@pentoshi/clai 4.11.7 → 4.11.8

package/dist/llm/provider-info-text.js

@@ -0,0 +1,587 @@
+export const providerInfo = {
+    free: `Free (opencode zen + kilo gateway) — keyless OpenAI-compatible models
+
+WHAT IT IS
+  The default provider for a fresh clai install. It bundles two keyless
+  gateways behind one provider id, namespaced by source:
+    free-1/<model>   opencode zen   https://opencode.ai/zen/v1
+    free-2/<model>   kilo gateway   https://api.kilo.ai/api/gateway
+  Both serve free models with NO API key — requests are forwarded without an
+  Authorization header. Chat Completions with SSE streaming, native tool
+  calling and reasoning_content thinking all work. A bare model id with no
+  free-N/ prefix routes to free-1.
+
+  Auth       none for free models (Bearer only if you add a key)
+  Endpoints  /models · /chat/completions on both gateways
+
+MODELS
+  /model lists the live catalogs from both gateways (each cached for an
+  hour), namespaced by source:
+    free-2/kilo-auto/free                      (clai default)
+    free-1/mimo-v2.5-free
+    free-1/hy3-free
+    free-1/x-preview-f-free
+    free-2/stepfun/step-3.7-flash:free
+    free-2/nvidia/nemotron-3-ultra-550b-a55b:free
+  zen free ids end in -free; kilo free ids end in :free or /free (the kilo
+  catalog also flags them with isFree). Premium models stay hidden unless
+  you add a key. The free sets rotate upstream and models can be delisted
+  without notice — treat availability as transient and just pick another id
+  from /model.
+
+COST
+  Free. No signup, no key, no card. The trade-off is reliability: free tiers
+  are capacity-constrained, rate limited and occasionally down. If a request
+  fails, retry once — and for dependable daily use set a key for any other
+  provider (clai set <provider> <key>, then clai use <provider>).
+
+SETUP
+  None. A fresh install already uses this provider.
+  Optional: clai set free <key>   unlock premium zen models on your account
+  Optional env var: FREE_API_KEY  (used when nothing is stored)
+
+GOOD TO KNOW
+  - Premium models without a key fail fast with a 402-style message instead
+    of proxying an upstream 401.
+  - Reasoning models stream thinking as reasoning_content; clai folds it
+    into the usual thinking block, so /think and /effort behave normally.
+  - Classed as free-cloud, so /freeonly on keeps it in the fallback chain.`,
+    tokenrouter: `TokenRouter — one key for frontier open models
+
+WHAT IT IS
+  An OpenAI-compatible gateway that fronts Kimi, DeepSeek, Qwen, GLM, GPT-OSS
+  and MiniMax behind a single bearer key and base URL. Chat Completions plus
+  the Responses API upstream; clai uses Chat Completions with SSE streaming,
+  native tool calling and JSON mode.
+
+  Base URL   https://api.tokenrouter.com/v1   (override if your account uses
+             a different host — see ENDPOINTS below)
+  Auth       Authorization: Bearer <key>
+  Endpoints  /models · /chat/completions
+
+MODELS
+  Ids are namespaced by vendor and case-sensitive (moonshotai/kimi-k3, not
+  "Kimi K3"). /model reads the live list from /models, filtered to the
+  channels your key can actually reach.
+
+COST
+  Prepaid balance billed per token; the dashboard header shows what is left.
+  There is no published free tier, so treat it as paid: clai classes it
+  paid-cloud and /freeonly on keeps it out of the fallback chain. Prices vary
+  a lot between the flash/fast variants and the pro ones.
+
+SETUP
+  1. Create an API key in your TokenRouter account under API Keys.
