@bitkyc08/opencodex
Version:
Universal provider proxy for OpenAI Codex & Claude Code — use any LLM with Codex CLI/App/SDK and Claude Code
1,165 lines • 61.6 kB
text/typescript
import type { ProviderModelDiscoverySpec } from "./types";
// Shared between the OAuth (Claude account) and API-key Anthropic entries so both expose the
// same static model seed.
// 260710 context refresh: Tier-2 evidence in
// devlog/_plan/260710_provider_hardening/001_research_frontier.md.
// 260902 Claude Fable 5.1 (`claude-fable-5-1`): 1M context / 128K output / adaptive thinking
// always on, per the official models overview and pricing page (platform.claude.com).
// 260923 Claude Opus 5.5 (`claude-opus-5-5`, released 2026-09-22): 1M context / 128K output /
// adaptive thinking always on / effort low..max with a medium default, per the Opus 5.5
// overview, effort and pricing pages (platform.claude.com).
// 260929 Claude Sonnet 5.5 (`claude-sonnet-5-5`, released 2026-09-28): 1M context / 128K output /
// adaptive thinking / effort low..max with a high default, same price as Sonnet 5, per the Sonnet 5.5
// overview, migration guide, effort and pricing pages (platform.claude.com).
export const ANTHROPIC_MODELS = ["claude-fable-5-1", "claude-fable-5", "claude-sonnet-5-5", "claude-sonnet-5", "claude-opus-5-5", "claude-opus-5", "claude-opus-4-8", "claude-opus-4-7", "claude-opus-4-6", "claude-sonnet-4-6", "claude-haiku-4-5"];
export const ANTHROPIC_MODEL_CONTEXT_WINDOWS: Record<string, number> = { "claude-fable-5-1": 1_000_000, "claude-sonnet-5-5": 1_000_000, "claude-sonnet-5": 1_000_000, "claude-fable-5": 1_000_000, "claude-opus-5-5": 1_000_000, "claude-opus-5": 1_000_000, "claude-opus-4-8": 1_000_000, "claude-opus-4-7": 1_000_000, "claude-opus-4-6": 1_000_000, "claude-sonnet-4-6": 1_000_000, "claude-haiku-4-5": 200_000 };
// All seeded Claude models support vision: https://platform.claude.com/docs/en/models/overview
export const ANTHROPIC_MODEL_INPUT_MODALITIES: Record<string, string[]> = Object.fromEntries(
ANTHROPIC_MODELS.map(id => [id, ["text", "image"]]),
);
// Every current Claude family accepts at least 64k output tokens (Haiku 4.5 / Sonnet 4.x
// through Opus 5 and Fable 5), so this stays the floor for a model id not listed in
// ANTHROPIC_MODEL_MAX_OUTPUT_TOKENS below.
export const ANTHROPIC_DEFAULT_MAX_OUTPUT_TOKENS = 64_000;
/**
* Per-model synchronous Messages API output maxima, each read from that model's own overview
* page on platform.claude.com (`/docs/en/models/<slug>/overview`), which states the number in
* its header line and again in the Capabilities table. Deliberately NOT the 300K Message
* Batches extended-output beta, which needs `output-300k-2026-03-24` and does not apply to the
* synchronous route these providers use.
*
* Every seeded Claude model except Haiku 4.5 is 128K. Haiku 4.5 is genuinely 64K and is left
* out so it keeps inheriting ANTHROPIC_DEFAULT_MAX_OUTPUT_TOKENS.
*
* Without this the 64000 default stood in for all of them, which understates the real ceiling
* by half on every Opus, Sonnet and Fable model. `resolveOutputCeiling` is the value the
* adapter budgets against when a caller declares no `max_tokens`, so a 128K-capable model
* stopped at 64000 with `stop_reason: max_tokens`; the same figure feeds the combo admission
* reserve in `checkComboTargetInputAdmission`, where understating a target's output makes its
* headroom arithmetic wrong.
*/
export const ANTHROPIC_MODEL_MAX_OUTPUT_TOKENS: Record<string, number> = {
"claude-fable-5-1": 128_000,
"claude-fable-5": 128_000,
"claude-sonnet-5-5": 128_000,
"claude-sonnet-5": 128_000,
"claude-opus-5-5": 128_000,
"claude-opus-5": 128_000,
"claude-opus-4-8": 128_000,
"claude-opus-4-7": 128_000,
"claude-opus-4-6": 128_000,
"claude-sonnet-4-6": 128_000,
};
/**
* The effort rungs opencodex exposes for native Anthropic models. Without this the
* providers advertised no ladder at all, so every client that keys its effort control off
* `reasoningEfforts` — Aside and the rest of the Pi-shaped exports — wrote these models
* with no control, while the SAME Claude models routed through `cursor` or
* `google-antigravity` had one.
*
* This is an opencodex ladder, not a claim that each model takes `output_config.effort`.
