TheStreamCode · Oct 14, 2026 · Oct 14, 2026 · Oct 14, 2026 · Oct 14, 2026 · Oct 14, 2026
@@ -48,7 +48,7 @@ src/router.mjs            The local proxy: model injection, provider routing, pa
 src/catalog.mjs           Turns providers into the model entries Copilot expects from GET /models
 src/launcher.mjs          Starts the router on an ephemeral port and spawns Copilot against it
 src/upstream.mjs          Resolves the Copilot API tier (individual / business / enterprise)
-src/discovery.mjs         Asks each provider for its live model list; falls back to the catalog
+src/discovery.mjs         Asks each provider for its live model list; falls back to the remembered list
 src/config.mjs            Provider config loading and validation; loads providers.default.json
 src/providers.default.json  Generated catalog — never edit by hand, run npm run catalog:update
 src/keystore.mjs          Optional on-disk key store (0600 / owner-only ACL)
@@ -60,8 +60,8 @@ src/copilot-bin.mjs       Resolves the Copilot executable, skipping the stale VS
 src/model-ranking.mjs     Pure ranking of provider model catalogs (legacy mode)
 src/process-env.mjs       Strips stale/secret variables from the child environment
 src/provider-env.mjs      Builds the COPILOT_PROVIDER_* variables (legacy mode)
-scripts/update-catalog.mjs  Rebuilds the catalog from models.dev
-scripts/catalog-sources.json  Curated provider/model selection feeding the generator
+scripts/update-catalog.mjs  Rebuilds the endpoint catalog (offline, no network)
+scripts/catalog-sources.json  Provider endpoints feeding the generator
 extensions/byok/extension.mjs  The /byok slash command, copied into ~/.copilot/extensions/
 test/*.test.mjs           One suite per src module
 schemas/providers.schema.json  Published JSON Schema for providers.json
@@ -80,7 +80,7 @@ All commands run from the repository root.
 | Test | `npm test` |
 | Test with coverage | `npm run test:coverage` |
 | Full quality gate | `npm run check` (lint + test) |
-| Verify curated models still exist | `npm run catalog:audit` |
+| Verify providers serve usable models | `npm run catalog:audit` |
 | Package contents check | `npm pack --dry-run` |
 | Dependency audit | `npm audit --omit=dev --audit-level=high` |
 | Local install for manual testing | `npm link` then `copilot-byok --help` |
@@ -132,20 +132,18 @@ This project handles third-party provider API keys. Every change must preserve t
 The catalog is generated. **Never hand-edit `src/providers.default.json`** — it is overwritten.
 
 1. Add the provider to `scripts/catalog-sources.json`, using only endpoints documented by the
-   provider, then run `npm run catalog:update`. List model ids under `models`; the generator fills
-   in context window and capabilities from models.dev and drops anything without tool-calling
-   support. Leave `models` empty when names are account-specific (Volcengine endpoint ids) or the
-   catalog is too large to curate (OpenRouter).
-2. Run `npm run catalog:audit` after changing any curated list. A model that does
-   not exist on the provider is worse than a missing one: it is offered to the user
-   and fails only when picked.
+   provider, then run `npm run catalog:update`. List no models: the picker is filled
+   by live discovery from the provider's `GET /models`.
+2. Run `npm run catalog:audit` after changing the catalog. A provider that answers
+   but yields no usable chat model fails the audit; providers without a configured
+   key are reported as unverifiable.
 3. Verify the endpoint yourself before adding it: `POST {baseUrl}/chat/completions` with an invalid
    key must answer `400`/`401`/`403`, not `404`. A `404` usually means the path is wrong.
-3. For authenticated providers, add every credential environment name to `DEFAULT_SECRET_SOURCE_ENV`
+4. For authenticated providers, add every credential environment name to `DEFAULT_SECRET_SOURCE_ENV`
    in `src/process-env.mjs`. For intentionally authless providers set `authRequired: false` and do
    not invent a credential environment name.
-4. Add the row to the provider table in `README.md`.
-5. Update `docs/provider-verification.md` honestly: endpoint reachability, authenticated catalog
+5. Add the row to the provider table in `README.md`.
+6. Update `docs/provider-verification.md` honestly: endpoint reachability, authenticated catalog
    access, and end-to-end inference are three distinct evidence levels. Never claim a level that was
    not actually verified.
 

