Skip to content

Claude Code. Any Model. The most powerful AI coding agent now speaks every language.

fix: ask an older Ollama daemon for capabilities it lists only in /api/show

Jack Rudenko50c7f0ed71ab
diff --git a/ai-docs/architecture/providers/qwen-alibaba.md b/ai-docs/architecture/providers/qwen-alibaba.md
index 36bcee59..a3bc133f 100644
--- a/ai-docs/architecture/providers/qwen-alibaba.md
+++ b/ai-docs/architecture/providers/qwen-alibaba.md
@@ -177,8 +177,10 @@ The vendor docs do carry capability hints in prose ("qwen3.7-plus (vision)",
`happyhorse-1.1-t2v` = text-to-video), but transcribing those into source is the
hardcoded model list the repo invariant forbids. `NON_CHAT_PATTERNS` caught every non-chat
id on both live lists and missed exactly three, all video generators on editions the
-measured account does not hold; only a cross-edition check found them. The video name
-patterns in `providers/transport/probe-discovery.ts` cover them now.
+measured account does not hold; only a cross-edition check found them. The name patterns
+are gone since v10.1.1: chat capability now comes from the provider's own listing or the
+catalog's published modalities (text among inputs AND outputs), and an id nothing
+describes is not offered at all, so an undescribed video generator no longer slips through.
## Measurement traps, all of which cost real time
diff --git a/packages/cli/src/model-selector.test.ts b/packages/cli/src/model-selector.test.ts
index d7df467d..4267d54a 100644
--- a/packages/cli/src/model-selector.test.ts
+++ b/packages/cli/src/model-selector.test.ts
@@ -34,6 +34,7 @@ import {
type DiscoveredModel,
type DiscoveryFailureKind,
discoverProviderModels,
+ discoverProviderModelsCatalog,
getDiscoveryFailure,
invalidateModelDiscovery,
registerModelDiscoveryFetcher,
@@ -43,6 +44,7 @@ import { BUILTIN_PROVIDERS, getProviderByName } from "./providers/provider-defin
import { clearRuntimeRegistry, registerRuntimeProvider } from "./providers/runtime-providers.js";
import {
_clearChatCapabilityIndex,
+ isReportedChatCapable,
ollamaReported,
} from "./providers/transport/probe-discovery.js";
import { discoverProviderProbeModel } from "./providers/transport/provider-model-discovery.js";
@@ -268,43 +270,97 @@ describe("provider-reported chat capability in discovered model callers", () =>
});
describe("Ollama discovery capability reports", () => {
- const realFetch = globalThis.fetch;
+ let realFetch: typeof globalThis.fetch;
+ let realOllamaHost: string | undefined;
beforeEach(() => {
+ realFetch = globalThis.fetch;
+ realOllamaHost = process.env.OLLAMA_HOST;
+ process.env.OLLAMA_HOST = "http://ollama-picker.test";
invalidateModelDiscovery("ollama");
});
afterEach(() => {
globalThis.fetch = realFetch;
+ if (realOllamaHost === undefined) delete process.env.OLLAMA_HOST;
+ else process.env.OLLAMA_HOST = realOllamaHost;
invalidateModelDiscovery("ollama");
});
- test("the Ollama fetcher carries completion reports and preserves an absent capability", async () => {
- globalThis.fetch = mock(
- async () =>
- new Response(
- JSON.stringify({
- models: [
- { name: "llama3.2:3b-q4_K_M", capabilities: ["completion", "tools"] },
- { name: "granite3.2-vision:2b-instruct-q4_K_M" },
- ],
- }),
- { status: 200 }
- )
- ) as unknown as typeof fetch;
-
- expect(await discoverProviderModels("ollama")).toEqual([
+ test("enriches missing tag capabilities before the picker filters discovered models", async () => {
+ const showRequests: string[] = [];
+ globalThis.fetch = mock(async (input: string | URL | Request, init?: RequestInit) => {
+ const url = new URL(input instanceof Request ? input.url : String(input));
+ if (url.pathname === "/api/tags") {
+ return Response.json({
+ models: [{ name: "llama3.2:3b" }, { name: "nomic-embed-text:latest" }],
+ });
+ }
+ if (url.pathname === "/api/show") {
+ const body = JSON.parse(String(init?.body)) as { name: string };
+ showRequests.push(body.name);
