cartethyia 2.5.4 → 2.5.7

package/src/providers/integrations/cursor/cursor.ts

@@ -37,7 +37,6 @@ import {
   buildCancelAction,
   buildExecStreamClose,
   buildExecThrow,
-  buildHostExecResult,
   buildInteractionResponse,
   buildKvGetBlobResult,
   buildKvSetBlobResult,
@@ -56,7 +55,7 @@ import {
 } from "./cursor-request";
 import { cursorRunTerminationError, decodeServerMessages, openCursorRun } from "./cursor-http2";
 import { cursorRunSessions } from "./cursor-session-manager";
-import type { CursorContinuationScope, CursorHostExecDetail, CursorRunSession } from "./cursor-session-manager";
+import type { CursorContinuationScope, CursorRunSession } from "./cursor-session-manager";
 import { fetchCursorRunOrigin } from "./cursor-catalog";
 import { defaultParameterValues, fetchCursorModels } from "./cursor-catalog";
 import type { CursorModelInfo } from "./cursor-catalog";
@@ -289,126 +288,6 @@ function readExecutableMcpCall(
   return { toolName, arguments: decodeMcpArguments(mcpArgs) };
 }
 
-/** Host exec variants this gateway can forward to the client's own tools. */
-const HOST_TOOL_ALIASES: Record<CursorHostExecDetail["variant"], readonly string[]> = {
-  shell: ["bash", "shell", "sh"],
-  write: ["write", "file_write", "create_file"],
-  read: ["read", "read_file", "view", "cat"],
-};
-
-/** Canonical exec value → property names a client tool schema may declare. */
-const HOST_ARG_ALIASES: Readonly<Record<string, readonly string[]>> = {
-  intent: ["i", "intent", "description"],
-  command: ["command", "cmd"],
-  workingDirectory: ["working_directory", "cwd", "workdir"],
-  path: ["path", "file_path", "filename"],
-  content: ["content", "text", "file_text", "contents"],
-};
-
-function fileBytesToText(value: unknown): string {
-  return value instanceof Uint8Array ? new TextDecoder().decode(value) : "";
-}
-
-function byteLengthOf(text: string): number {
-  return new TextEncoder().encode(text).length;
-}
-
-/**
- * Projects one host exec request onto a client-declared tool descriptor.
- *
- * The argument keys are taken from the schema the tool itself declared, so a
- * strict schema (`additionalProperties: false`) never rejects the call; a
- * property the schema demands but no exec value fills refuses the forward —
- * calling the tool with a half-built record would strand the model on an
- * error it can only work around by circling.
- */
-function hostCallFor(
-  descriptors: readonly CursorToolDescriptor[],
-  variant: CursorHostExecDetail["variant"],
-  values: Readonly<Record<string, string | undefined>>,
-  detail: CursorHostExecDetail,
-): { toolName: string; arguments: string; detail: CursorHostExecDetail } | undefined {
-  const descriptor = descriptors.find((entry) =>
-    HOST_TOOL_ALIASES[variant].includes(entry.toolName.toLowerCase().replace(/^_+/, "")),
-  );
-  if (descriptor === undefined) return undefined;
-  const schema = recordOf(descriptor.inputSchema);
-  const properties = recordOf(schema?.["properties"]);
-  const required = Array.isArray(schema?.["required"])
-    ? (schema?.["required"] as unknown[]).filter((value): value is string => typeof value === "string")
-    : [];
-  const subject =
-    values.command !== undefined
-      ? values.command
-      : `${variant} ${values.path ?? ""}`.trim();
-  const filled: Record<string, string> = {};
-  for (const [key, value] of Object.entries(values)) {
-    if (value === undefined) continue;
-    const target = HOST_ARG_ALIASES[key]?.find((alias) => properties?.[alias] !== undefined)
-      ?? (properties === undefined ? key : undefined);