@@ -4,6 +4,40 @@ All notable changes to this project are documented in this file.
 
 ## [Unreleased]
 
+### Added
+
+- Meta Muse (`muse`) and NVIDIA NIM (`nim`) providers, both verified with
+  OpenAI-compatible chat and model-list endpoints.
+- Image-input support is now read from each provider's live catalog, the same
+  way reasoning support already was.
+- `--no-discovery` is now an offline mode: providers use their pinned models or
+  the list each key returned last time, without any network request.
+
+### Changed
+
+- The built-in catalog ships endpoints only; model lists always come from live
+  discovery. There are no shipped model lists to go stale.
+- Model ranking no longer prefers any model family; order is recency, version,
+  context window, then name.
+- `npm run catalog:update` is offline and instant: no more models.dev fetch.
+- `npm run catalog:audit` now checks that every provider serves at least one
+  usable chat model instead of checking pinned model ids.
+- `--list-providers` marks providers with a configured key as usable and shows
+  pinned or remembered model counts.
+
+### Removed
+
+- **Breaking:** `stepfun`, `doubao`, `hunyuan` and `qianfan` providers were
+  removed: the catalog keeps only coding-oriented frontier providers. Configs
+  referencing them need updating; account-specific endpoint ids still work
+  through a custom provider entry.
+
+### Fixed
+
+- The `/byok` in-session command lists every provider needing a credential
+  again instead of an empty list, and no longer reports model counts that only
+  exist after discovery.
+
 ## [1.4.2] - 2026-08-12
 
 ### Changed

@@ -100,18 +100,18 @@ Every provider below is reached through its OpenAI-compatible endpoint, verified
 |---|---|---|---|
 | OpenAI | `OPENAI_API_KEY` | DeepSeek | `DEEPSEEK_API_KEY` |
 | Anthropic | `ANTHROPIC_API_KEY` | Alibaba Qwen | `DASHSCOPE_API_KEY` |
-| Google Gemini | `GEMINI_API_KEY` | Z.AI GLM | `ZAI_API_KEY` |
-| xAI Grok | `XAI_API_KEY` | Zhipu AI (CN) | `ZHIPU_CN_API_KEY` |
-| Mistral | `MISTRAL_API_KEY` | Moonshot Kimi | `MOONSHOT_API_KEY` |
-| Groq | `GROQ_API_KEY` | MiniMax | `MINIMAX_API_KEY` |
-| Cerebras | `CEREBRAS_API_KEY` | StepFun | `STEPFUN_API_KEY` |
-| Together AI | `TOGETHER_API_KEY` | ByteDance Doubao | `ARK_API_KEY` |
-| DeepInfra | `DEEPINFRA_API_KEY` | Tencent Hunyuan | `HUNYUAN_API_KEY` |
-| Fireworks AI | `FIREWORKS_API_KEY` | Baidu Qianfan | `QIANFAN_API_KEY` |
-| OpenRouter | `OPENROUTER_API_KEY` | SiliconFlow | `SILICONFLOW_API_KEY` |
-| Chutes | `CHUTES_API_KEY` | ModelScope | `MODELSCOPE_API_KEY` |
-| OpenCode Zen | `OPENCODE_ZEN_API_KEY` | Alibaba Token Plan | `ALIBABA_TOKEN_PLAN_API_KEY` |
-| OpenCode Go | `OPENCODE_GO_API_KEY` | Ollama / LM Studio | no key needed |
+| Google Gemini | `GEMINI_API_KEY` | Alibaba Qwen (CN) | `DASHSCOPE_API_KEY_CN` |
+| xAI Grok | `XAI_API_KEY` | Z.AI GLM | `ZAI_API_KEY` |
+| Mistral | `MISTRAL_API_KEY` | Zhipu AI (CN) | `ZHIPU_CN_API_KEY` |
+| Meta Muse | `MODEL_API_KEY` | Moonshot Kimi | `MOONSHOT_API_KEY` |
+| NVIDIA NIM | `NVIDIA_API_KEY` | MiniMax | `MINIMAX_API_KEY` |
+| Groq | `GROQ_API_KEY` | SiliconFlow | `SILICONFLOW_API_KEY` |
+| Cerebras | `CEREBRAS_API_KEY` | ModelScope | `MODELSCOPE_API_KEY` |
+| Together AI | `TOGETHER_API_KEY` | OpenRouter | `OPENROUTER_API_KEY` |
+| DeepInfra | `DEEPINFRA_API_KEY` | Chutes | `CHUTES_API_KEY` |
+| Fireworks AI | `FIREWORKS_API_KEY` | OpenCode Zen | `OPENCODE_ZEN_API_KEY` |
+| Alibaba Token Plan | `ALIBABA_TOKEN_PLAN_API_KEY` | OpenCode Go | `OPENCODE_GO_API_KEY` |
+| Ollama | no key needed | LM Studio | no key needed |
 