* The adapter serves two wire shapes (src/adapters/anthropic.ts): adaptive families
* (fable, sonnet >= 5, opus >= 4.7) send the effort directly, while opus 4.6, sonnet 4.6
* and haiku 4.5 take the legacy path where `reasoningBudget` TRANSLATES each rung into
* `thinking.budget_tokens`. Anthropic documents `low|medium|high|max` for the 4.6 models
* and no effort parameter at all for haiku 4.5; the budget translation is what makes five
* rungs meaningful there, and it clamps below `max_tokens` so none of them 400.
*
* Deliberately excluded, each because advertising it would offer a control that does not
* do what it says:
* - `minimal`: `adaptiveEffort` rewrites it to `low` (the adaptive wire 400s on it), so
* it is not a distinct setting.
* - `none`: only sonnet >= 5 accepts an explicit thinking disable
* (`EXPLICIT_THINKING_DISABLE_FAMILY_MINIMUMS`); Fable rejects one outright.
* - `ultra`: not an Anthropic concept, and it is degraded to `max` at the request
* boundary anyway (src/responses/parser.ts).
*/
export const ANTHROPIC_REASONING_EFFORTS = ["low", "medium", "high", "xhigh", "max"];
export const ANTHROPIC_MODEL_REASONING_EFFORTS: Record<string, string[]> = Object.fromEntries(
ANTHROPIC_MODELS.map(id => [id, [...ANTHROPIC_REASONING_EFFORTS]]),
);
// 260814 GLM-5.3 is registered pre-emptively alongside 5.2 everywhere 5.2 appears. Z.AI's
// devpack "How to Switch Models" page (docs.z.ai/devpack/latest-model) lists glm-5.3 and
// glm-5.3[1m] as Coding Plan ids on the unchanged endpoints; the capability and pricing
// tables were not published yet, so every 5.3 row mirrors its 5.2 sibling until they settle.
// The non-Z.AI providers below are speculative on purpose: they carry 5.2 today and are
// expected to pick 5.3 up on their usual lag. Providers whose live /v1/models discovery is
// enabled self-correct on the next successful fetch; static ones need a follow-up refresh.
// Every 5.3 family member, so the effort ladder, the default effort and the output
// cap are derived in ONE place. `glm-5.3-flash` was seeded into the model list and
// the context map by hand and left out of this constant, which meant it advertised
// a 1M context with a null effort ladder, no default effort and no output cap while
// its siblings carried three tiers, a `max` default and 131072 tokens. A member
// added to the list but not to the family is a model whose metadata silently
// disappears.
export const ZAI_GLM_53_MODELS = ["glm-5.3", "glm-5.3[1m]", "glm-5.3-flash"];
export const ZAI_GLM_52_MODELS = ["glm-5.2", "glm-5.2[1m]"];
export const ZAI_GLM_5X_MODELS = [...ZAI_GLM_53_MODELS, ...ZAI_GLM_52_MODELS];
/**
* The 5.x rows whose images the PROXY has to describe, which is NOT the same set as
* the 5.x rows themselves.
*
* `glm-5.3-flash` is a native VLM (docs.z.ai/guides/vlm/glm-5.3-flash), so listing it
* in `noVisionModels` sent an image through the vision sidecar and handed the model a
* text description of a picture it could have read itself - no error, worse answer,
* extra call. The correction commit fixed the Alibaba entries and left the eight
* providers that reach this constant behind.
*
* Kept separate from ZAI_GLM_5X_MODELS rather than filtered at each use site: that
* constant also drives `modelSupportsReasoningSummaries` and
* `preserveReasoningContentModels`, where flash DOES belong.
*/
export const ZAI_GLM_5X_SIDECAR_VISION_MODELS = ZAI_GLM_5X_MODELS.filter(id => id !== "glm-5.3-flash");
/**
* Positive input-modality declaration for the Chat-path GLM rows.
*
* `noVisionModels` already keeps Flash out of the vision sidecar, but that is a NEGATIVE
* statement: it stops a detour without telling the catalog what the model can read. With
* no `modelInputModalities` entry, `configuredInputModalities` returns undefined and the
* catalog falls through to the `["text"]` floor, so every client export (ZCode, Pi, OMP)
* listed a native VLM as text-only and its picker refused to attach an image.
*
* The Responses sibling row below already declares this positively, so the same model was
* described two different ways in one registry.
*
* Authoritative source: `GET https://api.z.ai/api/v1/models` returns `input_modalities:
* ["text"]` for glm-5.3 and `["text", "image"]` for glm-5.3-flash (captured in
* devlog/_plan/260912_zcode_protocol_and_catalog/evidence/zai-responses-models.json).
* docs.z.ai/devpack/latest-model says the same in prose: "GLM-5.3 is a text-only model...
* GLM-5.3-FLASH is a multimodal model". Upstream also lists video and file for Flash;
* neither the internal vocabulary nor the export vocabulary can express them, so `image`
* is where this stops.