+  2. clai set tokenrouter sk-yourKey
+  3. clai use tokenrouter
+  4. /model moonshotai/kimi-k3      (or any id from /model)
+
+MANAGING KEYS AND ENDPOINTS IN clai
+  clai set tokenrouter <key>            add a key (up to 10, rotated on failure)
+  clai keys                             masked keys + the active endpoint
+  clai unset tokenrouter                remove every stored key
+  /set tokenrouter                      TUI: endpoint editor, then key editor
+  /info tokenrouter                     this page
+
+  Base URL (optional — defaults to api.tokenrouter.com):
+  clai set tokenrouter --url https://tokenrouter.me/v1
+  clai set tokenrouter --url <a> --url <b>   store several, sticky active one
+  clai unset tokenrouter --url               back to the default
+  TOKENROUTER_BASE_URL overrides the whole list.
+
+GOOD TO KNOW
+  - Reasoning models return their thinking in reasoning_content; clai folds it
+    into the usual thinking block, so /think and /effort behave normally.
+  - max_tokens above a model's ceiling is clamped by the gateway rather than
+    rejected, but prompt + max_tokens must still fit the context window.
+  - The model field in responses may echo a fully-qualified upstream path
+    instead of the id you sent. That is expected.
+  - Env var: TOKENROUTER_API_KEY (used when nothing is stored).
+
+Docs: https://docs.tokenrouter.me`,
+    lightning: `Lightning AI Model APIs — one key for OpenAI, Anthropic, Google
+and Lightning-hosted open models
+
+WHAT IT IS
+  An OpenAI-compatible gateway at https://lightning.ai/api/v1 that fronts
+  frontier models from several vendors plus open models Lightning serves
+  itself. One account, one key, no per-vendor subscriptions, billed by the
+  token. Model ids are vendor-namespaced:
+    openai/gpt-5, openai/o3, openai/gpt-5.6-sol
+    anthropic/claude-opus-4-8, anthropic/claude-sonnet-4-6
+    google/gemini-3.5-flash, google/gemini-2.5-pro
+    lightning-ai/gpt-oss-120b, lightning-ai/deepseek-v4-pro,
+    lightning-ai/nemotron-3-ultra-550b-a55b, lightning-ai/gemma-4-31B-it
+
+FREE TIER — WHAT YOU GET
+  New accounts get up to 40 million free tokens to start, and the docs
+  advertise no subscription or credit card to begin. After that it is
+  pay-per-token at each model's rate — the /models endpoint returns the
+  per-token input and output price for every id, so nothing is hidden.
+
+  Rate limits by plan (requests/min · tokens/min):
+    Free         15 ·  120,000
+    Pro          20 ·  120,000
+    Teams        30 ·  150,000
+    Enterprise  300 ·  unlimited
+
+  Signing up needs a non-virtual phone number, and the free grant is once
+  per person — a second account does not get a second grant.
+
+SETUP — STEP BY STEP
+  1. Create an account at https://lightning.ai
+  2. Open the Model APIs page and reveal your key:
+       https://lightning.ai/lightning-ai/model-apis/models?showApiKey=true
+  3. Give it to clai:
+       clai set lightning <your-api-key>
+       clai use lightning
+  4. Pick a model — /model lists the live catalog from the gateway:
+       /model openai/gpt-5
+       /model lightning-ai/gpt-oss-120b     (cheapest open-weight option)
+
+  Or export it instead of storing it:
+       LIGHTNING_API_KEY=<your-api-key>
+
+MANAGING KEYS AND ENDPOINTS IN clai
+  clai set lightning <key>            add a key (up to 10, rotated on failure)
+  clai set lightning <key2>           add another; the last that worked is sticky
+  clai keys                           masked keys + the active endpoint
+  clai unset lightning                remove every stored key
+  /set lightning                      TUI: endpoint editor, then multi-key editor
+  /info lightning                     this page
+
+  Base URL (optional — defaults to the shared gateway above):
+  clai set lightning --url <url>       add an endpoint, make it active
+  clai set lightning --url <a> --url <b>   add several
+  clai unset lightning --url           back to the default gateway
+  /set lightning https://...           add + activate one endpoint
+
+  Point it at a private Lightning Inference deployment or a proxy that speaks
+  the same OpenAI routes; up to 10 URLs are stored with a sticky active one.
