {}) };
+}
+
+/** A title needs no deliberation: reasoning off when the model allows it, else its lowest level. */
+function titleEffort(reasoning) {
+  return reasoning.wire.off !== undefined ? 'off' : reasoning.levels.find(level => level !== 'ultra');
+}
+
+const textOf = blocks => blocks.filter(block => block.type === 'text').map(block => block.text).join('');
+
+function contentParts(blocks, images) {
+  const parts = [];
+  for (const block of blocks) {
+    if (block.type === 'text' && block.text.length > 0) parts.push({ type: 'text', text: block.text });
+    else if (block.type === 'image') {
+      const version = images?.versions.get(block.attachment.attachmentId);
+      if (version === undefined) { parts.push({ type: 'text', text: textOnlyImageText(block.attachment) }); continue; }
+      parts.push({ type: 'text', text: requestImageHandleText(block.attachment, version, images.access?.(block.attachment)) });
+      parts.push({ type: 'image_url', image_url: { url: `data:${version.mediaType};base64,${Buffer.from(version.data).toString('base64')}` } });
+    }
+  }
+  return parts;
+}
+
+const collapse = parts => parts.every(part => part.type === 'text') ? parts.map(part => part.text).join('') : parts;
+
+/** Reasoning details this adapter (or the pi-ai adapter before it) stored for a same-model replay. */
+function replayedDetails(message) {
+  const blocks = message.source?.replayState?.blocks;
+  if (!Array.isArray(blocks) || blocks.length !== message.content.length) return [];
+  const details = [];
+  blocks.forEach((block, index) => {
+    if (message.content[index]?.type !== 'reasoning') return;
+    if (Array.isArray(block?.reasoningDetails)) details.push(...block.reasoningDetails);
+    else if (typeof block?.thinkingSignature === 'string') {
+      try {
+        const parsed = JSON.parse(block.thinkingSignature);
+        if (Array.isArray(parsed)) details.push(...parsed);
+      } catch { /* a plain field-name signature carries no details */ }
+    }
+  });
+  return details;
+}
+
+function serializeAssistant(message, model) {
+  const text = textOf(message.content);
+  const calls = message.content.filter(block => block.type === 'tool-call')
+    .map(block => ({ id: block.id, type: 'function', function: { name: block.name, arguments: block.arguments } }));
+  // An empty assistant turn (reasoning only) has nothing a provider accepts.
+  if (text.length === 0 && calls.length === 0) return undefined;
+  // Reasoning replays only to the model that produced it.
+  const sameModel = message.source?.provider === PROVIDER && message.source?.model === model;
+  const details = sameModel ? replayedDetails(message) : [];
+  const reasoning = sameModel ? message.content.filter(block => block.type === 'reasoning').map(block => block.text).join('') : '';
+  return {
+    role: 'assistant',
+    content: text,
+    ...(calls.length > 0 ? { tool_calls: calls } : {}),
+    ...(details.length > 0 ? { reasoning_details: details } : reasoning.length > 0 ? { reasoning } : {}),
+    ...(REASONING_CONTENT_MODELS.test(model) ? { reasoning_content: '' } : {}),
+  };
+}
+
+/**
+ * Harness history as OpenRouter chat messages. Each tool result becomes a `tool`
+ * message; its images follow in one user message, which `tool` content cannot carry.
+ * @param images - prepared request images (`versions` by attachment id, `access`), when the request has any.