*/
export const ZAI_GLM_5X_INPUT_MODALITIES: Record<string, string[]> = {
...Object.fromEntries(ZAI_GLM_5X_SIDECAR_VISION_MODELS.map(id => [id, ["text"]])),
"glm-5.3-flash": ["text", "image"],
};
export const ZAI_GLM_52_REASONING_EFFORTS = ["low", "medium", "high", "xhigh", "max"];
/**
* GLM-5.3 does NOT share 5.2's five-tier ladder. docs.z.ai/devpack/latest-model folds every
* incoming effort into three effective tiers — low/minimal/light -> low, medium/high -> high,
* xhigh/max/ultra -> max — with max as both the default and the unknown-value fallback.
* Advertising five levels would publish two picker rows that are indistinguishable on the wire,
* so only the effective tiers are exposed (same treatment Cursor and Baseten already give GLM).
*/
export const ZAI_GLM_53_REASONING_EFFORTS = ["low", "high", "max"];
/** Per-model ladders for the Coding Plan rows: 5.3 gets its three effective tiers, 5.2 keeps five. */
export const ZAI_GLM_5X_REASONING_EFFORTS: Record<string, string[]> = {
...Object.fromEntries(ZAI_GLM_53_MODELS.map(id => [id, ZAI_GLM_53_REASONING_EFFORTS])),
...Object.fromEntries(ZAI_GLM_52_MODELS.map(id => [id, ZAI_GLM_52_REASONING_EFFORTS])),
};
// 260710 MiniMax models and context windows: Tier-2 evidence in
// devlog/_plan/260710_provider_hardening/002_research_cn.md.
// 260930 MiniMax-M3.1-Flash-Preview: Token Plan / MiniMax Code only, 1M context, thinking
// always on (effort none or thinking disabled answers 400 code 2013), omitted effort = max.
// It returns thinking as reasoning_content and ignores reasoning_split. The live /v1/models
// roster does not list it yet. Evidence: devlog/_plan/260930_minimax_m31_flash_preview/.
export const MINIMAX_M31_FLASH_PREVIEW = "MiniMax-M3.1-Flash-Preview";
export const MINIMAX_MODELS = [
MINIMAX_M31_FLASH_PREVIEW,
"MiniMax-M3",
"MiniMax-M2.7", "MiniMax-M2.7-highspeed",
"MiniMax-M2.5", "MiniMax-M2.5-highspeed",
"MiniMax-M2.1", "MiniMax-M2.1-highspeed",
"MiniMax-M2",
];
/** The eight-id roster every MiniMax preset seeded from 2026-07-10 until the preview landed. */
export const MINIMAX_MODELS_BEFORE_M31 = MINIMAX_MODELS.filter(id => id !== MINIMAX_M31_FLASH_PREVIEW);
/** Models that honour reasoning_split and answer with structured reasoning_details. */
export const MINIMAX_REASONING_SPLIT_MODELS = MINIMAX_MODELS_BEFORE_M31;
export const MINIMAX_MODEL_CONTEXT_WINDOWS: Record<string, number> = Object.fromEntries(
MINIMAX_MODELS.map(id => [id, id === "MiniMax-M3" || id === MINIMAX_M31_FLASH_PREVIEW ? 1_000_000 : 204_800]),
);
export const MINIMAX_M3_REASONING_EFFORTS = ["low", "medium", "high", "xhigh", "max"];
/** Identity efforts on the wire; no map, so none omits the field instead of disabling thinking. */
export const MINIMAX_M31_REASONING_EFFORTS = ["low", "medium", "high", "xhigh", "max"];
export const MINIMAX_M31_DEFAULT_REASONING_EFFORT = "max";
export const MINIMAX_M3_REASONING_EFFORT_MAP: Record<string, string> = {
none: "disabled",
minimal: "disabled",
low: "disabled",
medium: "adaptive",
high: "adaptive",
xhigh: "adaptive",
max: "adaptive",
};
export const OPENAI_GPT56_MODELS = ["gpt-5.6", "gpt-5.6-sol", "gpt-5.6-terra", "gpt-5.6-luna"];
export const OPENAI_GPT56_PRO_MODELS = ["gpt-5.6-sol-pro", "gpt-5.6-terra-pro", "gpt-5.6-luna-pro"];
export const OPENAI_API_GPT56_CONTEXT_WINDOW = 1_050_000;
export const OPENAI_API_GPT56_CONTEXT_WINDOWS: Record<string, number> = {
...Object.fromEntries([...OPENAI_GPT56_MODELS, ...OPENAI_GPT56_PRO_MODELS].map(id => [id, OPENAI_API_GPT56_CONTEXT_WINDOW])),
"gpt-5.5": OPENAI_API_GPT56_CONTEXT_WINDOW,
};
export const OPENAI_API_GPT56_MAX_INPUT_TOKENS: Record<string, number> = {
...Object.fromEntries([...OPENAI_GPT56_MODELS, ...OPENAI_GPT56_PRO_MODELS].map(id => [id, 922_000])),
"gpt-5.5": 922_000,
};
export const OPENAI_API_GPT56_VIRTUAL_MODELS: Record<string, { wireModelId: string; reasoningMode: "pro" }> = {
"gpt-5.6-sol-pro": { wireModelId: "gpt-5.6-sol", reasoningMode: "pro" },
"gpt-5.6-terra-pro": { wireModelId: "gpt-5.6-terra", reasoningMode: "pro" },
"gpt-5.6-luna-pro": { wireModelId: "gpt-5.6-luna", reasoningMode: "pro" },
};
export const OPENAI_API_GPT56_REASONING_EFFORTS = ["low", "medium", "high", "xhigh", "max"];
/**
* GPT-6 Sol and Luna on the OpenAI API (released 2026-09-22,
* https://developers.openai.com/api/docs/changelog). Added 2026-09-23 ahead of live discovery; the
* API window is not published yet, so the rows mirror gpt-6-astra's 1,050,000 / 922,000 API seed.
