Expand and verify provider presets

[jellologic]

Goal

Ship verified presets for every provider exposing an Anthropic-compatible /v1/messages endpoint. We currently have three (Anthropic, z.ai, OpenRouter) plus a custom escape hatch.

Ground rule

Every preset must be verified against primary sources, not blog posts. The research pass found widely-repeated claims that are simply false and actively break users. Each preset needs: exact base URL (paths differ — some bare host, some /anthropic, some /v1), which env var carries the credential, real model IDs, and confirmed tool-calling support.

Candidates

Verified so far:

ModelScope — https://api-inference.modelscope.cn (bare host; the /v1 on the same doc page is the OpenAI route, a common mistake). Token keeps its ms- prefix — the widespread "strip the prefix" advice is wrong and breaks auth. 2,000 calls/day free, requires a real-name-verified Alibaba Cloud link. No prompt caching (verified: cache_control is silently discarded). Tool calling is per-model and undocumented; maintainer confirmed it does not work for deepseek-v3.1 or kimi-k2. Labeled beta, and ModelScope never documents Claude Code at all.

⚠️ Serious quirk: with stream: true (which Claude Code always uses), an invalid token or exhausted quota returns HTTP 200 with a silent empty reply rather than an error. Users will see a hang, not a message. If we ship this preset, doctor (#PLACEHOLDER_8) must probe non-streaming to get a real status code.

Rejected:

iFlow — a /v1/messages route provably exists on apis.iflow.cn but is entirely undocumented, unused publicly, and missing count_tokens. Decisive blocker: API keys expire after 7 days, which breaks any persisted config by design. Not worth a preset.

Still to assess (survey in flight): Moonshot/Kimi, DeepSeek, Groq, Cerebras, Together, Fireworks, Nebius, DeepInfra, local runners (Ollama / LM Studio / llama.cpp / vLLM), and gateways (LiteLLM, Portkey, Helicone, Cloudflare AI Gateway).

Separately: Bedrock, Vertex and Foundry are not base-URL providers. Claude Code has native support via CLAUDE_CODE_USE_BEDROCK / USE_VERTEX / USE_FOUNDRY, each with its own required vars (AWS_REGION, ANTHROPIC_VERTEX_PROJECT_ID, ANTHROPIC_FOUNDRY_RESOURCE) and provider-specific model IDs — Bedrock needs full inference-profile ARNs. These need a different preset shape and probably belong in their own issue.

Acceptance criteria

• Each preset cites the primary source it was verified against
• Each records: base URL, credential env var, model IDs, tool support, caching support, free-tier terms
• Known-broken combinations are marked, not silently offered
• Providers with disqualifying quirks are documented as rejected, with the reason
• Presets carry a verifiedOn date so staleness is visible

---

Addendum — provider survey results (2026-07-21)

SiliconFlow — https://api.siliconflow.com/ intl, https://api.siliconflow.cn/ CN. Bare host. Verified trap: https://api.siliconflow.com/v1 is the OpenAI URL and yields /v1/v1/messages → 404. Models zai-org/GLM-5.2, MiniMaxAI/MiniMax-M3; a Pro/ prefix marks the paid variant, unprefixed is free tier at 1,000 RPM. Docs say ANTHROPIC_API_KEY, but both headers work — prefer ANTHROPIC_AUTH_TOKEN, which avoids Claude Code's one-time interactive approval prompt for ANTHROPIC_API_KEY.

Volcengine / Doubao ARK — recommend skipping. Two products with non-interchangeable keys (/api/coding vs /api/plan; a /v3 suffix marks the OpenAI variant). Three reasons for caution:

1. Probe could not confirm routing — ARK returns identical 401s on nonsense control paths, so this rests on docs alone unlike every other entry here.
2. Pointing at /api/v3 silently bills pay-as-you-go instead of consuming plan quota.
3. Docs reportedly warn that driving these endpoints from your own scripts may be flagged as abuse and lead to account suspension — and a launcher is exactly the "own script" shape. Shipping a preset that can get users banned is not worth the breadth.

iFlow — confirmed do-not-ship.

⚠️ Hold before finalising z.ai and DeepSeek

Two findings are still outstanding from the survey and both are likely to cause user-facing bugs:

• A [1m] / auto-compact-window clamping interaction. Directly relevant to z.ai: glm-5.2[1m] needs CLAUDE_CODE_AUTO_COMPACT_WINDOW, but there may be a clamping floor that makes the naive value not behave as expected. This also affects Set context-window env vars from catalog data #2.
• A DeepSeek role: "system" breakage.

Do not finalise either preset until these land.

[jellologic]

Hold resolved — both outstanding findings landed.

1. [1m] clamping → split out as #14 since it is a live bug in the shipped z.ai preset, not just a planning item.

2. DeepSeek role: "system" — do not ship a DeepSeek direct preset yet.

api.deepseek.com/anthropic returns 400 unknown variant "system". Opus 4.8 introduced mid-conversation system messages (role: "system" inside messages[] rather than only the top-level system field); Claude Code began emitting these at v2.1.154, and DeepSeek's compatibility layer rejects the schema. This is outright request rejection, not degraded output, and it affects current Claude Code by default.

• DeepSeek Anthropic endpoint: 400 error on role: "system" after v2.1.154 deepseek-ai/awesome-deepseek-agent#167
• Anthropic endpoint rejects Opus 4.8 mid-conversation system messages deepseek-ai/awesome-deepseek-integration#639

No verified workaround. Specifically do not claim CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1 fixes it — that strips anthropic-beta headers and beta tool-schema fields, whereas this is message-array structure. Very likely inapplicable. Version-pinning below v2.1.154 is the only mitigation identified, and that is inferred rather than doc-backed.

Scope: this is DeepSeek's own endpoint. DeepSeek weights served via OpenRouter/DeepInfra/Novita go through those providers' translation layers and are not affected — marked INFERRED by the researcher, reasoned from the failure mechanism, not tested.

Adjacent, also unfixable by env var: the Bash security classifier sends a non-streaming request; deepseek-v4-pro defaults to thinking-on; TTFB of 28–32s against a ~30s timeout stalls Bash commands (DeepSeek-V3#1464).

3. Volcengine — stale config is circulating widely. kimi-k2.5, doubao-seed-code, deepseek-v3.2, glm-4.7 are all wrong. Reinforces skipping it.

4. Two gaps needing a live key (~2 min each for whoever has one):

• Are SiliconFlow .cn and .com keys interchangeable? Differing error formats suggest separate backends; no doc states it.
• Does iFlow's /v1/messages genuinely speak Anthropic format? Its auth gate fires before protocol handling, so this is unknowable externally. Recommendation stands: do not ship.