+  LIGHTNING_BASE_URL overrides the whole list.
+
+GOOD TO KNOW
+  - The catalog lists one entry per published agent/preset, so the same model
+    id appears more than once upstream; clai dedupes it for /model.
+  - Cost varies enormously across the list — claude-opus and gpt-5.x cost
+    dollars per million tokens while lightning-ai/gpt-oss-* are cents. Check
+    the price on https://lightning.ai/models before long agent runs.
+  - Streaming, native tool calling and reasoning_effort all work; clai sends
+    the standard OpenAI knobs and retries without them if a model objects.
+  - Classed as paid-cloud, so /freeonly on keeps it out of the fallback
+    chain even while the free tokens last.
+  - Model APIs are separate from Lightning Studios/GPU credits; the token
+    grant is not the same balance as Studio compute credits.
+
+Docs:   https://lightning.ai/docs/platform/inference/model-apis
+Models: https://lightning.ai/models
+API:    https://lightning.ai/api/v1 (OpenAI-compatible; /models, /chat/completions)`,
+    modal: `Modal Endpoints — your own serverless OpenAI-compatible endpoint
+
+WHAT IT IS
+  Modal Endpoints deploy an open-weight model (Kimi, Qwen, DeepSeek, GLM,
+  Gemma, GPT-OSS, Nemotron, or your own fine-tune) behind a low-latency
+  request proxy. The endpoint serves the standard Chat Completions API under
+  /v1, autoscales under load, and scales to zero when idle. The endpoint
+  belongs to your workspace, so the base URL is unique to you:
+    https://<workspace>--ep-<endpoint>.<region>.modal.direct/v1
+
+FREE TIER — WHAT YOU GET AND WHAT IT COSTS
+  Signing up is free. The Starter plan is $0/month + compute and includes
+  $30 of free compute credit every month; you unlock it by adding a payment
+  method, and nothing is charged until a month's usage passes the credit.
+  Starter also includes 3 workspace seats, 100 containers + 10 GPU
+  concurrency, region selection, and real-time metrics and logs.
+  Team is $250/month + compute with $100/month of credit, 1000 containers
+  and 50 GPU concurrency. Credit grants exist for early-stage startups and
+  for academics (up to $10k).
+
+  Billing is per-second compute (GPU + CPU) at standard Modal rates — not
+  per token. Because endpoints scale to zero you pay only while a container
+  is starting or serving; an idle endpoint costs nothing. Pinning compute to
+  a region applies a price multiplier. The credit is granted per month, so
+  keep an eye on the Credits figure in the dashboard header.
+
+SETUP — STEP BY STEP
+  1. Create a free account at https://modal.com
+  2. Add a payment method (dashboard → settings → billing) to activate the
+     $30/month credit.
+  3. Install the CLI and log in:
+       pip install modal
+       modal setup      (browser login; writes tokens to ~/.modal.toml)
+  4. Deploy an endpoint from the Endpoints tab in the dashboard, or:
+       modal endpoint create --model moonshotai/Kimi-K3 --name kimi-k3
+       modal endpoint create --model Qwen/Qwen3.5-4B --routing-region us-east
+     Provisioning takes a few minutes. Region default is us-west.
+  5. Copy the endpoint URL from the endpoint Overview page, or run
+       modal endpoint list
+  6. Create a proxy token pair (endpoints are authenticated by default):
+       modal workspace proxy-tokens create
+     This prints a token ID (wk-...) and a token secret (ws-...). THE SECRET
+     IS SHOWN ONLY ONCE — copy it now; it cannot be retrieved later. You can
+     also create one in the dashboard under workspace settings, and the
+     endpoint Quickstart panel lists the token IDs you already have.