*
* GPT-6.1 Sol (released 2026-09-29) publishes the same numbers on its own model page:
* 1,050,000 context, 922,000 max input, 128,000 max output, efforts low..max
* (https://developers.openai.com/api/docs/models/gpt-6.1-sol, checked 2026-09-30).
*/
export const OPENAI_GPT6_MODELS = ["gpt-6.1-sol", "gpt-6-sol", "gpt-6-luna"];
/*
* Meta Model API (https://api.meta.ai/v1) — published ladder, deliberately NOT the
* house set. dev.meta.ai/docs/reasoning lists "none", "minimal", "low", "medium",
* "high", "xhigh" and then excludes "none" for this family: "not supported by Muse
* Spark and returns HTTP 400". "max" and "ultra" are absent from the vendor's list
* entirely, so appending one by family resemblance would invent a wire value.
*
* Corroborated on a second surface: an unauthenticated OpenCode Zen probe of
* muse-spark-1.3-contributor-free (2026-09-03) accepted minimal..xhigh, rejected
* max/ultra with `unknown variant`, and rejected none with "does not support none
* with this model".
*/
export const META_MUSE_REASONING_EFFORTS = ["minimal", "low", "medium", "high", "xhigh"];
/*
* Identity wire map. `requestToCodexEffort` (src/reasoning-effort.ts) rewrites
* `minimal` to `low` unless a model-scoped wire map says otherwise, so without this
* the picker would advertise an effort the wire never sends — and a registry-array
* assertion would pass while the request body was wrong. Identity because Meta's
* values ARE the Codex names.
*/
export const META_MUSE_REASONING_EFFORT_MAP: Record<string, string> = Object.fromEntries(
META_MUSE_REASONING_EFFORTS.map(effort => [effort, effort]),
);
/*
* Muse Code credentials have a separate capability contract. Meta's authenticated
* /muse-code/models roster advertises max for both 1.3 models, and the Responses API
* accepts it when the request identifies the Muse client surface.
*/
export const META_MUSE_CODE_REASONING_EFFORTS = [...META_MUSE_REASONING_EFFORTS, "max"];
export const META_MUSE_CODE_REASONING_EFFORT_MAP: Record<string, string> = Object.fromEntries(
META_MUSE_CODE_REASONING_EFFORTS.map(effort => [effort, effort]),
);
/** Both Muse Spark 1.3 tiers publish a 1,048,576-token window (dev.meta.ai/docs/models). */
export const META_MUSE_CONTEXT_WINDOW = 1_048_576;
export const META_MUSE_MODELS = ["muse-spark-1.3", "muse-spark-1.3-contributor"];
/**
* Daybreak program aliases. These `-latest` ids are the stable contract: OpenAI repoints
* them at newer snapshots over time (red -> gpt-5.6-cyber, blue -> gpt-5.6-sol as of
* 2026-08-11), so registering the ALIAS inherits future model swaps while a pinned
* snapshot id would silently go stale. Snapshot ids are deliberately absent here.
* Responses-only per both published endpoint tables (`v1/chat/completions` is marked
* Not supported) — never add these to a chat-completions provider. Access needs separate
* Daybreak approval and provisioning, so neither is ever a default.
* Verified 2026-08-11: developers.openai.com/api/docs/models/daybreak-red-latest.md
* and .../daybreak-blue-latest.md
*/
export const OPENAI_DAYBREAK_MODELS = ["daybreak-red-latest", "daybreak-blue-latest"];
export const OPENAI_DAYBREAK_CONTEXT_WINDOWS: Record<string, number> = {
"daybreak-red-latest": 400_000,
"daybreak-blue-latest": 1_050_000,
};
export const OPENAI_DAYBREAK_MAX_INPUT_TOKENS: Record<string, number> = {
"daybreak-red-latest": 272_000,
"daybreak-blue-latest": 922_000,
};
/**
* Neither Daybreak page publishes a reasoning-effort ladder. An explicit empty array means
* "expose no effort control"; OMITTING the key would instead fall back to the full routed
* ladder (`configuredReasoningEfforts` returns undefined -> `applyReasoningLevels` uses
* ROUTED_REASONING_LEVELS), which would advertise efforts the models never documented.
