Provider setup and conformance matrix

This is the user-facing mirror of the 70-entry PROVIDERS catalog in src/providers.rs. The base URL, request style, default model, and environment variable below are the exact current catalog values, not recommendations inferred from a vendor's marketing page.

Set the listed environment variable and use /login <id> to select the row.

/login also pins one entry that is deliberately not in this catalog: TypeSafe · Jev (id typesafe), shown first with its own borders. It is a credential that gives every provider typed judgments, not a chat model, so selecting it stores a key and never becomes your active provider. It can also run with no key at all against a local engine (NUR_JEV_LOCAL_URL; see jev-local.md). See typesafe.md. Enclave · security agents (id enclave) is pinned beside it on the same terms: Enclave serves MCP, not chat, so its key (ENCLAVE_MCP_API_KEY) gives every provider Enclave's security agents through the enclave tool. See enclave.md. R means Nur sends the Responses shape (/responses); CC means Chat Completions (/chat/completions); AM means the native Anthropic Messages shape (/v1/messages); and GCC means Google's Cloud Code native v1internal:streamGenerateContent path. A vendor's OpenAI-compatible claim is not, by itself, evidence that every model supports tools, streaming, or authoritative usage. Unless the caveat says otherwise, verify those three features for the selected model and account before relying on them for an agent/budget workflow.

Official links are first-party documentation only. Not established is intentional: it means the linked public documentation does not establish the capability for this exact Nur route/model, so Nur must treat it as unverified.