+ */
+export function serializeMessages(messages, { model, system, images } = {}) {
+  const wire = [];
+  if (system !== undefined && system.length > 0) wire.push({ role: 'system', content: system });
+  let pendingImages = [];
+  const flush = () => {
+    if (pendingImages.length === 0) return;
+    wire.push({ role: 'user', content: [{ type: 'text', text: TOOL_RESULT_IMAGE_TEXT }, ...pendingImages] });
+    pendingImages = [];
+  };
+  for (const message of messages) {
+    if (message.role === 'system') {
+      flush();
+      const text = textOf(message.content);
+      if (text.length > 0) wire.push({ role: 'system', content: text });
+      continue;
+    }
+    if (message.role === 'assistant') {
+      flush();
+      const entry = serializeAssistant(message, model);
+      if (entry) wire.push(entry);
+      continue;
+    }
+    const results = message.content.filter(block => block.type === 'tool-result');
+    const parts = contentParts(message.content.filter(block => block.type !== 'tool-result'), images);
+    if (parts.length > 0 || results.length === 0) {
+      flush();
+      wire.push({ role: 'user', content: collapse(parts) });
+    }
+    for (const result of results) {
+      const nested = contentParts(result.content, images);
+      const text = nested.filter(part => part.type === 'text').map(part => part.text).join('');
+      const attached = nested.filter(part => part.type !== 'text');
+      wire.push({ role: 'tool', tool_call_id: result.toolCallId, content: text || (attached.length > 0 ? '(see attached image)' : '(no output)') });
+      pendingImages.push(...attached);
+    }
+  }
+  flush();
+  return wire;
+}
+
+const cacheMarked = content => typeof content === 'string'
+  ? [{ type: 'text', text: content, cache_control: { type: 'ephemeral' } }]
+  : content.map((part, index) => index === content.findLastIndex(candidate => candidate.type === 'text') ? { ...part, cache_control: { type: 'ephemeral' } } : part);
+
+/** Explicit cache breakpoints for models that cache only at them: the system prompt and the latest user turn. */
+export function withCacheBreakpoints(messages, model) {
+  if (!EXPLICIT_CACHE_MODELS.test(model)) return messages;
+  const targets = new Set([messages.findIndex(message => message.role === 'system'), messages.findLastIndex(message => message.role === 'user')]);
+  return messages.map((message, index) => targets.has(index) && message.content !== '' ? { ...message, content: cacheMarked(message.content) } : message);
+}
+
+/**
+ * The chat-completions body for one harness request.
+ * @param options - harness request (images already prepared into `images`).
+ * @param context - `{ entry, images }`: the model's listing entry and prepared request images.
+ */
+export function requestBody(options, { entry, images } = {}) {
+  const reasoning = modelReasoning(options.model, entry);
+  const effort = options.purpose === 'session-title' && reasoning ? titleEffort(reasoning) : options.reasoningEffort;
+  const wireEffort = effort === undefined ? undefined : reasoning?.wire[effort];
+  if (effort !== undefined && reasoning !== undefined && wireEffort === undefined) {
+    throw new LlmError(`OpenRouter model "${options.model}" does not offer reasoning effort "${effort}"`, 'UNSUPPORTED_REASONING_EFFORT');
+  }
+  let messages = serializeMessages(options.messages, { model: options.model, system: options.system, images });
+  messages = withCacheBreakpoints(ultraRequest(options, messages), options.model);
+  // Delegation tools are offered at every effort; workflow and ralph never are.
+  const tools = (options.tools ?? []).filter(tool => tool.name !== 'workflow' && tool.name !== 'ralph')
+    .map(tool => ({ type: 'function', function: { name: tool.name, description: tool.description, parameters: tool.parameters } }));
+  return {
+    model: options.model,
+    messages,
+    stream: true,
+    ...(tools.length > 0 ? { tools } : {}),
+    ...(wireEffort !== undefined ? { reasoning: { effort: wireEffort } } : {}),
+    // Endpoints declare `max_tokens`; `max_completion_tokens` would fail `require_parameters` almost everywhere.
+    ...(options.maxTokens !== undefined ? { max_tokens: options.maxTokens } : {}),
+    ...(options.temperature !== undefined ? { temperature: options.temperature } : {}),
+    ...(options.stop !== undefined ? { stop: options.stop } : {}),
+    // The sticky-routing key: one session keeps one upstream endpoint and its warm cache.
+    ...(options.sessionId !== undefined ? { session_id: String(options.sessionId).slice(0, 256) } : {}),
+    provider: { require_parameters: ROUTING.require_parameters, quantizations: [...ROUTING.quantizations] },
+  };
+}
+
+// OpenRouter's typed `metadata.error_type`, which it asks clients to route on before the status.