* `noReasoningModels` is wrong here: both pages document reasoning-token support, so these
* are reasoning models with no *selectable* ladder.
*/
export const OPENAI_DAYBREAK_REASONING_EFFORTS: Record<string, string[]> = Object.fromEntries(
OPENAI_DAYBREAK_MODELS.map(id => [id, [] as string[]]),
);
export const OPENROUTER_GPT56_MODELS = OPENAI_GPT56_MODELS.map(id => `openai/${id}`);
export const XAI_MODELS = [
"grok-4.7",
"grok-4.6",
"grok-4.5",
"grok-4.3",
"grok-4.20-multi-agent-0309",
"grok-4.20-0309-reasoning",
"grok-4.20-0309-non-reasoning",
"grok-build-0.1",
"grok-composer-2.5-fast",
];
// OpenRouter's live /endpoints routes report 1,050,000; keep this separate from the
// unverified OpenAI API-key seed. Evidence: devlog/_plan/260710_provider_hardening/003_research_aggregators.md.
export const OPENROUTER_GPT56_CONTEXT_WINDOW = 1_050_000;
export const OPENROUTER_GPT56_CONTEXT_WINDOWS = {
"openai/gpt-5.6-sol": OPENROUTER_GPT56_CONTEXT_WINDOW,
"openai/gpt-5.6-terra": OPENROUTER_GPT56_CONTEXT_WINDOW,
"openai/gpt-5.6-luna": OPENROUTER_GPT56_CONTEXT_WINDOW,
// 260923 preemptive: GPT-6 Sol/Luna ahead of OpenRouter's own listing; same window as GPT-5.6.
"openai/gpt-6-sol": OPENROUTER_GPT56_CONTEXT_WINDOW,
"openai/gpt-6-luna": OPENROUTER_GPT56_CONTEXT_WINDOW,
// Live /api/v1/models on 2026-09-30: context_length 1,050,000, max_completion_tokens 128,000.
"openai/gpt-6.1-sol": OPENROUTER_GPT56_CONTEXT_WINDOW,
};
/**
* Vendor thinking-toggle models (MiMo v2.x, GLM 5/5.1 on Zen Go): the wire knob is
* `thinking: {type: enabled|disabled}` — a binary. Advertise the full Codex picker ladder
* and map efforts onto the toggle. Zen Go
* pass-through probed live 2026-07-07 (glm-5.2 toggle verified; mimo/minimax accept shape).
*/
export const THINKING_TOGGLE_EFFORTS = ["low", "medium", "high", "xhigh", "max"];
export const THINKING_TOGGLE_MAP: Record<string, string> = {
none: "disabled",
minimal: "disabled",
low: "disabled",
medium: "enabled",
high: "enabled",
xhigh: "enabled",
max: "enabled",
};
export const OPENCODE_GO_THINKING_TOGGLE_MODELS = [
"mimo-v2.5", "mimo-v2.5-pro", "glm-5", "glm-5.1",
// V2.6 keeps the vendor's thinking toggle; listed ahead of a Go probe (preemptive, 2026-09-23).
"mimo-v2.6-pro", "mimo-v2.6-flash",
];
/**
* Zhipu's domestic BigModel platform. Text families first, then the vision member: modalities are
* declared per model because `noVisionModels` means the opposite of "text only" here — it routes
* images through the proxy's vision sidecar (src/codex/catalog/provider-fetch.ts), a claim nobody
* has verified for BigModel-hosted GLM.
*/
// `glm-5.3-flash` is deliberately absent: it is a native VLM
// (docs.z.ai/guides/vlm/glm-5.3-flash), unlike glm-5.3 itself.
export const ZHIPU_BIGMODEL_TEXT_MODELS = ["glm-4.6", "glm-4.7", "glm-4.7-flash", "glm-5", "glm-5.1", "glm-5.2", "glm-5.3"];
export const ZHIPU_BIGMODEL_MODELS = [...ZHIPU_BIGMODEL_TEXT_MODELS, "glm-4.6v"];
export const ZHIPU_BIGMODEL_INPUT_MODALITIES: Record<string, string[]> = {
...Object.fromEntries(ZHIPU_BIGMODEL_TEXT_MODELS.map(id => [id, ["text"]])),
"glm-4.6v": ["text", "image"],
};
export const ZHIPU_BIGMODEL_THINKING_TOGGLE_MODELS = ["glm-4.6", "glm-4.7", "glm-5", "glm-5.1", "glm-5.2", "glm-5.3", "glm-5.3-flash"];
export const THINKING_BUDGET_EFFORTS = ["low", "medium", "high", "xhigh", "max"];
// Qwen3.8-Max is the first Qwen3.x model with official direct `reasoning_effort` support.