-    if (target !== undefined) filled[target] = value;
-  }
-  // The client surfaces carry a concise-intent property the exec frame has no
-  // counterpart for: derive it from the subject and project it like the rest.
-  const intentTarget = HOST_ARG_ALIASES["intent"]?.find((alias) => properties?.[alias] !== undefined);
-  const intent = subject.length > 80 ? `${subject.slice(0, 79)}…` : subject;
-  if (intentTarget !== undefined) filled[intentTarget] = intent;
-  else if (properties === undefined) filled.intent = intent;
-  if (required.some((name) => !(name in filled))) return undefined;
-  return { toolName: descriptor.toolName, arguments: JSON.stringify(filled), detail };
-}
-
-/**
- * Reads one Cursor host exec request the canonical client executes locally.
- *
- * Cursor's edge asks its executor for a native host tool (shell, write, read);
- * the canonical client declared the same surfaces as its own tools, so the
- * gateway forwards the work instead of refusing it. Exec variants without a
- * faithful typed result frame (grep, ls, subagent, …) return `undefined` and
- * keep the throw-refusal path.
- */
-function readHostExecCall(
-  execServer: Record<string, unknown>,
-  descriptors: readonly CursorToolDescriptor[],
-): { toolName: string; arguments: string; detail: CursorHostExecDetail } | undefined {
-  const shell = recordOf(execServer["shell_args"]) ?? recordOf(execServer["shell_stream_args"]);
-  if (shell !== undefined) {
-    const command = nonEmptyStringOf(shell["command"]);
-    if (command === undefined) return undefined;
-    const workingDirectory = nonEmptyStringOf(shell["working_directory"]);
-    return hostCallFor(
-      descriptors,
-      "shell",
-      { command, workingDirectory },
-      {
-        variant: "shell",
-        command,
-        ...(workingDirectory === undefined ? {} : { workingDirectory }),
-      },
-    );
-  }
-  const write = recordOf(execServer["write_args"]);
-  if (write !== undefined) {
-    const path = nonEmptyStringOf(write["path"]);
-    if (path === undefined) return undefined;
-    const content =
-      typeof write["file_text"] === "string" ? write["file_text"] : fileBytesToText(write["file_bytes"]);
-    return hostCallFor(descriptors, "write", { path, content }, {
-      variant: "write",
-      path,
-      lines: content.split("\n").length,
-      fileSize: byteLengthOf(content),
-    });
-  }
-  const read = recordOf(execServer["read_args"]);
-  if (read !== undefined) {
-    const path = nonEmptyStringOf(read["path"]);
-    if (path === undefined) return undefined;
-    return hostCallFor(descriptors, "read", { path }, { variant: "read", path });
-  }
-  return undefined;
-}
-
 /**
  * Decodes Cursor's `McpArgs.args` — a `map<string, google.protobuf.Value>`
  * carried as repeated map-entry messages — into a canonical arguments string.
@@ -685,10 +564,6 @@ class CursorAdapter implements ProviderAdapter {
       for (const item of pending) {
         const entry = cursorRunSessions.resolveCall(item.callId, scope);
         if (entry === undefined) continue;
-        // A forwarded host exec answers the variant's typed result frame (an
-        // mcp_result leaves the server unmatched on a frame it did rank);
-        // MCP-channel calls answer the generic mcp_result.
-        if (entry.call.host === undefined) {
           await handle.write(
             buildMcpResult({
               execId: entry.call.execId,
@@ -698,18 +573,6 @@ class CursorAdapter implements ProviderAdapter {
               ...(item.images.length > 0 ? { images: item.images } : {}),