+const ERROR_TYPES = Object.freeze({
+  context_length_exceeded: CONTEXT_WINDOW_EXCEEDED_CODE, token_limit_exceeded: CONTEXT_WINDOW_EXCEEDED_CODE,
+  payment_required: QUOTA_EXCEEDED_CODE, authentication: 'AUTH', permission_denied: 'AUTH',
+  rate_limit_exceeded: 'RATE_LIMIT', provider_overloaded: 'SERVER', provider_unavailable: 'SERVER', server: 'SERVER', unmapped: 'SERVER',
+  timeout: 'TIMEOUT', content_policy_violation: 'CONTENT_POLICY', refusal: 'CONTENT_POLICY',
+});
+
+/**
+ * Harness failure code for an OpenRouter error.
+ * @param status - HTTP status of a rejected request, or undefined for an error inside a 200 stream.
+ * @param error - the `error` object (`code`, `message`, `metadata`).
+ */
+export function errorCode(status, error) {
+  const type = error?.metadata?.error_type;
+  if (typeof type === 'string' && Object.hasOwn(ERROR_TYPES, type)) return ERROR_TYPES[type];
+  const code = Number.isInteger(status) ? status : Number(error?.code);
+  // A 402 names token counts ("fewer max_tokens"); its status decides before any wording does.
+  if (code === 402) return QUOTA_EXCEEDED_CODE;
+  const detail = [error?.message, error?.metadata?.raw].filter(value => typeof value === 'string').join(' ');
+  if (isContextWindowExceededError(detail)) return CONTEXT_WINDOW_EXCEEDED_CODE;
+  if (isQuotaExceededError(detail)) return QUOTA_EXCEEDED_CODE;
+  if (code === 401 || code === 403) return 'AUTH';
+  if (code === 429) return 'RATE_LIMIT';
+  if (code === 408 || code === 504) return 'TIMEOUT';
+  if (code >= 500) return 'SERVER';
+  if (code >= 400) return 'INVALID_REQUEST';
+  // A stream that failed after its 200 without a numeric code is an upstream failure worth retrying.
+  return 'SERVER';
+}
+
+/** Human-readable message for an OpenRouter error object. */
+export function errorMessage(error, fallback) {
+  const message = typeof error?.message === 'string' && error.message.length > 0 ? error.message : fallback;
+  const provider = error?.metadata?.provider_name;
+  return typeof provider === 'string' && provider.length > 0 ? `${message} (provider: ${provider})` : message;
+}
+
+/** Retry-After as milliseconds (seconds or an HTTP date), when valid. */
+export function retryAfterMs(value) {
+  if (value === null || value === undefined) return undefined;
+  const delay = /^\d+$/.test(value) ? Number(value) * 1e3 : Date.parse(value) - Date.now();
+  return Number.isFinite(delay) && delay > 0 ? delay : undefined;
+}
+
+/**
+ * SSE `data:` payloads from a response body, skipping `:` keep-alive comments.
+ * @param onActivity - called for every received chunk, comments included.
+ */
+export async function* sseData(body, onActivity) {
+  const decoder = new TextDecoder();
+  let buffer = '', data = [];
+  const lines = function* (final) {
+    let newline;
+    while ((newline = buffer.indexOf('\n')) >= 0 || final && buffer.length > 0) {
+      let line = newline >= 0 ? buffer.slice(0, newline) : buffer;
+      buffer = newline >= 0 ? buffer.slice(newline + 1) : '';
+      if (line.endsWith('\r')) line = line.slice(0, -1);
+      if (line === '') {
+        if (data.length > 0) yield data.join('\n');
+        data = [];
+      } else if (line.startsWith('data:')) data.push(line.slice(line.startsWith('data: ') ? 6 : 5));
+    }
+  };
+  for await (const chunk of body) {
+    onActivity?.();
+    buffer += decoder.decode(chunk, { stream: true });