ID / provider	Auth, exact route, default model	First-party docs	Tool / stream / usage caveat	Setup note
meta - Meta Model API	META_API_KEY / MODEL_API_KEY; https://api.meta.ai/v1; R; muse-spark-1.2	Meta Model API models	Muse Spark is multimodal (text, image, video, PDF); verify tools/stream/usage for your account.	Catalog default vendor. Browser: Muse Code (muse login) or paste a key. /model lists muse-spark-1.1, muse-spark-1.2, muse-spark-1.2-contributor.
openai - OpenAI	OPENAI_API_KEY; https://api.openai.com/v1; R; gpt-5.5	OpenAI API	Official Responses docs cover streaming, tools, and usage, subject to model availability.	API key or Nur's ChatGPT OAuth login.
openai-cc - OpenAI (Chat Completions)	OPENAI_API_KEY; https://api.openai.com/v1; CC; gpt-5.5	Chat Completions	Chat route is retained for compatibility; feature availability is model-specific.	Same key as openai; prefer openai unless a Chat-only endpoint is required.
anthropic - Anthropic	ANTHROPIC_API_KEY; https://api.anthropic.com/v1; AM; claude-sonnet-5	Messages API	Official Messages docs cover streaming, tools, and usage.	API key or supported browser OAuth.
google - Google Gemini	GEMINI_API_KEY; https://generativelanguage.googleapis.com/v1beta/openai; CC; gemini-3.1-pro-preview	Gemini OpenAI compatibility	Compatibility and feature support are model-specific; verify tool and usage fields.	Gemini key or gcloud SSO; keep the -preview suffix for this first-party route.
antigravity - Antigravity	GEMINI_API_KEY; https://cloudcode-pa.googleapis.com; GCC; gemini-2.5-flash	Gemini Code Assist	Cloud Code is a native OAuth route. Public Gemini API docs do not establish tool/stream/usage conformance for it.	Use agy/gcloud OAuth and a project ID, not a Gemini key against the OpenAI-compatible host.
xai - xAI Grok	XAI_API_KEY; https://api.x.ai/v1; CC; grok-4.5	xAI API	Verify tool/stream/usage on the chosen Grok model.	API key or supported browser login.
deepseek - DeepSeek	DEEPSEEK_API_KEY; https://api.deepseek.com/v1; CC; deepseek-v4-flash	DeepSeek API	Verify OpenAI-compatible tools and usage for the chosen model. Official models: deepseek-v4-flash, deepseek-v4-pro.	Import $DSH_HOME/.credentials.yaml from DeepSeek Harness, or paste a platform key. No official OAuth.
mistral - Mistral	MISTRAL_API_KEY; https://api.mistral.ai/v1; CC; mistral-large-latest	Mistral API	Tool/stream/usage support depends on model and endpoint.	Create an API key in Mistral's console.
cohere - Cohere	COHERE_API_KEY; https://api.cohere.ai/compatibility/v1; CC; command-a-03-2025	Cohere OpenAI compatibility	Compatibility coverage is not a blanket tools/usage guarantee.	Use a Cohere API key.
ai21 - AI21	AI21_API_KEY; https://api.ai21.com/studio/v1; CC; jamba-large	AI21 Studio	Not established publicly for this exact CC route/model.	Create a Studio key.
reka - Reka	REKA_API_KEY; https://api.reka.ai/v1; CC; reka-core	Reka API	Not established publicly for full tools/stream/usage conformance.	Create a Reka API key.
inception - Inception (Mercury)	INCEPTION_API_KEY; https://api.inceptionlabs.ai/v1; CC; mercury-2	Inception API	Verify CC tools/stream/usage for the account's live catalog.	The catalog's current public default is mercury-2.
writer - Writer (Palmyra)	WRITER_API_KEY; https://api.writer.com/v1; CC-compatible; palmyra-x5	Writer chat endpoint	Writer uses its first-party POST /v1/chat path, not /chat/completions; Nur routes it explicitly.	Create a Writer API key.
upstage - Upstage (Solar)	UPSTAGE_API_KEY; https://api.upstage.ai/v1; CC; solar-pro3	Upstage console	Tool and detailed usage fields remain model-specific.	Create an Upstage key for Solar.
thinkingmachines - Thinking Machines	TINKER_API_KEY; https://tinker.thinkingmachines.dev/services/tinker-prod/oai/api/v1; CC; thinkingmachines/Inkling	Tinker docs	Live model list is key-scoped; public docs do not establish all CC features.	Set the key, then use /model to inspect models available to it.
poolside - Poolside	POOLSIDE_API_KEY; https://inference.poolside.ai/v1; CC; poolside/laguna-m.1	Poolside Platform	Verify tools/stream/usage for the selected Platform deployment.	For self-hosted Platform use https://<domain>/openai/v1.
huggingface - Hugging Face	HF_TOKEN; https://router.huggingface.co/v1; CC; meta-llama/Llama-3.3-70B-Instruct	Inference Providers	Features vary by routed provider/model; no universal tools or usage guarantee.	Create an HF access token.
azure - Azure OpenAI	AZURE_OPENAI_API_KEY; https://YOUR_RESOURCE.openai.azure.com/openai/v1; CC; gpt-5.5	Azure OpenAI REST	Deployment, API version, tools, and usage are resource/model-specific.	Replace YOUR_RESOURCE; API key or supported az login flow.
bedrock - Amazon Bedrock	AWS_BEARER_TOKEN_BEDROCK; https://bedrock-runtime.YOUR_REGION.amazonaws.com/openai/v1; CC; us.amazon.nova-lite-v1:0	Bedrock OpenAI compatibility	Compatibility and tool/usage support are model and region-specific.	Replace YOUR_REGION; use /model for models entitled in that region.