+     On RBAC workspaces, also scope it to the endpoint's environment:
+       modal workspace proxy-tokens allow <token-id> <environment>
+  7. Point clai at it (URL from step 5, pair from step 6):
+       clai set modal --url <endpoint-url>
+       clai set modal wk-yourTokenId:ws-yourTokenSecret
+       clai use modal
+  8. Verify: "clai keys" shows the masked pair and the endpoint it points at,
+     and /model lists the models the endpoint actually serves.
+
+KEYS — WHAT IS NEEDED AND WHERE IT GOES
+  Endpoint URL        stored in config as modalBaseUrl (not a secret)
+  Proxy token ID      wk-...  sent as the Modal-Key header
+  Proxy token secret  ws-...  sent as the Modal-Secret header
+
+  There is no bearer API key. clai stores the pair as ONE secret shaped
+  "<token-id>:<token-secret>" in the OS keychain (or ~/.clai/keys.json with
+  restricted permissions when no keychain is available), so masking,
+  rotation and /unset behave exactly like every other provider.
+
+  Proxy tokens (wk-/ws-) are NOT Modal API tokens (ak-/as-). API tokens
+  authenticate the CLI and SDK; they will not work as endpoint headers.
+
+MANAGING ENDPOINTS AND KEYS IN clai
+  Both are multi-entry lists with a sticky active choice, up to 10 each.
+
+  clai set modal --url <endpoint>          add an endpoint, make it active
+  clai set modal --url <a> --url <b>       add several in one call
+  clai set modal --url <known endpoint>    re-activate one already stored
+  clai set modal wk-id:ws-secret           add a token pair
+  clai set modal                           prompt for the pair (input hidden)
+  clai keys                                every endpoint + masked pairs, ★ active
+  clai unset modal --url                   clear the endpoint list only
+  clai unset modal                         clear the token pairs only
+  /set modal                               TUI: endpoint editor, then keys
+  /set modal https://...                   add + activate one endpoint
+  /info modal                              this page
+
+  One endpoint serves one model, so keep an endpoint per model you deploy and
+  switch with clai set modal --url <that endpoint> — or star a row in the
+  /set endpoint editor. Endpoints are not auto-rotated on failure, because a
+  different endpoint serves a different model; token pairs DO rotate like any
+  other provider (401 / 403 / 429 / 5xx / empty response moves to the next).
+
+ENVIRONMENT VARIABLES (used only when nothing is stored)
+  MODAL_BASE_URL              endpoint URL; overrides the stored list entirely
+                              ("/v1" is appended if missing)
+  MODAL_PROXY_TOKEN_ID        wk-...  both halves needed, or it is ignored
+  MODAL_PROXY_TOKEN_SECRET    ws-...
+  MODAL_SESSION_ID            optional sticky-session id. When unset, clai
+                              generates one per run so a whole conversation
+                              lands on the same warm container.
+
+GOOD TO KNOW
+  - Cold start: the first request after idle pays container start-up, which
+    can take tens of seconds on a large model. clai allows up to 3 minutes
+    for the first token before treating a stream as stalled.
+  - The model name on the wire is the source repo id (e.g.
+    moonshotai/Kimi-K3), not a Modal alias. /model reads the live list.
+  - Endpoint URLs are stored as a list, so several deployments can live side
+    by side; only the active one is used. "clai keys" shows them all.
+  - Streaming, native tool calling, structured outputs and thinking all work.
+    /effort on|off maps to Modal's reasoning toggle.
+  - Finished experimenting? "modal endpoint stop <name>" tears the endpoint
+    down so no stray request can wake it and spend credit.
+  - Modal counts as paid-cloud, so /freeonly on keeps it out of the
+    cross-provider fallback chain.