+ return Response.json({
+ capabilities: body.name === "llama3.2:3b" ? ["completion", "tools"] : ["embedding"],
+ });
+ }
+ throw new Error(`Unexpected fetch: ${url}`);
+ }) as unknown as typeof fetch;
+
+ const outcome = await discoverProviderModelsCatalog("ollama");
+
+ expect(showRequests.sort()).toEqual(["llama3.2:3b", "nomic-embed-text:latest"].sort());
+ expect(outcome.kind).toBe("served");
+ if (outcome.kind !== "served") throw new Error("unreachable");
+ expect(outcome.models).toEqual([
{
- id: "llama3.2:3b-q4_K_M",
- displayName: "llama3.2:3b-q4_K_M",
+ id: "llama3.2:3b",
+ displayName: "llama3.2:3b",
supportsTools: true,
reported: "chat",
},
+ ]);
+ expect(
+ outcome.models.filter((model) => isReportedChatCapable(model.id, model.reported))
+ ).toEqual(outcome.models);
+ expect(outcome.models.map((model) => model.id)).not.toContain("nomic-embed-text:latest");
+ });
+
+ test("uses inline tag capabilities without requesting per-model details", async () => {
+ let showRequests = 0;
+ globalThis.fetch = mock(async (input: string | URL | Request) => {
+ const url = new URL(input instanceof Request ? input.url : String(input));
+ if (url.pathname === "/api/tags") {
+ return Response.json({
+ models: [
+ { name: "llama3.2:3b", capabilities: ["completion", "tools"] },
+ { name: "qwen2.5-coder:7b", capabilities: ["completion"] },
+ ],
+ });
+ }
+ if (url.pathname === "/api/show") {
+ showRequests++;
+ return Response.json({ capabilities: ["completion"] });
+ }
+ throw new Error(`Unexpected fetch: ${url}`);
+ }) as unknown as typeof fetch;
+
+ const outcome = await discoverProviderModelsCatalog("ollama");
+
+ expect(showRequests).toBe(0);
+ expect(outcome.kind).toBe("served");
+ if (outcome.kind !== "served") throw new Error("unreachable");
+ expect(outcome.models).toEqual([
{
- id: "granite3.2-vision:2b-instruct-q4_K_M",
- displayName: "granite3.2-vision:2b-instruct-q4_K_M",
+ id: "llama3.2:3b",
+ displayName: "llama3.2:3b",
+ supportsTools: true,
+ reported: "chat",
+ },
+ {
+ id: "qwen2.5-coder:7b",
+ displayName: "qwen2.5-coder:7b",
supportsTools: false,
- reported: undefined,
+ reported: "chat",
},
]);
});
diff --git a/packages/cli/src/model-selector.ts b/packages/cli/src/model-selector.ts
index 0f33a086..e94d202c 100644
--- a/packages/cli/src/model-selector.ts
+++ b/packages/cli/src/model-selector.ts
@@ -472,9 +472,11 @@ function sortModelsNewestFirst(models: ModelInfo[]): ModelInfo[] {
* accepts text and writes text is a chat model.
*
* What is left to watch is the catalog being WRONG, so a row that survives but
- * whose name reads like a generator — `seedance-2.5`, `flux-2-pro` — is named
- * ONCE per process in the debug log. That decides nothing; it points at a
- * catalog row to correct in models-index.
+ * whose name reads like a generator family (seedance, flux, veo, imagen) is
+ * named ONCE per process in the debug log. It decides nothing. Measured on
+ * 2026-09-23, 3 of the 53 such rows publish `in: ["text"]`, `out: ["text"]`
+ * (`flux-dev-finetuner`, `flux-fill`, `seedance-2.5-el`) — models-index rows
+ * that need correcting, which is what this line is for.
*/
export function toPickerRows(models: ModelInfo[]): ModelInfo[] {
const kept = sortModelsNewestFirst(dedupeModels(models.filter((m) => isChatCapable(m.id))));
@@ -1470,7 +1472,8 @@ async function pickModelFromList(
* - **Not everything served is chat.** Alibaba's plan host answers with image
* and TTS models alongside chat ones. Filtered with the SAME predicate the
* probe path uses, so the picker and `--probe` agree on what is chat-capable
- * — a name-based rule, never a model-id skip list. An all-non-chat dynamic models catalog
+ * — evidence from the provider's own listing or the catalog's modalities,
+ * never a name rule or a model-id skip list. An all-non-chat dynamic models catalog
* returns [] so the caller falls through to the catalog / free-text path
* instead of showing an empty list.