// Evidence: https://qwen.ai/blog?id=qwen3.8
export const QWEN38_REASONING_EFFORTS = ["low", "medium", "xhigh"];
export const THINKING_BUDGET_MODELS = [
"qwen3.5-397b", "qwen3.6-35b",
"qwen3.5-plus", "qwen3.6-plus", "qwen3.7-max", "qwen3.7-plus",
];
export const OPENCODE_GO_THINKING_BUDGET_MODELS = ["qwen3.5-plus", "qwen3.6-plus", "qwen3.7-max", "qwen3.7-plus"];
/*
* DeepSeek moved the whole V4 name set on 2026-09-10. V4.1-Flash ships as deepseek-flash
* on the first-party API; deepseek-v4-flash and the vision preview retire as models but
* keep routing there as compatibility aliases, and deepseek-v4-pro follows from
* 2026-09-14 04:00 UTC. Evidence: https://api-docs.deepseek.com/news/news260910/.
*
* The spelling differs by who serves it, so one shared list cannot express it: the
* first-party API answers to deepseek-flash, while the Zen gateway exposes the route as
* deepseek-v4.1-flash (issue #4253, PR #4258). Vendor-hosted rosters (Volcengine plan
* snapshots, Alibaba) publish on their own schedule and keep the legacy set until they say
* otherwise - a first-party retirement notice does not end their deployment.
*/
export const DEEPSEEK_V4_LEGACY_MODELS = ["deepseek-v4-flash"];
/*
* `deepseek-v4-pro` is deliberately absent from both live sets. DeepSeek retires it from
* 2026-09-14 04:00 UTC and routes its requests to V4.1-Flash until a V4.1 Pro exists, so a
* row here would advertise a Pro context window and Pro pricing for a route that serves
* Flash. The retirement is followed through every roster in this file, including the
* vendor-hosted ones; providers that discover their models live are handled by
* `ROUTED_MODEL_COMPATIBILITY_EXCLUSIONS` because deleting a row there removes the
* model's capabilities rather than the model.
*/
export const DEEPSEEK_NATIVE_THINKING_MODELS = ["deepseek-flash", "deepseek-v4-flash"];
export const DEEPSEEK_GATEWAY_THINKING_MODELS = ["deepseek-v4.1-flash", "deepseek-v4-flash"];
/*
* DeepSeek's legacy vision preview id (released 2026-08-21). First-party probes
* in #4436 resolve it to image-capable `deepseek-flash`; retain the existing
* declarations because gateway support is specific to each served identifier.
*/
export const DEEPSEEK_VISION_PREVIEW_MODEL = "deepseek-v4-flash-vision-exp";
/**
* CommandCode routes verified to accept image input end-to-end (#2406).
*
* Verified-negative and therefore deliberately ABSENT: deepseek/deepseek-v4-flash,
* zai-org/GLM-5.2, zai-org/GLM-5.3. Those
* routes accept the request and drop the image, which is worse than declining it — the
* model answers about an image it never saw. Do not add an id here on family resemblance;
* capability intersection trusts this map.
*/
export const COMMAND_CODE_IMAGE_MODELS = [
`deepseek/${DEEPSEEK_VISION_PREVIEW_MODEL}`,
// Probed 2026-09-18 through a running 2.58.0 proxy: a 3x3 random-color grid
// (180x180 PNG, six candidate colors) came back 9/9 correct both as a user
// message and as a tool_result, and the request logs show the route served
// the image natively — no vision-sidecar call in either window. #4505 asked
// for exactly this upstream probe before promoting the id. The sibling
// deepseek/deepseek-v4-flash route remains verified-negative above.
"deepseek/deepseek-v4.1-flash",
"gpt-5.6-luna",
"gpt-5.6-sol",
"MiniMaxAI/MiniMax-M3",
"moonshotai/Kimi-K3",
"meta/muse-spark-1.3",
"meta/muse-spark-1.3-contributor",
"meta/muse-spark-1.2",
"meta/muse-spark-1.2-contributor",
// Live 2026-09-23 3x3 random-color grid (180x180) via ocx 2.62.0:
// 4.7 read 9/9 in user messages and tool results; 4.6 read 9/9 and 8/9.
// Neither route requested a vision sidecar. Evidence:
// devlog/_plan/260923_grok47_parity/010_probe-evidence.md.
"xai/grok-4.6",
"xai/grok-4.7",
// Native Z.AI VLM (docs.z.ai/guides/vlm/glm-5.3-flash). This exact id is already
// classified as natively vision-capable in NVIDIA_NIM_VISION_MODELS in this file;
// it is not one of the verified-negative ids the header names (those are
// deepseek/deepseek-v4-flash, zai-org/GLM-5.2, zai-org/GLM-5.3 —
// different ids). Adding it on the shared GLM-5.3 prefix would be the family-
// resemblance mistake the header forbids; the VLM docs are the evidence (#4505).