groq - Groq	GROQ_API_KEY; https://api.groq.com/openai/v1; CC; llama-3.3-70b-versatile	Groq API	Verify tools and reported usage for the selected model.	Create a Groq API key.
liquidinference - Liquid Inference	LIQUID_API_KEY; https://router.inference.ai.exchange/v1; CC; liquid.auto:balanced	Liquid Inference	Auction exchange: each request is quoted across providers and capped before the first token (x-liquid-cap-usd); the settled charge is in the exchange block of the response. Tool support is per model; GET /v1/models lists models that accept tools. :smartest / :fastest / :cheapest suffixes re-rank.	Create a spend-scope key (sk-liquid-…).
cerebras - Cerebras	CEREBRAS_API_KEY; https://api.cerebras.ai/v1; CC; gpt-oss-120b	Cerebras Inference	Verify compatibility features by model tier.	Create a Cerebras API key.
together - Together AI	TOGETHER_API_KEY; https://api.together.xyz/v1; CC; meta-llama/Llama-3.3-70B-Instruct-Turbo	Together API	Tool and usage support varies by served model.	Create a Together API key.
fireworks - Fireworks AI	FIREWORKS_API_KEY; https://api.fireworks.ai/inference/v1; CC; accounts/fireworks/models/llama-v3p3-70b-instruct	Fireworks API	Verify tool calling and usage on this model.	Create a Fireworks key.
deepinfra - DeepInfra	DEEPINFRA_API_KEY; https://api.deepinfra.com/v1/openai; CC; meta-llama/Llama-3.3-70B-Instruct-Turbo	DeepInfra API	Compatibility does not establish full feature parity.	Use the -Turbo model spelling used by this host.
novita - NovitaAI	NOVITA_API_KEY; https://api.novita.ai/openai; CC; minimaxai/minimax-m1-80k	Novita API	Not established publicly for full tools/stream/usage conformance.	Create a Novita key.
hyperbolic - Hyperbolic	HYPERBOLIC_API_KEY; https://api.hyperbolic.xyz/v1; CC; meta-llama/Llama-3.3-70B-Instruct	Hyperbolic API	Verify model-level support.	Create a Hyperbolic API key.
nebius - Nebius Token Factory	NEBIUS_API_KEY; https://api.tokenfactory.nebius.com/v1; CC; Qwen/Qwen3.5-397B-A17B-fast	Nebius Token Factory	Verify compatibility features for the selected endpoint/model.	Create a Nebius API key.
sambanova - SambaNova	SAMBANOVA_API_KEY; https://api.sambanova.ai/v1; CC; Meta-Llama-3.3-70B-Instruct	SambaNova Cloud API	Not established publicly for full tools/usage conformance.	Create a SambaNova key.
nvidia - NVIDIA NIM	NVIDIA_API_KEY; https://integrate.api.nvidia.com/v1; CC; meta/llama-3.3-70b-instruct	NVIDIA API catalog	Tool/stream/usage is model-specific.	Create an NVIDIA API key.
perplexity - Perplexity	PERPLEXITY_API_KEY; https://api.perplexity.ai/v1; R (Agent API); perplexity/sonar	Agent API	Sonar's Chat Completions endpoint was retired on 2026-09-27. The Agent API takes provider/model ids (perplexity/sonar, anthropic/..., openai/..., google/..., xai/...) and a narrower Responses schema; nur fits each request to it (no replayed reasoning, text-only assistant turns, image URLs up to 2048 characters). Perplexity's hosted web_search tool is not enabled; nur's own web tools are offered instead.	Create a Perplexity API key.
baseten - Baseten	BASETEN_API_KEY; https://inference.baseten.co/v1; CC; deepseek-ai/DeepSeek-V3-0324	Baseten API	Not established publicly for full OpenAI-tool conformance.	Create a Baseten API key.
friendli - Friendli	FRIENDLI_TOKEN; https://api.friendli.ai/serverless/v1; CC; Qwen/Qwen3-235B-A22B-Instruct-2507	Friendli API	Live serverless catalog and features can change; verify before use.	Use a Friendli token.
chutes - Chutes	CHUTES_API_TOKEN; https://llm.chutes.ai/v1; CC; deepseek-ai/DeepSeek-V3.2-TEE	Chutes API	Verify tools/stream/usage for the chosen chute.	Use a Chutes API token; this model requires its -TEE suffix.
venice - Venice AI	VENICE_API_KEY; https://api.venice.ai/api/v1; CC; zai-org-glm-5-1	Venice Chat Completions	Current docs expose function calls, required usage, and prompt-cache affinity; model capabilities still vary.	Create a Venice key. Nur forwards its stable session cache key on this route.
kimi - Kimi Code (kimi.com)	KIMI_API_KEY; https://api.kimi.com/coding/v1; CC; kimi-for-coding	Kimi Code	Managed coding route; public docs do not establish full OpenAI-tool conformance.	API key or supported browser OAuth.
moonshot - Moonshot AI	MOONSHOT_API_KEY; https://api.moonshot.ai/v1; CC; kimi-k2.7-code	Moonshot API	Verify model-specific tools/stream/usage.	Platform API key, distinct from Kimi Code OAuth.
zhipu - Z.AI	ZAI_API_KEY; https://api.z.ai/api/paas/v4; CC; glm-5.2	Z.AI API · ZCode connect	Coding Plan / ZCode OAuth uses https://api.z.ai/api/coding/paas/v4 and glm-5.3.	zcode login, or paste a key. Do not point a Coding Plan key at /api/paas/v4.
paperinstruments - Paper Instruments	PAPERINSTRUMENTS_API_KEY; https://api.paperinstruments.com/v1; CC; ultramarine-1-flash	Ultramarine API · Chat completions	One model, a GLM-5.3 Flash post-train for finance, law, consulting and operations deliverables (Excel, PowerPoint, Word). Docs cover function tools, streaming with usage and cached_tokens. Reasoning is mandatory: the effort ladder is low/high/max (/effort medium lands on high), and the output ceiling is 32,768 tokens including reasoning. Also speaks Responses and Messages; nur uses Chat Completions.	Create a key at platform.paperinstruments.com. Nur sends the sticky x-session-id header so turns stay on one deployment.