+    yield* lines(false);
+  }
+  buffer += decoder.decode();
+  yield* lines(true);
+  if (data.length > 0) yield data.join('\n');
+}
+
+/** Map OpenRouter usage to disjoint harness counts (`prompt_tokens` includes cache reads and writes). */
+export function mapUsage(usage) {
+  const valid = value => Number.isSafeInteger(value) && value >= 0;
+  const prompt = usage?.prompt_tokens, completion = usage?.completion_tokens;
+  if (!valid(prompt) || !valid(completion)) return undefined;
+  const read = valid(usage.prompt_tokens_details?.cached_tokens) ? usage.prompt_tokens_details.cached_tokens : 0;
+  const write = valid(usage.prompt_tokens_details?.cache_write_tokens) ? usage.prompt_tokens_details.cache_write_tokens : 0;
+  const reasoning = usage.completion_tokens_details?.reasoning_tokens;
+  return {
+    inputTokens: Math.max(0, prompt - read - write),
+    outputTokens: completion,
+    ...(usage.total_tokens === undefined || usage.total_tokens === prompt + completion ? { totalTokens: prompt + completion } : {}),
+    ...(read > 0 ? { cacheReadTokens: read } : {}),
+    ...(write > 0 ? { cacheWriteTokens: write } : {}),
+    ...(valid(reasoning) && reasoning > 0 ? { reasoningTokens: reasoning } : {}),
+  };
+}
+
+/** Streamed reasoning details merged into the blocks a replay sends back unmodified. */
+function mergeDetails(items) {
+  const merged = [];
+  for (const item of items) {
+    if (item === null || typeof item !== 'object') continue;
+    const last = merged.at(-1);
+    if (last && last.type === item.type && item.type !== 'reasoning.encrypted' && (last.index ?? 0) === (item.index ?? 0)) {
+      if (typeof item.text === 'string') last.text = (last.text ?? '') + item.text;
+      if (typeof item.summary === 'string') last.summary = (last.summary ?? '') + item.summary;
+      for (const key of ['id', 'format', 'signature']) if (item[key] !== undefined && item[key] !== null && item[key] !== '') last[key] = item[key];
+    } else merged.push({ ...item });
+  }
+  return merged;
+}
+
+function closeBlock(block) {
+  if (block.kind === 'tool-call') return { type: 'tool-call', id: block.callId ?? '', name: block.name ?? '', arguments: block.text };
+  return { type: block.kind, text: block.text };
+}
+
+function finishReason(reason, blocks) {
+  if (reason === 'length') return { kind: 'max-tokens' };
+  if (reason === 'content_filter') return { kind: 'error', failure: { message: 'OpenRouter stopped the response for content filtering', code: 'CONTENT_FILTER' } };
+  if (reason === 'tool_calls' || blocks.some(block => block.kind === 'tool-call') && (reason === undefined || reason === 'stop')) return { kind: 'tool-calls' };
+  if (reason === undefined || reason === null || reason === 'stop' || reason === 'end') {
+    return blocks.length === 0 ? { kind: 'error', failure: { message: 'model returned a completed response with no content', code: EMPTY_RESPONSE_CODE } } : { kind: 'stop' };
+  }
+  return { kind: 'error', failure: { message: `model stopped: ${reason}`, code: String(reason).toUpperCase() } };
+}
+
+/**
+ * Translate SSE payloads into harness chunks. Block ends, usage and the finish are
+ * held until `[DONE]`; the finish of a successful response carries the replay state:
+ * reasoning details per block, the generation id, the serving provider and the billed cost.
+ * An error inside the stream throws with its routed code.