 Each provider also accepts a `COPILOT_BYOK_<PROVIDER>_API_KEY` variant, useful when you already use the plain name for something else.
 
@@ -126,17 +126,17 @@ The store lives at `~/.config/copilot-byok/keys.json` (`%APPDATA%\copilot-byok\k
 
 Either way the keys stay out of the Copilot process: the router uses them in its own, and strips them from the environment it hands to the agent.
 
-Some entries ship without a preset model list, either because their model names are account-specific (Volcengine endpoint ids) or because the catalog is too large to curate (OpenRouter). Add the ones you want in your own config — see below.
+No entry ships a model list: the picker is filled by live discovery. To pin exact models instead — for an offline setup or a fixed line-up — add them in your own config — see below.
 
 ### Where the model list comes from
 
 **Providers are asked directly.** At startup each configured provider is queried
 for the models it serves right now, so the picker shows today's line-up rather
-than whatever was curated at release time — a model released this morning is
-there this morning. All 30 providers were verified to expose an OpenAI-style
-`GET /models`; the ones that also publish capabilities (Chutes, OpenRouter and
-others) supply the real context window, output limit and tool-calling support
-along with it.
+than whatever was pinned at release time — a model released this morning is
+there this morning. All 28 providers expose an OpenAI-style `GET /models`
+(verified in `docs/provider-verification.md`); the ones that also publish
+capabilities (Chutes, OpenRouter, NIM and others) supply the real context
+window, output limit and tool-calling support along with it.
 
 While this happens the startup shows what it is waiting on, since the pause comes
 before Copilot has drawn anything:
@@ -153,31 +153,28 @@ Only the providers you have a key for are queried, in parallel, with a hard
 timeout. Measured on three configured providers: about a second. Results are
 cached for ten minutes, failures for one.
 
-Anything the provider does not report is filled in from the shipped catalog,
-which is generated from [models.dev](https://models.dev) and refreshed with:
-
-```sh
-npm run catalog:update
-```
+Anything the provider does not report falls back to conservative defaults (128k
+context, 32k output); image input and the reasoning selector appear only when
+the provider's own metadata declares them.
 
 If a request fails — a rate limit, a blip, an expired key — the list that key
 returned last time is used instead, kept in `~/.cache/copilot-byok/models.json`
-(`%LOCALAPPDATA%` on Windows). That matters: the shipped list is generic, while a
-provider's catalog belongs to the plan behind your key, so falling back to it can
-offer models your plan does not include and hide ones it does. Only when nothing
-was ever remembered does the shipped list step in, so a hiccup never empties the
-picker. To skip discovery entirely and always use the shipped lists:
+(`%LOCALAPPDATA%` on Windows). That matters: a provider's catalog belongs to the
+plan behind your key, so only what your key actually returned is ever offered —
+never a generic guess. When nothing was ever remembered, the provider stays out
+of the picker and the log says why. To skip discovery entirely and work from
+remembered lists:
 
 ```sh
 copilot-byok --no-discovery
 ```
 
-Models that cannot call tools are excluded, whether that is known from the
-provider's own metadata or from the catalog: Copilot's harness trusts what a
-model entry declares, and an agent that cannot call tools breaks halfway through
-a session instead of failing cleanly. Embedding, image, audio and moderation
-models are filtered out for the same reason. Each provider contributes at most
-twelve models, so a catalog like OpenRouter's 400+ cannot swamp the picker.
+Models that cannot call tools are excluded based on the provider's own metadata:
+Copilot's harness trusts what a model entry declares, and an agent that cannot
+call tools breaks halfway through a session instead of failing cleanly.
+Embedding, image, audio and moderation models are filtered out for the same
+reason. Each provider contributes at most twelve models, so a catalog like
+OpenRouter's 460+ cannot swamp the picker.
 