             }),
           );
-        } else {
-          for (const frame of buildHostExecResult({
-            execId: entry.call.execId,
-            ...(entry.call.execIdString === undefined ? {} : { execIdString: entry.call.execIdString }),
-            variant: entry.call.host.variant,
-            text: item.text,
-            isError: item.isError,
-            host: entry.call.host,
-          })) {
-            await handle.write(frame);
-          }
-        }
         cursorRunSessions.clearPendingCall(runId, item.callId);
       }
       yield* this.#pump({
@@ -1051,7 +914,6 @@ class CursorAdapter implements ProviderAdapter {
             execId: handled.execId,
             ...(handled.execIdString === undefined ? {} : { execIdString: handled.execIdString }),
             toolName: handled.toolName,
-            ...(handled.host === undefined ? {} : { host: handled.host }),
           })) {
             await handle.write(buildMcpRejected(handled.execId, "tool call is not answerable", handled.execIdString));
             continue;
@@ -1201,7 +1063,6 @@ class CursorAdapter implements ProviderAdapter {
         execIdString: string | undefined;
         toolName: string;
         arguments: string;
-        host?: CursorHostExecDetail;
       }
   > {
     const execId = typeof execServer["id"] === "number" ? execServer["id"] : 0;
@@ -1249,29 +1110,16 @@ class CursorAdapter implements ProviderAdapter {
         arguments: call.arguments,
       };
     }
-    // Host exec variants the canonical client declares as its own tools are
-    // forwarded as tool calls — the client executes them locally and the
-    // result leg answers the variant's typed result frame. This is Cursor's
-    // own architecture: its executor is the local CLI, and silencing a host
-    // tool the client could have run just makes the model circle before
-    // falling back to the MCP channel.
-    const host = readHostExecCall(execServer, session.descriptors);
-    if (host !== undefined) {
-      return {
-        kind: "tool_call",
-        execId,
-        execIdString,
-        toolName: host.toolName,
-        arguments: host.arguments,
-        host: host.detail,
-      };
-    }
-    // Every other exec variant is a Cursor host tool (shell, read, edit, glob,
-    // subagent, …) or a frame this gateway has no handler for. Stranding the
-    // exec id stalled the turn forever, and answering it as an mcp_result left
-    // the server unmatched on a frame it did rank — the protocol's failure
-    // channel (throw + stream_close, exactly what Cursor's own executor sends
-    // for an unclaimed frame) surfaces the refusal to the model instead.
+    // Host exec frames (shell_args, write_args, read_args, and every other
+    // Cursor-native variant) are not executed here. Cursor's own executor on
+    // the local CLI handles them; the MCP channel is the correct path for tool
+    // calls routed through this gateway.
+    //
+    // A direct-forward path (`readHostExecCall`) was implemented experimentally
+    // but produced protocol errors in practice: the typed result frames it sent
+    // back (shell_result, write_result, read_result) were not matched correctly
+    // by the server under all code paths, causing the Run to stall or the model
+    // to receive malformed tool output. The path has been removed; use MCP.
     const refusal = "Cursor host tools are not executed by this gateway";
     await handle.write(buildExecThrow(execId, refusal, execIdString));
     await handle.write(buildExecStreamClose(execId, execIdString));