qwen - Alibaba Qwen (DashScope)	DASHSCOPE_API_KEY; https://dashscope-intl.aliyuncs.com/compatible-mode/v1; CC; qwen-max	DashScope OpenAI compatibility	Tool/stream/usage depend on Qwen model and region.	Use an international DashScope key for this endpoint.
minimax - MiniMax (minimaxi.com)	MINIMAX_API_KEY; https://api.minimaxi.com/anthropic/v1; AM; MiniMax-M2.7	MiniMax API overview	Anthropic compatibility does not establish every Anthropic telemetry field; the direct catalog currently documents M2.7/M2.5 families.	Use a MiniMax API key; preserve model casing.
stepfun - StepFun (China)	STEPFUN_API_KEY; https://api.stepfun.ai/v1; CC; step-3.5-flash-2603	StepFun models	Reasoning and tool behavior are model-specific; Nur uses the current documented agent model instead of the nonexistent step-3.7-flash.	Create a StepFun API key.
baichuan - Baichuan	BAICHUAN_API_KEY; https://api.baichuan-ai.com/v1; CC; Baichuan4	Baichuan API	Not established publicly for full tools/usage conformance.	Create a Baichuan API key.
openrouter - OpenRouter	OPENROUTER_API_KEY; https://openrouter.ai/api/v1; CC; openai/gpt-5.5	OpenRouter API	Router features and usage vary by upstream model/provider.	Create one OpenRouter key; use namespaced model ids.
experiential - Experiential Labs	EXPLABS_API_KEY; https://api.experientiallabs.ai/v1; CC; claude-sonnet-5	Experiential Labs docs · Model catalog	1,100+ models behind per-model provider waterfalls; also speaks Responses, /messages and the System One judgment schema on the same key. Model slugs are bare (claude-opus-5, kimi-k3), not namespaced.	Create an xpl_ key at platform.experientiallabs.ai; BYOK or credits, with free tiers. One key also unlocks nur jev use --engine jev-via-experiential.
opendesign - OpenDesign	OPENDESIGN_API_KEY; https://amr-link.open-design.ai/v1; CC; deepseek-v4.1-flash	OpenDesign model credits & API · Pricing	The console's "Connection setup" offers Chat Completions, Responses and Anthropic Messages on the same host; Nur uses Chat Completions. API calls spend the same Design Plan / plan model credits as the desktop app (Go: 5-hour and 7-day windows). Ids use dotted versions (deepseek-v4.1-flash, kimi-k2.7-code, claude-sonnet-5.5); /model lists what your plan can call.	Create a key at open-design.ai/cloud/api-keys (any paid plan, Go upward).
cline - Cline	CLINE_API_KEY; https://api.cline.bot/api/v1; CC; anthropic/claude-sonnet-5	Cline API · Models	Official docs cover streaming, tool calling, usage.cost, and delta.reasoning; capability is per model (supportsImages / supportsReasoning).	Account sign-in (cline auth cline) or an API key from app.cline.bot. Ids are OpenRouter-style vendor/model and must not be rewritten.
commandcode - Command Code	COMMAND_CODE_API_KEY / CMD_API_KEY; https://api.commandcode.ai/provider/v1; CC with per-model AM; deepseek/deepseek-v4-flash	Command Code Provider API	One gateway, 70+ models. Claude ids cannot use /chat/completions on this host; Nur routes those to /messages automatically.	First-party studio sign-in (same flow as cmd login) or a Studio key. See authentication.md for the credential map.
requesty - Requesty	REQUESTY_API_KEY; https://router.requesty.ai/v1; CC; openai/gpt-5.5	Requesty docs	Router compatibility does not establish upstream tools/usage.	Create a Requesty key and select a routed model.
glama - Glama	GLAMA_API_KEY; https://gateway.glama.ai/v1; CC; openai/gpt-5.6-terra	Glama AI Gateway API	Gateway support does not guarantee identical upstream tool/usage fields.	Create a Glama key; catalog default is namespaced.
portkey - Portkey	PORTKEY_API_KEY; https://api.portkey.ai/v1; CC; @openai/gpt-5.5	Portkey API	Virtual-key routing governs upstream features and billing usage.	Replace @openai with your virtual-key slug via /model.
litellm - LiteLLM Proxy	LITELLM_API_KEY optional; http://localhost:4000/v1; CC; gpt-5.5	LiteLLM Proxy	Local proxy configuration determines tools/stream/usage.	Start/configure your proxy; key is optional in this catalog row.
vercel - Vercel AI Gateway	AI_GATEWAY_API_KEY; https://ai-gateway.vercel.sh/v1; CC; openai/gpt-5.5	Vercel AI Gateway	Gateway passes model-specific feature and usage semantics.	Create an AI Gateway key.
cloudflare - Cloudflare AI Gateway	CLOUDFLARE_API_TOKEN; https://api.cloudflare.com/client/v4/accounts/YOUR_ACCOUNT_ID/ai/v1; CC; openai/gpt-5.5	Cloudflare REST API	The old unified gateway.ai.cloudflare.com compatibility path is deprecated; upstream tools/usage remain model-specific.	Replace YOUR_ACCOUNT_ID; authenticate with a Cloudflare API token.
featherless - Featherless	FEATHERLESS_API_KEY; https://api.featherless.ai/v1; CC; Qwen/Qwen3.5-397B-A17B	Featherless tool calling	Native function calling is documented for Qwen 3 and Kimi K2 families, not every hosted model.	Create a Featherless key; Nur defaults to a current Qwen 3.5 model instead of a Llama model that required a JSON workaround.