+
+TROUBLESHOOTING
+  401 / 403         wrong, revoked or unscoped pair — or an API token (ak-)
+                    was pasted instead of a proxy token (wk-)
+  404               endpoint URL is wrong, or the endpoint was stopped
+  "endpoint URL is not configured"
+                    run clai set modal --url <endpoint>
+  model not found   send the repo id reported by /model or modal endpoint list
+  slow first token  cold start; the endpoint had scaled to zero
+
+Docs:    https://modal.com/docs/guide/endpoints
+Auth:    https://modal.com/docs/guide/webhook-proxy-auth
+Pricing: https://modal.com/pricing`,
+    bynara: `Current Plan
+
+Free
+Daily token cap
+0 / 7,000,000 used
+7,000,000 remaining
+
+Rate limit
+
+10 req/min
+Reset time
+
+07.00 WIB
+Plan expires
+
+No expiry`,
+    meta: `Meta Model API — Muse Spark (Meta Superintelligence Labs)
+
+WHAT IT IS
+  Meta's OpenAI-compatible API for agentic and coding workflows. It serves the
+  Muse Spark lineup behind a single bearer key at https://api.meta.ai/v1.
+
+  Base URL   https://api.meta.ai/v1
+  Auth       Authorization: Bearer <key>
+  Endpoints  /models · /chat/completions
+  Context    1,048,576 tokens (1M)
+
+MODELS
+  muse-spark-1.2             current general model (clai default)
+  muse-spark-1.1             previous generation
+  muse-spark-1.2-contributor contributor tier (see /model for the live list)
+
+  Muse Spark is a reasoning-first model: it always thinks internally before
+  answering. /effort maps clai's effort onto the API's reasoning_effort
+  (minimal/low/medium/high/xhigh; "off" degrades to minimal because Muse does
+  not support disabling reasoning — "none" returns HTTP 400).
+
+COST
+  Pay-as-you-go per token. Cached input tokens bill at a lower rate than
+  uncached input. clai classes it paid-cloud, so /freeonly on keeps it out of
+  the fallback chain.
+
+SETUP
+  1. Create an API key in your Meta Model API dashboard (MODEL_API_KEY).
+  2. clai set meta <your-key>
+  3. clai use meta
+  4. /model muse-spark-1.2      (or any id from /model)
+
+MANAGING KEYS IN clai
+  clai set meta <key>            add a key (up to 10, rotated on failure)
+  clai set meta <key2>           add another; the last that worked is sticky
+  clai keys                      masked keys + the active endpoint
+  clai unset meta                remove every stored key
+  /set meta                      TUI: multi-key editor
+  /info meta                     this page
+
+GOOD TO KNOW
+  - Reasoning, native tool calling, image understanding and prompt caching all
+    work. Cached tokens arrive as usage.prompt_tokens_details.cached_tokens and
+    show up in the usual usage footer.
+  - Env var: MODEL_API_KEY (used when nothing is stored).
+
+Docs: https://dev.meta.ai/docs`,
+    fireworks: `Fireworks AI — OpenAI-compatible inference for open models
+
+WHAT IT IS
+  Fireworks serves open models behind an OpenAI-compatible API at
+  https://api.fireworks.ai/inference/v1. One key, many models, billed per
+  token. Chat Completions with SSE streaming, native tool calling, vision
+  via image_url, structured outputs and reasoning all work.
+
+  Base URL   https://api.fireworks.ai/inference/v1
+  Auth       Authorization: Bearer <key>
+  Endpoints  /models · /chat/completions
+
+MODELS (ids are fully-qualified — accounts/fireworks/models/<name>)
+  accounts/fireworks/models/kimi-k2p6              (clai default) 256K ctx
+  accounts/fireworks/models/kimi-k2-instruct-0905  256K ctx
+  accounts/fireworks/models/deepseek-v3p1          128K ctx
+  accounts/fireworks/models/glm-5p2                200K ctx
+  accounts/fireworks/models/qwen3-235b-a22b        128K ctx
+  accounts/fireworks/models/gpt-oss-120b           128K ctx
+  /model reads the live list from /models (cached 1h).