 ## The /byok command
 
@@ -193,9 +190,9 @@ Then `/byok` opens a provider list and asks for the key:
 
 ```
 Choose a provider to configure:
-❯ Anthropic — not configured (4 models)
-  Google Gemini — not configured (3 models)
-  DeepSeek — not configured (4 models)
+❯ Anthropic — not configured
+  Google Gemini — not configured
+  DeepSeek — not configured
   OpenAI — set via environment
   Chutes — set via environment
 
@@ -356,13 +353,13 @@ Everything has a working default; these are for when you need to change one.
 ### Keeping the catalog honest
 
 ```sh
-npm run catalog:update    # refresh models and limits from models.dev
-npm run catalog:audit     # check every curated model still exists on its provider
+npm run catalog:update    # regenerate the endpoint catalog from scripts/catalog-sources.json
+npm run catalog:audit     # check every provider serves a usable model list
 ```
 
-The audit exits non-zero when a curated entry is not served by its provider.
-Curated lists are only a fallback, but a wrong entry is worse than a missing one:
-it is offered in the picker and fails only when selected.
+The audit queries each provider's live catalog and exits non-zero when one that
+answers yields no usable chat model. Providers that need a key you have not
+configured are reported as unverifiable rather than failing.
 
 ### Single-provider mode
 

@@ -1,6 +1,6 @@
 # Provider verification
 
-Last verified: 2026-08-12 (model-list endpoints re-verified the same day)
+Last verified: 2026-10-14 for endpoints and model lists; end-to-end runs below are from 2026-08-12.
 
 This document separates three different levels of evidence, because they are not
 interchangeable:
@@ -19,6 +19,8 @@ interchangeable:
 - GitHub Copilot CLI 1.0.79, personal (free) plan
 - Endpoint probes: `POST /chat/completions` with a deliberately invalid key and a
   25-second timeout. No credentials were printed.
+- Endpoint and model-list tables were re-probed on 2026-10-14 (Windows,
+  Node.js 24.21, invalid-key probes, no credentials).
 
 ## Endpoint reachability
 
@@ -45,15 +47,13 @@ All entries below answered on their OpenAI-compatible route.
 | Zhipu AI (CN) | `https://open.bigmodel.cn/api/paas/v4` | 401 |
 | Moonshot Kimi | `https://api.moonshot.ai/v1` | 401 |
 | MiniMax | `https://api.minimax.io/v1` | 401 |
-| StepFun | `https://api.stepfun.com/v1` | 401 |
+| Meta Muse | `https://api.meta.ai/v1` | 401 |
 | SiliconFlow | `https://api.siliconflow.com/v1` | 401 |
 | ModelScope | `https://api-inference.modelscope.cn/v1` | 400 |
-| ByteDance Doubao | `https://ark.cn-beijing.volces.com/api/v3` | 401 |
-| Tencent Hunyuan | `https://api.hunyuan.cloud.tencent.com/v1` | 401 |
-| Baidu Qianfan | `https://qianfan.baidubce.com/v2` | 401 |
 | Baichuan | `https://api.baichuan-ai.com/v1` | 401 |
 | Chutes | `https://llm.chutes.ai/v1` | 401 |
 | OpenCode Zen | `https://opencode.ai/zen/v1` | 401 |
+| OpenCode Go | `https://opencode.ai/zen/go/v1` | 400 |
 | Baseten | `https://inference.baseten.co/v1` | 403 |
 
 Two endpoints answered `404` to the probe because they validate the model name
@@ -67,40 +67,46 @@ before the credential, and the probe deliberately sends a non-existent model:
 ## Model-list endpoints
 
 Discovery reads each provider's `GET /models`, so that route was probed for every
-catalog entry with an invalid key. **None returned 404** — every provider exposes
-it. Four serve it without authentication, which is how the model counts below
-were observed:
+catalog entry. Sending no credentials returns the public catalogs below;
+sending an invalid key is rejected instead, so a 401 proves the route exists
+but says nothing about catalog contents. Observed 2026-10-14 via
+`npm run catalog:audit` with no keys configured:
 
 | Provider | Result |
 |---|---|
-| ModelScope | 200 — 43 models |
-| OpenCode Go | 200 — 25 models |
-| OpenCode Zen | 200 — 61 models |
-| OpenRouter | 200 — 408 models, 340 declaring tool support |