package/src/providers/integrations/factory/factory-models.ts

@@ -8,15 +8,13 @@
  *
  * Gemini ids are listed below on the `gemini` wire family.
  *
- * Settled from the reference client source: the
- * research was right and the old rows were wrong. Factory serves Gemini on a
+ * Factory serves Gemini on a
  * native generateContent wire at `/api/llm/g/v1/generate`
  * (`packages/catalog/src/wire/factory-droid.ts:83-84`,
  * `packages/ai/src/providers/factory-droid/gemini.ts:28`) with
  * `x-api-provider: google` (`packages/ai/test/factory-droid-gemini.test.ts:25`).
- * The rows used to say `chat`, which posted Gemini requests to an endpoint that
- * does not serve them — so they failed at the endpoint, not at routing, which
- * is why nothing caught it.
+ * A Gemini id routed onto `chat` posts to an endpoint that does not serve it —
+ * so it fails at the endpoint, not at routing, which is why nothing catches it.
  *
  * `garnet-07-15`, `gemini-3.1-pro-preview` and `gemini-3.8-flash` remain absent:
  * the donor's catalog does not list them either.

package/src/providers/integrations/grok/grok-install-id.ts

@@ -2,9 +2,8 @@
  * Grok's per-installation agent id.
  *
  * The persistence and failure handling live in the shared install-id module;
- * this file only names the id. Previously each provider carried its own copy of
- * the read/mkdir/open dance, and a failure to write it escaped into dispatch and
- * failed the request.
+ * this file only names the id, so a failure to write it never escapes into
+ * dispatch and fails the request.
  */
 import { getOrCreateInstallId } from "../shared/install-id";
 

package/src/providers/integrations/grok/grok.ts

@@ -373,11 +373,10 @@ function normalizeTools(payload: Record<string, unknown>): void {
       normalized.push(tool);
       continue;
     }
-    // Tools arrive factory-normalized (flat `{type:"function",name,parameters}`;
-    // canonical tools are flat, so the legacy nested `function.*` resolution
-    // duplicated the factory payload path and is deleted). Only grok wire
-    // strictness stays here: trimmed names, empty-name drop, parameters
-    // fallback, custom-tool synthesis.
+    // Tools arrive factory-normalized (flat `{type:"function",name,parameters}`
+    // — canonical tools are flat, so there is no nested `function.*` payload to
+    // resolve). Only grok wire strictness stays here: trimmed names,
+    // empty-name drop, parameters fallback, custom-tool synthesis.
     const name = typeof tool.name === "string" ? tool.name : "";
     if (!name.trim()) continue;
     const parameters =
@@ -413,8 +412,8 @@ function normalizeTools(payload: Record<string, unknown>): void {
     if (type === "function" || type === "custom") {
       // Unlike `tools`, `tool_choice` genuinely arrives nested: the chat
       // encoder writes `{type:"function", function:{name}}` and
-      // `{type:"custom", custom:{name}}`, so these reads are not a legacy
-      // leftover even though the tools loop above dropped its own nested path.
+      // `{type:"custom", custom:{name}}`, so these nested reads are required
+      // even though the tools loop above has no nested path.
       const fn = value.function;
       const nested = fn && typeof fn === "object" && !Array.isArray(fn) ? fn as Record<string, unknown> : undefined;
       const custom = value.custom;

package/src/providers/integrations/hermes.ts

@@ -28,17 +28,15 @@ const HERMES_PROVIDER_ID = "hermes" as const;
  * (`agent/portal_tags.py` in `NousResearch/hermes-agent`). Neither is required
  * for dispatch — a request carrying only `user=` answers 200 — but they are
  * sent because they are the client's real attribution and cost nothing. The
- * client version tracks the released tag `v2026.9.24`, which is date-based:
- * Hermes does not publish a semver, so the `0.13.0` this used to claim was not
- * a version Hermes ever had.
+ * client version is read from `getHermesVersion()` and sent in the shape the
+ * released tags use (`v<version>`): Hermes does not publish a semver, so its
+ * versions are date-based release tags.
  */
 /**
  * Builds the portal tags for one dispatch.
  *
- * `client=` carries the version in the shape the released tag used
- * (`v<version>`), but the version itself is now the resolved one: the previous
- * constant named a date tag (`v2026.9.24`) that was neither a release nor the
- * version the project publishes. `getHermesVersion()` tracks the real release.
+ * `client=` carries the resolved version in the shape the released tags use
+ * (`v<version>`); `getHermesVersion()` tracks the real release.
  */
 function hermesPortalTags(): readonly string[] {
   return [
@@ -66,25 +64,23 @@ import type { ModelDefinition } from "../provider-model-definition";
  *
  * `loadModelDiscovery` supersedes this list whenever `/v1/models` answers, so
  * these numbers are only read by a fetch that failed. They are the upstream's
- * own `context_length` and `top_provider.max_completion_tokens` — the previous
- * values were a uniform `8192` output for every row, which understated every
- * one of them.
+ * own `context_length` and `top_provider.max_completion_tokens`, stated per row
+ * because they differ per row.
  *
- * Two ids that used to live here are gone because upstream no longer carries
- * them: `tencent/hy3:free` (absent from the catalog entirely) and
- * `upstage/solar-pro4:free` (only the billed `upstage/solar-pro4` and the free
- * `upstage/solar-mini4:free` exist).
+ * Two ids upstream no longer carries are absent: `tencent/hy3:free` (missing
+ * from the catalog entirely) and `upstage/solar-pro4:free` (only the billed
+ * `upstage/solar-pro4` and the free `upstage/solar-mini4:free` exist).
  */
 export const HERMES_MODELS: readonly ModelDefinition[] = [
   // The live `/models` free rows, with the limits the catalog states. Verified
   // against the API: every id here answers 200.
   //
-  // `meituan/longcat-2.0:free` is gone: the catalog still lists the id, but
-  // dispatching it answers 404 "This model is no longer free", so keeping it
+  // `meituan/longcat-2.0:free` is excluded: the catalog still lists the id, but
+  // dispatching it answers 404 "This model is no longer free", so listing it
   // would be a card that can only ever fail. `meituan/longcat-2.5-preview:free`
-  // replaces it and answers 200.
+  // is listed instead and answers 200.
   //
-  // `stepfun/step-5-preview:free` is gone: it is absent from the live catalog
+  // `stepfun/step-5-preview:free` is excluded: it is absent from the live catalog
   // entirely and dispatching it answers 404 "This model's free period has
   // ended" — a retired row, not a capacity blip, so it cannot be left in.
   //