+
+COST
+  Pay-as-you-go per token. No free tier — clai classes it paid-cloud so
+  /freeonly on keeps it out of the fallback chain.
+
+SETUP
+  1. Create an API key at https://app.fireworks.ai/settings/users/api-keys
+  2. clai set fireworks <your-key>
+  3. clai use fireworks
+  4. /model accounts/fireworks/models/kimi-k2p6
+
+MANAGING KEYS IN clai
+  clai set fireworks <key>           add a key (up to 10, rotated on failure)
+  clai set fireworks <key2>          add another; last success is sticky
+  clai keys                          masked keys
+  clai unset fireworks               remove every stored key
+  /set fireworks                     TUI: multi-key editor
+  /info fireworks                    this page
+
+GOOD TO KNOW
+  - Reasoning models return thinking in reasoning_content; clai folds it
+    into the usual thinking block so /think and /effort work.
+  - Vision models accept image_url; clai sends images as OpenAI image_url
+    blocks when the model supports vision.
+  - Env var: FIREWORKS_API_KEY (used when nothing is stored).
+
+Docs: https://docs.fireworks.ai
+API:  https://docs.fireworks.ai/api-reference/introduction`,
+    hetzner: `Hetzner Inference — OpenAI-compatible inference on EU infrastructure
+
+WHAT IT IS
+  Hetzner's experimental inference API at https://inference.hetzner.com/api/v1.
+  OpenAI-compatible Chat Completions with SSE streaming, native tool calling
+  and vision via image_url. Servers in Germany and Finland, outside the US
+  CLOUD Act. Currently experimental — free during test phase, no SLA/DPA yet.
+
+  Base URL   https://inference.hetzner.com/api/v1
+  Auth       Authorization: Bearer <key>
+  Endpoints  /models · /chat/completions
+
+MODELS
+  Qwen/Qwen3.6-35B-A3B-FP8             current model (clai default) 262K ctx
+  Qwen/Qwen3.6-35B-A3B                 alias without FP8 suffix
+  /model reads the live list from /models (cached 1h). Additional models
+  will be added based on demand.
+
+COST
+  Free during experimental phase — no billing yet. clai classes it free-cloud
+  so /freeonly on keeps it in the fallback chain. Expect Hetzner pricing
+  well below US hyperscalers once billing arrives.
+
+SETUP
+  1. Create an API token at https://experiments.hetzner.com/
+  2. clai set hetzner <your-token>
+  3. clai use hetzner
+  4. /model Qwen/Qwen3.6-35B-A3B-FP8
+
+MANAGING KEYS IN clai
+  clai set hetzner <key>               add a key (up to 10, rotated on failure)
+  clai set hetzner <key2>              add another; last success is sticky
+  clai keys                            masked keys
+  clai unset hetzner                   remove every stored key
+  /set hetzner                         TUI: multi-key editor
+  /info hetzner                        this page
+
+GOOD TO KNOW
+  - Qwen3 is a reasoning model: it thinks before answering. /effort on|off
+    maps to chat_template_kwargs.enable_thinking so thinking can be disabled
+    and the completion budget is not spent on hidden reasoning.
+  - Vision: Qwen3.6 accepts images; clai sends them as OpenAI image_url blocks
+    when the model supports vision.
+  - Tool calling works via OpenAI tools; streaming and prompt caching work.
+  - Env var: HETZNER_API_KEY (also HETZNER_INFERENCE_API_KEY, used when
+    nothing is stored).