-| Gemini, xAI | 400 (route exists, request rejected) |
-| Qianfan | 403 |
+| Chutes | 200 — 14 models, all usable |
+| DeepInfra | 200 — 179 models, 137 usable after ranking |
+| ModelScope | 200 — 36 models, 34 usable |
+| Nvidia NIM | 200 — 80 models, 69 usable |
+| OpenCode Go | 200 — 40 models, all usable |
+| OpenCode Zen | 200 — 85 models, all usable |
+| OpenRouter | 200 — 462 models, 447 usable |
+| Gemini | 404 without a key (400 with an invalid one) |
+| xAI | 400 (route exists, request rejected) |
 | All others | 401 |
 | Ollama, LM Studio | connection refused — local servers, not running here |
 
 Some providers publish the metadata discovery needs alongside the list, which is
-what lets the picker show real limits rather than guesses:
+what lets the picker show real limits rather than defaults:
 
 | Provider | Fields observed |
 |---|---|
 | Chutes | `context_length`, `max_output_length`, `supported_features: ["json_mode","tools","structured_outputs","reasoning"]` |
 | OpenRouter | `context_length`, `top_provider.max_completion_tokens`, `supported_parameters` including `tools` |
+| Nvidia NIM | ids only (`id`, `created`, `owned_by`) |
+| ModelScope | ids only (`id`, `created`, `owned_by`) |
 
-Providers that publish only ids still work: their limits come from the shipped
-catalog, and a model is excluded only when a provider explicitly says it cannot
-call tools.
+Providers that publish only ids still work: their limits fall back to
+conservative defaults (128k context, 32k output), and a model is excluded only
+when a provider explicitly says it cannot call tools.
 
 ### Discovery against real keys
 
 | Provider | Result |
 |---|---|
 | Chutes | 12 models, with context and output read from the API |
 | Alibaba Token Plan | 7 models |
-| OpenAI | 401 — the key on this machine is not valid for `api.openai.com`; the curated list was used, as designed |
+| OpenAI | 401 — the key on this machine is not valid for `api.openai.com`; nothing remembered, so the provider stayed out of the picker, as designed |
 
 Measured wall-clock for the three configured providers, queried in parallel: 1.2s.
 

@@ -85,9 +85,8 @@ The list comes from the provider itself. Three things can hide a model:
   chosen by the same ranking used elsewhere. Raise it with `maxDiscoveredModels`
   in your config, or pin the exact models you want under `models`.
 - **Discovery fell back.** If the provider was unreachable or rejected the key,
-  the list it returned last time is used; failing that, the shipped one. The log
-  says which and why. A shipped list is generic, so it may name models your plan
-  does not include — if you see one that should not be there, that is the sign.
+  the list it returned last time is used; failing that, the provider stays out
+  of the picker. The log says which and why.
 
 Results are cached for ten minutes, so a model added upstream may take that long
 to appear — restart to pick it up immediately.
@@ -101,16 +100,15 @@ provider actually offers:
 copilot-byok --legacy --provider <id> --list-models
 ```
 
-Then update `models` in your config, or run `npm run catalog:update` if you use
-the built-in catalog.
+Then pin the working names under `models` in your own config file.
 
 ## The agent starts but fails mid-task
 
 Almost always a capability mismatch: the model does not really support tool
-calling, and the harness discovers it once it tries to use a tool. The generated
-catalog only publishes models that report tool support, but a hand-written config
-entry can claim anything. Verify the model supports function calling and
-streaming before adding it.
+calling, and the harness discovers it once it tries to use a tool. Discovery
+only publishes models that do not rule out tool support, but a hand-written
+config entry can claim anything. Verify the model supports function calling and
+streaming before pinning it.
 
 ## Copilot CLI is not found