package/src/providers/integrations/kiro/kiro-discovery.ts

@@ -31,7 +31,7 @@ const MAX_DISCOVERED_MODELS = 500;
 const KIRO_EFFORTS = ["low", "medium", "high", "xhigh"] as const;
 
 /**
- * Ids the seed catalog deliberately no longer carries.
+ * Ids the seed catalog deliberately excludes.
  *
  * Discovery adds every id the upstream lists, so without this a retired row
  * comes back the moment an account fetches its models. The seed is the owner

package/src/providers/integrations/kiro/kiro.ts

@@ -8,7 +8,7 @@
 // Two upstream behaviours drive the header and endpoint logic:
 //
 //  * There is exactly one generation surface, `q.{region}.amazonaws.com`. The
-//    legacy `codewhisperer.*` host and the vendor's own path-style gateway are
+//    `codewhisperer.*` host and the vendor's own path-style gateway are
 //    not what the real client talks to, and reaching for them is actively
 //    harmful: an account that has just been refused would immediately probe two
 //    other service entries with the same token, which is the shape of credential

package/src/providers/integrations/mistral.ts

@@ -6,12 +6,12 @@ const MISTRAL_PROVIDER_ID = "mistral" as const;
 /**
  * Mistral's own current generalist tiers.
  *
- * The provider is registered as a generic OpenAI-compatible API-key provider, so
- * it carried no catalog at all and every model it serves was absent until a
- * discovery run happened to fetch them. These are the two `latest` aliases
- * Mistral keeps pointed at its current generalist models — `mistral-large-latest`
- * (Large 4) and `mistral-medium-latest` (Medium 3.5), per the Mistral model
- * overview.
+ * The provider is registered as a generic OpenAI-compatible API-key provider,
+ * which carries no catalog of its own: without this list, every model it serves
+ * stays absent until a discovery run happens to fetch them. These are the two
+ * `latest` aliases Mistral keeps pointed at its current generalist models —
+ * `mistral-large-latest` (Large 4) and `mistral-medium-latest` (Medium 3.5),
+ * per the Mistral model overview.
  *
  * Limits and pricing come from the base catalog via `providerId`: `defineModel`
  * resolves `mistral:<id>` and falls back to the majority row, so no number is