+
+Docs: https://experiments.hetzner.com/docs/inference
+API:  https://inference.hetzner.com/api/v1`,
+    orcarouter: `OrcaRouter — one key for OpenAI, Anthropic, Google, DeepSeek,
+Grok, Qwen, Kimi, MiniMax and GLM at provider cost price
+
+WHAT IT IS
+  An OpenAI-compatible gateway at https://api.orcarouter.ai/v1 that routes
+  eleven upstream providers behind one bearer key, with zero token markup
+  (you pay each provider's published price). Model ids are vendor-prefixed:
+    openai/gpt-4o-mini, openai/gpt-5, openai/o3-mini
+    anthropic/claude-sonnet-4.6, anthropic/claude-opus-4.7
+    google/gemini-2.5-flash, google/gemini-3-pro-preview
+    deepseek/deepseek-reasoner, grok/grok-4-fast-reasoning
+    qwen/qwen3-max, kimi/kimi-k2.6, minimax/minimax-m2.7, z-ai/glm-5.1
+  Chat Completions with SSE streaming, native tool calling, structured
+  outputs (response_format json_schema), vision via image_url and a unified
+  reasoning_effort knob all work.
+
+  Base URL   https://api.orcarouter.ai/v1
+  Auth       Authorization: Bearer <key>
+  Endpoints  /models · /chat/completions
+
+MODELS
+  Ids are namespaced by upstream vendor and case-sensitive. /model reads the
+  live list from /models (cached 1h), filtered to models reachable over Chat
+  Completions — image/video/tts/embedding ids are hidden. orcarouter/auto
+  picks the cheapest model that fits the request.
+
+REASONING
+  One unified syntax: top-level reasoning_effort (low/medium/high, plus
+  minimal/max on some models). The gateway translates it per upstream —
+  OpenAI native, Claude thinking budgets, Gemini thinkingConfig, Grok
+  native. /think and /effort map onto it; thinking arrives as
+  reasoning_content and folds into the usual thinking block.
+
+COST
+  Zero markup: provider list price per token. Revenue comes from optional
+  subscription plans, not inflated token cost. clai classes it paid-cloud,
+  so /freeonly on keeps it out of the fallback chain.
+
+SETUP
+  1. Create an API key at https://www.orcarouter.ai/console (starts sk-…).
+  2. clai set orcarouter sk-yourKey
+  3. clai use orcarouter
+  4. /model openai/gpt-4o-mini      (or any id from /model)
+
+MANAGING KEYS IN clai
+  clai set orcarouter <key>          add a key (up to 10, rotated on failure)
+  clai set orcarouter <key2>         add another; the last that worked is sticky
+  clai keys                          masked keys + the active endpoint
+  clai unset orcarouter              remove every stored key
+  /set orcarouter                    TUI: multi-key editor
+  /info orcarouter                   this page
+
+GOOD TO KNOW
+  - Multi-key rotation, prompt caching, tool calling and compaction behave
+    exactly like every other OpenAI-compatible provider.
+  - Per-key options at creation: name, credit limit, expiration. Rate limits
+    are workspace-level, not per-key.
+  - Env var: ORCAROUTER_API_KEY (used when nothing is stored).
+
+Docs: https://docs.orcarouter.ai
+API:  https://api.orcarouter.ai/v1 (OpenAI-compatible; /models, /chat/completions)`,
+    "merge-gateway": `Merge Gateway — one key for OpenAI, Anthropic, Google and
+more through a unified gateway
+
+WHAT IT IS
+  A gateway at https://api-gateway.merge.dev/v1 that fronts several upstream
+  vendors behind one key. It exposes two surfaces: its own Responses-style
+  API at /v1 and a drop-in OpenAI-compatible surface at /v1/openai. clai
+  uses the OpenAI-compatible surface, so streaming, native tool calling,
+  prompt caching, reasoning effort and key rotation all behave exactly like
+  every other OpenAI-compatible provider. Model ids are vendor-prefixed:
+    openai/gpt-5.2, openai/gpt-4o-mini, openai/o4-mini
+    anthropic/claude-sonnet-4-6, anthropic/claude-opus-4-6
+    google/gemini-3.5-flash, google/gemini-2.5-pro
+    deepseek/deepseek-reasoner, meta/llama-3.3-70b-instruct
+
+  Base URL   https://api-gateway.merge.dev/v1/openai
+  Auth       Authorization: Bearer <key>   (X-API-Key also sent)
+  Endpoints  /models · /chat/completions
+
+MODELS
+  Ids are namespaced by upstream vendor and case-sensitive. /model reads the
+  live list from /models (cached 1h) and hides ids that are not reachable
+  over Chat Completions, so embedding/image/tts entries stay out of the
+  picker. If the catalog cannot be fetched, a documented offline subset is
+  shown instead.