* - **Neither price nor (always) context is reported.** Context comes from the
diff --git a/packages/cli/src/providers/model-discovery-builtins.ts b/packages/cli/src/providers/model-discovery-builtins.ts
index 6dff1e2c..958eaf75 100644
--- a/packages/cli/src/providers/model-discovery-builtins.ts
+++ b/packages/cli/src/providers/model-discovery-builtins.ts
@@ -203,10 +203,12 @@ async function fetchOllamaModelsCatalog(): Promise<FetcherResult> {
const { ollamaReported } = await import("./transport/probe-discovery.js");
const endpoint = `${ollamaBaseUrl()}/api/tags`;
try {
- const installed = await fetchOllamaModels({
- enrichCapabilities: false,
- throwOnError: true,
- });
+ // `enrichCapabilities` is left at its default. It used to be `false`, which
+ // cost nothing while an undescribed model still counted as chat — and hid
+ // every model from a daemon that lists `capabilities` only in `/api/show`
+ // once it stopped. A current daemon lists them inline, so the per-model
+ // `/api/show` fallback never fires there.
+ const installed = await fetchOllamaModels({ throwOnError: true });
return {
kind: "models",
endpoint,
diff --git a/packages/cli/src/providers/transport/anthropic-compat.ts b/packages/cli/src/providers/transport/anthropic-compat.ts
index 6c5fed1c..811caed4 100644
--- a/packages/cli/src/providers/transport/anthropic-compat.ts
+++ b/packages/cli/src/providers/transport/anthropic-compat.ts
@@ -118,7 +118,8 @@ export class AnthropicProviderTransport implements ProviderTransport {
* alphabetical tiebreak) — the same order the picker defaults to, so the
* probe exercises the model the user would actually get.
*
- * Non-chat rows are dropped with the shared `isChatCapable()` filter rather
+ * Non-chat rows are dropped by `discoverProviderProbeModel`, with the shared
+ * `isReportedChatCapable()` filter the picker uses too, rather
* than a per-provider skip list: qwen-token-plan's dynamic models catalog mixes
* image/TTS models (`wan2.7-image`, `qwen-audio-3.0-tts-plus`) in with the text models, and
* hardcoding model ids here would rot the moment Alibaba ships the next one.
diff --git a/packages/cli/src/providers/transport/probe-discovery.ts b/packages/cli/src/providers/transport/probe-discovery.ts
index f7b0926a..c6c9126f 100644
--- a/packages/cli/src/providers/transport/probe-discovery.ts
+++ b/packages/cli/src/providers/transport/probe-discovery.ts
@@ -81,10 +81,11 @@ function isSmallName(name: string): boolean {
*
* Two sources of evidence, and no guesses: the provider's own published
* capability for THIS deployment, and the cloud models catalog's published
- * output modality for the canonical model.
+ * input and output modalities for the canonical model — text must be among
+ * both.
*
- * - `"not-chat"` — a provider or catalog statement that it produces something
- * other than text, or a LiteLLM wildcard route. Never offered.
+ * - `"not-chat"` — a provider or catalog statement that it produces no text, or
+ * takes no text in, or a LiteLLM wildcard route. Never offered.
* - `"chat"` — a provider or catalog statement that it answers chat turns.
* - `"unknown"` — NOBODY said. Not offered either, and that is the point: an
* unknown used to be treated as chat, so an embedding model nothing described
@@ -157,8 +158,8 @@ function indexOutputModality(
* File a catalog row's keys when its published INPUT modalities leave out text.
*
* A chat model takes text in and gives text out; other modalities on either side
- * never exclude one, so `["file","image","text"]` (every Claude model) is a chat
- * model. What this catches is the other case: `gemini-3.5-transcribe` publishes
+ * never exclude one, so `["file","image","text"]` and `["image","text"]` (two of
+ * the lists Claude rows publish) are chat models. What this catches is the other case: `gemini-3.5-transcribe` publishes
* `in: ["audio"]`, `out: ["text"]` — it writes text, but nothing typed reaches it.
* Measured on the live catalog: 13 rows publish text output with no text input,
* all ASR, captioning or live-translation models.