+ */
+export async function* translate(payloads, { model }) {
+  let nextIndex = 0, textBlock, reasoningBlock, finish, usage, cost, id, provider;
+  const tools = new Map();
+  const order = [];
+  const open = kind => {
+    const block = { index: nextIndex++, kind, text: '', details: [] };
+    order.push(block);
+    return block;
+  };
+  for await (const payload of payloads) {
+    if (payload === '[DONE]') {
+      for (const block of order) yield { type: 'block-end', index: block.index, block: closeBlock(block) };
+      if (usage) yield { type: 'usage', usage };
+      const reason = finishReason(finish, order);
+      const succeeded = reason.kind === 'stop' || reason.kind === 'tool-calls' || reason.kind === 'max-tokens';
+      yield {
+        type: 'finish', reason,
+        ...(succeeded ? { replayState: {
+          response: { kind: REPLAY_KIND, version: 1, model, ...(id ? { id } : {}), ...(provider ? { provider } : {}), ...(cost !== undefined ? { cost } : {}) },
+          blocks: order.map(block => block.kind === 'reasoning' && block.details.length > 0 ? { type: 'reasoning', reasoningDetails: mergeDetails(block.details) } : { type: block.kind }),
+        } } : {}),
+      };
+      return;
+    }
+    let chunk;
+    try {
+      chunk = JSON.parse(payload);
+    } catch {
+      throw new LlmError(`malformed OpenRouter stream payload: ${payload.slice(0, 120)}`, 'MALFORMED_RESPONSE');
+    }
+    if (chunk?.error) {
+      const status = Number.isInteger(chunk.error.code) ? chunk.error.code : undefined;
+      throw new LlmError(errorMessage(chunk.error, 'OpenRouter stream error'), errorCode(undefined, chunk.error), status === undefined ? {} : { status });
+    }
+    if (typeof chunk?.id === 'string') id ??= chunk.id;
+    if (typeof chunk?.provider === 'string') provider ??= chunk.provider;
+    for (const choice of chunk?.choices ?? []) {
+      const delta = choice.delta ?? {};
+      const details = Array.isArray(delta.reasoning_details) ? delta.reasoning_details : [];
+      let reasoning = typeof delta.reasoning === 'string' ? delta.reasoning : typeof delta.reasoning_content === 'string' ? delta.reasoning_content : '';
+      if (reasoning.length === 0) reasoning = details.map(detail => detail?.type === 'reasoning.text' ? detail.text : detail?.type === 'reasoning.summary' ? detail.summary : '').filter(text => typeof text === 'string').join('');
+      if (reasoning.length > 0 || details.length > 0) {
+        if (!reasoningBlock) {
+          reasoningBlock = open('reasoning');
+          yield { type: 'block-start', index: reasoningBlock.index, blockType: 'reasoning' };
+        }
+        reasoningBlock.details.push(...details);
+        if (reasoning.length > 0) {
+          reasoningBlock.text += reasoning;
+          yield { type: 'reasoning-delta', index: reasoningBlock.index, text: reasoning };
+        }
+      }
+      if (typeof delta.content === 'string' && delta.content.length > 0) {
+        if (!textBlock) {
+          textBlock = open('text');
+          yield { type: 'block-start', index: textBlock.index, blockType: 'text' };
+        }
+        textBlock.text += delta.content;
+        yield { type: 'text-delta', index: textBlock.index, text: delta.content };
+      }
+      for (const call of Array.isArray(delta.tool_calls) ? delta.tool_calls : []) {
+        const key = call.index ?? call.id;
+        let block = tools.get(key);
+        if (!block) {
+          block = open('tool-call');
+          tools.set(key, block);
+          yield { type: 'block-start', index: block.index, blockType: 'tool-call' };
+        }
+        // id and name arrive once; an empty or null repeat is no update.
+        if (typeof call.id === 'string' && call.id.length > 0) block.callId = call.id;
+        if (typeof call.function?.name === 'string' && call.function.name.length > 0) block.name = call.function.name;
+        const fragment = typeof call.function?.arguments === 'string' ? call.function.arguments : '';
+        block.text += fragment;
+        yield { type: 'tool-call-delta', index: block.index, id: block.callId ?? '', ...(block.name !== undefined ? { name: block.name } : {}), argumentsDelta: fragment };
+      }
+      if (typeof choice.finish_reason === 'string') finish = choice.finish_reason;
+    }
+    if (chunk?.usage) {
+      usage = mapUsage(chunk.usage) ?? usage;
+      if (Number.isFinite(chunk.usage.cost) && chunk.usage.cost >= 0) cost = chunk.usage.cost;
+    }
+  }
+  throw new LlmError('OpenRouter stream ended without [DONE]', 'TRANSPORT');
+}

package/plugins/providers/catalog.mjs

@@ -1,83 +1,42 @@
 // Model providers `/provider` switches between. DeepSeek's official API is the
-// native `llm-deepseek` route; OpenRouter serves the DeepSeek models and the rest
-// of pi-ai's OpenRouter catalog through pi-ai's catalog route, which the base
-// composition mounts dormant until a `llm-pi-ai:` settings section declares it.
+// native `llm-deepseek` route; OpenRouter is DSCODE's own adapter
+// (plugins/openrouter), which serves OpenRouter's live model listing and is always
+// registered. This module ships beside the TUI too, so it imports nothing.