+
+REASONING
+  Top-level reasoning_effort (low/medium/high) is translated by the gateway
+  to each upstream's native reasoning shape. /think and /effort map onto it,
+  and thinking arrives as reasoning_content, folded into the usual thinking
+  block. clai only sends reasoning options to models that accept them, and
+  retries once without them if a model rejects them.
+
+COST
+  Billed per token by plan. The free tier has a budget: once exhausted the
+  gateway answers 402, which clai treats as a quota error and rotates to the
+  next key or provider. clai classes it paid-cloud, so /freeonly on keeps it
+  out of the fallback chain.
+
+SETUP
+  1. Create an API key at https://gateway.merge.dev (starts mg_…).
+  2. clai set merge-gateway mg_yourKey
+  3. clai use merge-gateway
+  4. /model openai/gpt-5.2           (or any id from /model)
+
+MANAGING KEYS IN clai
+  clai set merge-gateway <key>       add a key (up to 10, rotated on failure)
+  clai set merge-gateway <key2>      add another; the last that worked is sticky
+  clai keys                          masked keys + the active endpoint
+  clai unset merge-gateway           remove every stored key
+  /set merge-gateway                 TUI: multi-key editor
+  /info merge-gateway                this page
+
+GOOD TO KNOW
+  - Aliases: merge-gateway, mergegateway, merge, mg.
+  - Errors follow the documented gateway table — 400 bad request, 401 auth,
+    402 budget exhausted, 404 unknown model, 429 rate limit. Auth and quota
+    errors switch keys immediately; 429 backs off first.
+  - Env var: MERGE_GATEWAY_API_KEY (used when nothing is stored).
+
+Docs: https://gateway.merge.dev
+API:  https://api-gateway.merge.dev/v1/openai (OpenAI-compatible)`,
+};
+//# sourceMappingURL=provider-info-text.js.map

package/dist/llm/provider-info-text.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"provider-info-text.js","sourceRoot":"","sources":["../../src/llm/provider-info-text.ts"],"names":[],"mappings":"AAAA,MAAM,CAAC,MAAM,YAAY,GAA2B;IAClD,IAAI,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;2EA8CmE;IACzE,WAAW,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kCAoDmB;IAChC,SAAS,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;oFA4EuE;IAClF,KAAK,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;mCA2H0B;IACjC,MAAM,EAAE;;;;;;;;;;;;;;;UAeA;IACR,IAAI,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;+BA8CuB;IAC7B,SAAS,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;2DA+C8C;IACzD,OAAO,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;2CAgDgC;IACzC,UAAU,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;mFA4DqE;IACjF,eAAe,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;kEA6D+C;CACjE,CAAC"}

package/dist/llm/provider-model-layers.d.ts

@@ -1,3 +1,5 @@
 import type { ProviderId } from "../types.js";
 import type { ProviderProfileLayer } from "./provider-profile.js";
+import { QWEN3_MAX_GATEWAY_EFFORTS, QWEN3_MAX_PATTERN } from "./model-layers/model-patterns.js";
+export { QWEN3_MAX_GATEWAY_EFFORTS, QWEN3_MAX_PATTERN };
 export declare function modelLayerFor(provider: ProviderId, model: string): ProviderProfileLayer | undefined;