"z-ai/glm-5.3-flash",
] as const;
/**
* Native image stays sourced from COMMAND_CODE_IMAGE_MODELS. Text-only routes
* sit beside that list so the catalog can still advertise sidecar coverage
* without claiming the gateway itself accepts a picture. A positive text-only
* declaration makes the route a vision-sidecar consumer
* (src/vision/eligibility.ts), so the catalog advertises image input on its
* behalf — without claiming native vision — and modelInputModalities is
* per-key filled, so that reaches an existing install even when noVisionModels
* was persisted before the id joined a list.
*
* Empty as of 2026-09-18. Its only entry, deepseek/deepseek-v4.1-flash, moved
* to COMMAND_CODE_IMAGE_MODELS once the #4505-requested probe passed on both
* the user-message and tool-result paths (see the note at that entry). The
* mechanism stays for the next route that measures text-only.
*/
/**
* Command Code MiMo context windows from the live /provider/v1/models catalog (2026-09-23 fixture,
* tests/fixtures/commandcode-models.json). Model-keyed registry facts double as the router's native
* decode ids, so a cold start or failed discovery still turns `command-code/xiaomi-mimo-v2.6-pro`
* into `xiaomi/mimo-v2.6-pro` instead of sending the flattened slug upstream. They are not a roster.
*/
export const COMMAND_CODE_MIMO_CONTEXT_WINDOWS: Record<string, number> = {
"xiaomi/mimo-v2.6-pro": 1_048_576,
"xiaomi/mimo-v2.6-pro-ultraspeed": 1_048_576,
"xiaomi/mimo-v2.6-flash": 1_048_576,
"xiaomi/mimo-v2.5-pro": 1_000_000,
"xiaomi/mimo-v2.5": 1_000_000,
};
export const COMMAND_CODE_TEXT_ONLY_MODELS = [] as const;
export const COMMAND_CODE_MODEL_INPUT_MODALITIES: Record<string, ["text"] | ["text", "image"]> = {
...Object.fromEntries(COMMAND_CODE_IMAGE_MODELS.map(id => [id, ["text", "image"] as ["text", "image"]])),
...Object.fromEntries(COMMAND_CODE_TEXT_ONLY_MODELS.map(id => [id, ["text"] as ["text"]])),
};
export const OPENCODE_FREE_DEEPSEEK_MODELS = ["deepseek-v4-flash-free"];
/*
* Zen free models that reject `image_url` upstream (#1043, and the reproducible
* half of #1024).
*
* Zen publishes NO modality metadata — its `/v1/models` returns only id, object,
* created, owned_by — so this list is measured, not derived. Each id was probed
* once against https://opencode.ai/zen/v1 on 2026-08-05 with a text control first
* and then a 1x1 PNG; the six below failed the image request, four of them with
* `[404] No endpoints found that support image input` and `big-pickle` with the
* exact deserialize error quoted in #1043.
*
* `mimo-v2.5-free` and `longcat-2.0-free` ACCEPT images. They remain absent
* from the blind list and are recorded separately as positive input-modality evidence,
* so capability-positive dispatch can forward images without relying on blacklist absence.
*
* Zen's roster is discovered live while this list is static, so it is a dated
* exception list, not a capability model. Re-probe before extending it.
* Evidence: devlog/_fin/260805_bug_fix_stack/002_zen_modality_probe.md
*/
export const OPENCODE_ZEN_TEXT_ONLY_MODELS = [
"big-pickle",
"nemotron-3-ultra-free",
"ling-3.0-flash-free",
"north-mini-code-free",
"laguna-s-2.1-free",
"deepseek-v4-flash-free",
];
export const OPENCODE_ZEN_IMAGE_MODELS = ["mimo-v2.5-free", "longcat-2.0-free"] as const;
/*
* DeepSeek's Codex ladder is low/high/max. With the V4 Pro GA release
* (DeepSeek-V4-Pro-0813) the official thinking-mode table is IDENTICAL for both
* V4 models (api-docs.deepseek.com/guides/thinking_mode, verified 2026-08-13):
*
* requested | v4-flash | v4-pro
* low | low | low
* medium | high | high
* high | high | high
* xhigh | high | high
* max | max | max
*
* Before GA, Pro silently upgraded low->high and mapped xhigh->max (#1057-era
* table); the page's footnote about an early-August Pro mapping update landed
* with this GA, so Pro now advertises the same three real tiers as Flash.
*
* Two standing notes (#1057):
*
* - `xhigh` is a COMPATIBILITY ALIAS, not a native tier. It stays in the wire maps
* so existing requests and saved configs keep working, but it is not advertised.
* - `medium` has no row in the vendor table — mapping it to `high` is OUR
* compatibility choice for clients that only speak the OpenAI ladder.