 
 export const PROVIDERS = Object.freeze([
   { id: 'deepseek-official', name: 'DeepSeek', aliases: ['deepseek', 'deepseek-official', 'official'], credentialRef: 'DEEPSEEK_API_KEY', defaultModel: 'deepseek-flash' },
-  { id: 'openrouter', name: 'OpenRouter', aliases: ['openrouter', 'open-router'], credentialRef: 'OPENROUTER_API_KEY', defaultModel: 'deepseek/deepseek-v4-flash' },
+  // The optional management key reads account data only; it cannot call models.
+  { id: 'openrouter', name: 'OpenRouter', aliases: ['openrouter', 'open-router'], credentialRef: 'OPENROUTER_API_KEY', managementRef: 'OPENROUTER_MANAGEMENT_KEY', defaultModel: 'deepseek/deepseek-v4-flash' },
 ]);
 
+// The pi-ai adapter served OpenRouter until 0.7.6, from this settings section.
 const PI_AI_NS = 'llm-pi-ai';
 
 // OpenRouter serves DeepSeek V4 thinking as none/high/xhigh. DeepSeek itself
 // answers `low` as high and `max` as xhigh, so the route offers the official
 // low/high/max detents (and Ultra on top of max) with the wire spelling OpenRouter
 // accepts; session cards and delegated children that ask for `low` keep working.
-const OPENROUTER_EFFORTS = Object.freeze({ off: 'none', low: 'high', high: 'high', max: 'xhigh' });
+export const OPENROUTER_EFFORTS = Object.freeze({ off: 'none', low: 'high', high: 'high', max: 'xhigh' });
 
-/** The DeepSeek models the OpenRouter route declares, with their official-route counterparts. */
+/** The DeepSeek models the OpenRouter route serves, with their official-route counterparts. */
 export const OPENROUTER_MODELS = Object.freeze([
   { id: 'deepseek/deepseek-v4-flash', name: 'DeepSeek V4 Flash', official: ['deepseek-flash', 'deepseek-v4-flash'] },
   { id: 'deepseek/deepseek-v4-pro', name: 'DeepSeek V4 Pro', official: ['deepseek-v4-pro'] },
   { id: 'deepseek/deepseek-v4-flash-vision-exp', name: 'DeepSeek V4 Flash Vision Exp', official: ['deepseek-v4-flash-vision-exp'] },
 ]);
 
-/** The `llm-pi-ai` profile `/provider openrouter` writes: pi-ai's whole OpenRouter catalog. */
-export function openRouterProfile() {
-  return {
-    displayName: 'OpenRouter',
-    apiKeyEnv: 'OPENROUTER_API_KEY',
-    // Like the official route, requests default to high; a model without high uses its own default.
-    reasoning: 'high',
-    // No models list, so the route serves every catalog model; the overrides give the
-    // DeepSeek models the official detents (and /effort its detent bar).
-    modelOverrides: Object.fromEntries(OPENROUTER_MODELS.map(({ id, name }) => [id, { name, reasoningEfforts: { ...OPENROUTER_EFFORTS } }])),
-  };
-}
-
-/** The profile 0.7.3 to 0.7.5 wrote: the catalog narrowed to the three DeepSeek models. */
-export function narrowOpenRouterProfile() {
-  return {
-    displayName: 'OpenRouter',
-    apiKeyEnv: 'OPENROUTER_API_KEY',
-    reasoning: 'high',
-    models: OPENROUTER_MODELS.map(({ id, name }) => ({ id, name, reasoningEfforts: { ...OPENROUTER_EFFORTS } })),
-  };
-}
-
-// Fields a user sets to point the route elsewhere or shape its requests. The settings
-// service describes a profile with its resolved defaults, which fill these with empty
-// values (`input: []`, `compat: { chatTemplateKwargs: {}, ... }`), so empty counts as unset.