package/src/providers/integrations/opencode/fingerprint.ts

@@ -23,11 +23,11 @@ const SESSION_COUNTER_MAX = 63;
  * `ses_` + 12-char hex (bitwise inverted millisecond timestamp * 4096 + seq) + 14-char base62.
  *
  * The counter is capped rather than left to grow. Sixty-four dispatches inside
- * one millisecond used to emit ids the gate refuses, so a burst — exactly what
- * a proxy pool produces — 403'd on its own headers while every other field was
+ * one millisecond emit ids the gate refuses, so a burst — exactly what a proxy
+ * pool produces — would 403 on its own headers while every other field was
  * correct. On overflow the timestamp advances by one millisecond instead: the
  * id stays inside the window, and the only cost is that ids minted in the same
- * real millisecond no longer sort as the client's would.
+ * real millisecond do not sort as the client's would.
  */
 export function generateOpenCodeSessionId(now = Date.now()): string {
   if (now !== lastTimestamp) {

package/src/providers/integrations/opencode/provider.ts

@@ -79,7 +79,7 @@ const UNAVAILABLE_FREE_ZEN_IDS: readonly string[] = [
  * `CARTETHYIA_OPENCODE_UNAVAILABLE_FREE` lets an operator correct that
  * immediately: a comma-separated list that *replaces* the bundled one, or the
  * single token `none` to trust the listing completely. The bundled ids remain
- * the default so an unconfigured install behaves exactly as it does today.
+ * the default so an unconfigured install keeps these refusals.
  */
 function unavailableFreeZenIds(): readonly string[] {
   const override = resolveOpenCodeUnavailableFree();
@@ -163,11 +163,11 @@ function opencodeDesktopHeaders(
     // The free-tier gate refuses non-streaming dispatches with 403
     // FreeTierError, so free models always advertise SSE regardless of the
     // client's own Accept; billed models keep the JSON default. The value is
-    // what OpenCode 2.x's AI SDK sends, not the narrower
-    // `text/event-stream` we used to claim — the gate always accepted the
-    // narrow one, so this is alignment with the real client rather than a
-    // fix for a live failure. It costs nothing: `stream: true` is forced on
-    // the body either way, so the upstream answers SSE regardless.
+    // what OpenCode 2.x's AI SDK sends — a broader Accept than the narrower
+    // `text/event-stream` the gate is known to accept, so this is alignment
+    // with the real client rather than a fix for a live failure. It costs
+    // nothing: `stream: true` is forced on the body either way, so the upstream
+    // answers SSE regardless.
     ...(freeModel ? { accept: FREE_TIER_ACCEPT } : {}),
   };
 }
@@ -498,7 +498,7 @@ export function systemoneLimits(): Pick<ModelDefinition, "contextLimit" | "outpu
 export const OPENCODE_FREE_MODELS: readonly ModelDefinition[] = [
   ...sharedZenChat,
   // `nemotron-3-ultra-free`, `nemotron-3.5-lightning-free` and
-  // `longcat-2.5-preview-free` are gone: the live Zen roster no longer lists
+  // `longcat-2.5-preview-free` are excluded: the live Zen roster does not list
   // them. A row the provider's own roster dropped can disappear from the wire
   // without notice, which makes it a stale row — one that routes to a 404 at
   // dispatch time rather than failing at catalog load. `big-pickle` and the
@@ -514,7 +514,7 @@ export const OPENCODE_ZEN_MODELS: readonly ModelDefinition[] = [
 ];
 
 export const OPENCODE_GO_MODELS: readonly ModelDefinition[] = [
-  // `muse-spark-1.2` is retired upstream; the catalog files
+  // `muse-spark-1.2` is not served upstream; the catalog files
   // `muse-spark-1.2-contributor` at 1048576/131072, which is what this row
   // serves.
   defineModel({ id: "muse-spark-1.2-contributor", providerId: "opencode-go", wireFamily: "responses", endpoint: `${GO_PATH_PREFIX}/responses`, ctx: 1048576, out: 131072, vision: true, reasoning: true }),