nano-gpt - NanoGPT	NANOGPT_API_KEY; https://nano-gpt.com/api/v1; CC; openai/gpt-5.6-terra	NanoGPT API	Verify routed model support and usage fields.	Create a NanoGPT key; use namespaced modern model ids.
opencode - OpenCode	OPENCODE_API_KEY; https://opencode.ai/zen/v1; model-routed R/AM/Gemini-native/CC; claude-sonnet-5	OpenCode Zen endpoints	Nur follows the live model table: GPT/Grok use Responses; Claude and Qwen 3.5-3.7 use Messages; Gemini uses native GenerateContent; DeepSeek/MiniMax/GLM/Kimi/free models use CC.	Use opencode auth login; Go-family models route to /zen/go/v1 with the same model-specific protocol selection.
nous - Nous Portal	NOUS_API_KEY; https://inference-api.nousresearch.com/v1; CC; stealth/ox-alpha	Nous Portal	One endpoint reaches 300+ first-party and frontier ids (including a free tier); tools/stream/usage are per model.	Hermes Agent sign-in (hermes auth add nous) is imported automatically, or paste a NOUS_API_KEY.
github-models - GitHub Models	GITHUB_TOKEN; https://models.github.ai/inference; CC; openai/gpt-5	GitHub Models	Model availability and usage limits are account-specific.	Use gh auth or a PAT with GitHub Models access.
github-copilot - GitHub Copilot	COPILOT_GITHUB_TOKEN; https://api.githubcopilot.com; CC; gpt-5.5	Copilot API	Subscription APIs and model/tool availability are entitlement-specific.	Use supported gh OAuth or a fine-grained PAT.
cursor - Cursor	CURSOR_API_KEY; https://api2.cursor.sh; CC catalog placeholder; auto	Cursor CLI	Not CC: Nur uses cursor-agent -p; public docs do not establish API tools/stream/usage receipts.	Run cursor-agent login; no pasted key is required for CLI chat.
factory - Factory Droid	FACTORY_API_KEY (optional); https://app.factory.ai CC catalog placeholder; claude-sonnet-5	Droid Exec	Not HTTP: every round is droid exec --output-format json in read-only spec mode with nur's tools offered through the nur-tools fence; --session-id is reused so later rounds ship only the delta. Usage is reported only when the CLI returns it. NUR_DROID_NATIVE=1 delegates the task to Droid itself with --auto (NUR_DROID_AUTO, default medium). Model ids: /model asks the installed CLI (droid exec --help, cached per binary under ~/.nur/cache/droid-models.json, warmed at startup); the Factory-managed catalog is only the fallback.	Install the CLI (curl -fsSL https://app.factory.ai/cli \| sh or npm i -g @factory/cli), then sign in once with droid - /login → Factory Droid → browser imports that OAuth session (nothing secret is copied; the sealed keyring stays with the CLI). A pasted fk-… key still works. Nur looks for droid in ~/.factory/bin, ~/.local/bin, ~/bin, the local programs directory, the Factory desktop app's bundle (%LOCALAPPDATA%\Factory\app-*\resources\bin, /Applications/Factory.app/Contents/Resources/bin) and PATH, and uses the most recently built one - an installer launcher that is never run does not self-update, and an old build hides the models its catalog predates. Set NUR_DROID_BIN to pin a specific binary.
helicone - Helicone AI Gateway	HELICONE_API_KEY; https://ai-gateway.helicone.ai/v1; CC; gpt-5.4	Helicone Gateway	Gateway observability is not authoritative upstream usage unless its docs/provider response expose it.	Create a Helicone key; this gateway currently uses bare model ids.
aimlapi - AI/ML API	AIMLAPI_KEY; https://api.aimlapi.com/v1; CC; openai/gpt-5.6-terra	AI/ML API docs	Router conformance is model-specific.	Create an AI/ML API key.
ollama - Ollama (local)	OLLAMA_API_KEY optional; http://localhost:11434/v1; CC; local-model (resolved from /models)	Ollama OpenAI compatibility	Local server/model determines tools, streaming, and usage; no remote billing receipt.	Run Ollama and pull a model; Nur selects an installed model instead of assuming llama3.3.
lmstudio - LM Studio (local)	LMSTUDIO_API_KEY optional; http://localhost:1234/v1; CC; local-model	LM Studio API	Installed server/model determines feature and usage fields.	Start the local server; Nur resolves the first live model from /models, and /model remains available for an explicit choice.
llamacpp - llama.cpp (local)	LLAMACPP_API_KEY optional; http://localhost:8080/v1; CC; local-model	llama-server	Server build/model determines tools and usage; no remote billing receipt.	Start llama-server; Nur resolves its loaded model from /models.
vllm - vLLM (local)	VLLM_API_KEY optional; http://localhost:8000/v1; CC; local-model	vLLM tool calling	Installed version, model, chat template, and tool parser determine agent behavior.	Start the OpenAI-compatible server; Nur resolves the served model from /models. Configure the model's documented tool parser for agent use.
jan - Jan (local)	JAN_API_KEY optional; http://localhost:1337/v1; CC; local-model	Jan API server	Local runtime/model determines tool, stream, and usage behavior.	Start Jan's API server; Nur resolves the active model from /models.