-const USER_FIELDS = ['api', 'baseURL', 'modelOverrides', 'headers', 'compat', 'thinkingBudgets', 'cacheRetention', 'transport'];
-const empty = value => value === undefined || (Array.isArray(value) ? value.length === 0
-  : value !== null && typeof value === 'object' && Object.values(value).every(empty));
-const sameEfforts = (left, right) => left !== null && typeof left === 'object'
-  && Object.keys(left).length === Object.keys(right).length && Object.entries(right).every(([level, wire]) => left[level] === wire);
-
-/** Whether a stored profile is exactly the narrow one DSCODE wrote, so replacing it discards nothing the user chose. */
-export function isNarrowOpenRouterProfile(profile) {
-  if (profile === null || typeof profile !== 'object' || !Array.isArray(profile.models)) return false;
-  const narrow = narrowOpenRouterProfile();
-  return profile.displayName === narrow.displayName && profile.apiKeyEnv === narrow.apiKeyEnv && profile.reasoning === narrow.reasoning
-    && USER_FIELDS.every(field => empty(profile[field]))
-    && profile.models.length === narrow.models.length
-    && narrow.models.every((expected, index) => {
-      const { id, name, reasoningEfforts, ...rest } = profile.models[index] ?? {};
-      return id === expected.id && name === expected.name && sameEfforts(reasoningEfforts, expected.reasoningEfforts) && Object.values(rest).every(empty);
-    });
-}
-
 /**
- * Replace the narrow profile earlier builds wrote with the whole-catalog one.
- * Never declares a route and never throws: /model must open regardless.
+ * Remove the `openrouter` profile earlier builds wrote into the pi-ai section. The
+ * pi-ai adapter is no longer mounted, so the profile is inert, and it would claim the
+ * route a second time if that adapter were ever mounted again.
+ * Never throws: /model and /provider must open regardless.
  * @returns whether the settings changed.
  */
 export async function migrateOpenRouterProfile(settings) {
   try {
     const descriptor = settings?.describe?.({ redactSecrets: true }).find(entry => entry.ns === PI_AI_NS);
-    if (!descriptor || settings.writable !== true || !isNarrowOpenRouterProfile(descriptor.value?.providers?.openrouter)) return false;
-    await settings.mutate(PI_AI_NS, [{ op: 'set', path: ['providers', 'openrouter'], value: openRouterProfile() }], descriptor.revision);
+    if (!descriptor || settings.writable !== true || descriptor.value?.providers?.openrouter === undefined) return false;
+    await settings.mutate(PI_AI_NS, [{ op: 'unset', path: ['providers', 'openrouter'] }], descriptor.revision);
     return true;
   } catch {
     return false;
@@ -156,25 +115,16 @@ export function credentialState(row) {
 }
 
 /**
- * Declare a provider's route before it is used. Only OpenRouter needs one; a
- * profile the user already has (their own models or endpoint) is left alone,
- * while the narrow profile earlier builds wrote is replaced.
+ * Prepare a provider before a switch. Both routes are always registered; switching
+ * to OpenRouter only clears the inert pi-ai profile earlier builds wrote.
  * @param settings - the host settings service.
  * @returns whether the settings changed.
  */
 export async function ensureProviderRoute(settings, provider) {
-  if (provider !== 'openrouter') return false;
-  if (typeof settings?.describe !== 'function' || typeof settings.mutate !== 'function') throw new Error('settings are unavailable; OpenRouter cannot be configured in this profile');
-  const descriptor = settings.describe({ redactSecrets: true }).find(entry => entry.ns === PI_AI_NS);
-  if (!descriptor) throw new Error('the OpenRouter adapter (llm-pi-ai) is not mounted in this profile');
-  const existing = descriptor.value?.providers?.openrouter;
-  if (existing !== undefined) return migrateOpenRouterProfile(settings);
-  if (settings.writable !== true) throw new Error('settings are read-only; OpenRouter cannot be configured here');
-  await settings.mutate(PI_AI_NS, [{ op: 'set', path: ['providers', 'openrouter'], value: openRouterProfile() }], descriptor.revision);
-  return true;
+  return provider === 'openrouter' ? migrateOpenRouterProfile(settings) : false;
 }
 
-/** Wait for a freshly declared route to reach the model directory. */
+/** Wait for a route's models to reach the model directory. */
 export async function waitForModels(loadModels, provider, { attempts = 30, delayMs = 100 } = {}) {
   let directory;
   for (let attempt = 0; attempt < attempts; attempt++) {