Prompt caching

Every provider that caches prompts does so by prefix: a request is served from cache up to the first byte that differs from an earlier one. Nur shapes every request so that prefix is as long and as stable as it can be:

• The system prompt is static for the session. Permission mode, the todo list and the session inventories (goal, harness, context variables, routed memory) are not in it. They ride as the last input item of each request, headed [nur turn state], which is rebuilt per request and never persisted into the history. The prompt the model reads is byte-identical from the first request of a turn to the last, whatever mode you switch to in between.
• The full tool roster is sent on every request, in a stable order, so the schemas are cached once instead of re-read every time the offered set changes.
• Skill activation is a conversation item, not a prompt edit: when a skill fires, its body is appended after the user message and cached with the history.
• Anthropic cache breakpoints sit on the system prompt and the tool roster, then roll along the conversation; the turn-state block never carries one.
• Session affinity. Several backends shard their caches per session and only serve a prefix from the replica that already holds it. Nur sends one stable id, derived from the nur session id (so it survives /model, /resume and failover), as whichever header the route understands: x-opencode-session on the OpenCode Go route (gated on the URL, so a custom provider pointed at Go gets it too), session_id on the Codex OAuth path, x-session-affinity on Fireworks, and x-session-id on Paper Instruments. OpenAI-style routes also set prompt_cache_key to the session id.
• Compaction is cache-aware: pruning happens in the oldest part of the transcript only when the window is actually full, which is the one moment the prefix has to change anyway.

What you see: when a provider reports cached tokens, each request posts a transient status line, prompt cache · 84% of input read from cache (31k of 37k tok), in the TUI footer (printed in nur run --verbose). Cached tokens are priced at the model's cache-read rate in the cost estimate. E2E scenario 31 asserts the stable prefix across a whole turn.

Catalog conformance check

The matrix must have exactly 70 data rows, matching PROVIDERS.len(). A row is a catalog contract, not a promise that a remote model is available to every account. The primary expected catalog mismatches are intentional placeholders:

• azure requires YOUR_RESOURCE.
• bedrock requires YOUR_REGION and region/model entitlement.
• cloudflare requires YOUR_ACCOUNT_ID.
• cursor is a cursor-agent CLI integration despite its catalog CC marker.
• Local rows and LiteLLM intentionally use optional keys and depend on a user-operated server.