*/
export const DEEPSEEK_FLASH_THINKING_EFFORTS = ["low", "high", "max"];
export const DEEPSEEK_PRO_THINKING_EFFORTS = ["low", "high", "max"];
export const DEEPSEEK_PRO_REASONING_MAP: Record<string, string> = {
low: "low",
medium: "high",
high: "high",
xhigh: "high",
max: "max",
};
export const DEEPSEEK_FLASH_REASONING_MAP: Record<string, string> = {
low: "low",
medium: "high",
high: "high",
xhigh: "high",
max: "max",
};
/**
* Flash-versus-Pro classification for DeepSeek V4 model ids, including prefixed
* (`deepseek/deepseek-v4.1-flash`) and suffixed (`deepseek-v4-flash-free`) forms.
* `tests/providers/provider-registry-parity.test.ts` enumerates every id the registry
* actually passes here, so a future id this substring test would misread cannot
* land silently.
*/
export const isDeepseekFlashModel = (modelId: string): boolean =>
modelId.toLowerCase().includes("flash");
export const deepseekThinkingEffortsFor = (modelId: string): string[] =>
isDeepseekFlashModel(modelId) ? DEEPSEEK_FLASH_THINKING_EFFORTS : DEEPSEEK_PRO_THINKING_EFFORTS;
export const deepseekReasoningMapFor = (modelId: string): Record<string, string> =>
isDeepseekFlashModel(modelId) ? DEEPSEEK_FLASH_REASONING_MAP : DEEPSEEK_PRO_REASONING_MAP;
// 260719 Alibaba Token Plan Personal Edition (China/Beijing). Keep it distinct from
// Coding Plan: the products use different exact allowlists and different base URLs.
// Evidence: https://help.aliyun.com/en/model-studio/token-plan-personal-overview
// https://help.aliyun.com/en/model-studio/token-plan-quickstart
// 260909 refresh, re-probed against the live gateway (both regions, both tiers):
// https://github.com/oliver-mee/alibaba-token-plan-wiki (machine-readable catalog).
// 260918: glm-5.3 returns. The 260909 removal was correct at the time (the id
// 404'd on every plan key), but the gateway started serving glm-5.3 on 260917:
// it now appears on /models for global Team, global Personal, and CN Team, and
// answers a completion on a Personal key (probed 260918). Contract on the plan
// gateway: effort low/high/max (default max), thinking always-on (the gateway
// rejects enable_thinking:false with 400), 1M context, 131,072 max output,
// text-only input, strict json_schema accepted. glm-5.3-flash REMAINS OUT:
// still never served by the Token Plan gateway (docs.z.ai VLM id, not plan
// entitlement).
// The Beijing preset keeps the Personal Edition subset; non-chat ids (audio/image/
// video families) stay out: they answer only on async endpoints openai-chat cannot
// reach. deepseek-v4-pro-0813 is callable but NOT listed by /models, which is the
// reason liveModels must stay false for this provider. deepseek-v4.1-flash is the
// 260910 DeepSeek rename row: listed on /models on both tiers and regions from 260915,
// hybrid thinking, vision via user message and tool result, json_object but not
// json_schema (see noJsonSchemaModels on the entries).
// Beijing serves the Personal Edition, so this is the Personal-tier roster probed
// 260909 (a strict subset of Team). deepseek-v4-pro-0813 stays out of the Beijing
// entry: its callability is only proven on Team keys, and no Personal key has been
// shown to reach it. The Beijing entry also shares the intl maps, so it carries a
// few orphan keys (kimi/glm-5/MiniMax rows); harmless, and one map beats two
// drifting ones.
export const ALIBABA_TOKEN_PLAN_MODELS = [
"qwen3.8-max", "qwen3.8-flash", "qwen3.7-max", "qwen3.7-plus", "qwen3.6-flash",
"deepseek-v4-pro", "deepseek-v4-flash-0731", "deepseek-v4.1-flash", "glm-5.2", "glm-5.3",
];
export const ALIBABA_TOKEN_PLAN_QWEN_MODELS = [
"qwen3.8-max", "qwen3.8-flash", "qwen3.7-max", "qwen3.7-plus", "qwen3.6-flash",
];
export const ALIBABA_TOKEN_PLAN_INPUT_MODALITIES: Record<string, string[]> = {
"qwen3.8-max": ["text", "image"],
"qwen3.8-flash": ["text", "image"],
"qwen3.7-max": ["text"],
"qwen3.7-plus": ["text", "image"],
"qwen3.6-flash": ["text", "image"],
"deepseek-v4-pro": ["text"],
"deepseek-v4-pro-0813": ["text"],
"deepseek-v4-flash-0731": ["text"],
// Vision probed on the plan gateway 260915 (user message and tool result, both 200).
"deepseek-v4.1-flash": ["text", "image"],
"glm-5.2": ["text"],
"glm-5.3": ["text"],
};
// 260721 Alibaba Token Plan International (ap-southeast-1 / Singapore, hardened 260721).
// Multi-vendor lineup distinct from Beijing — includes DeepSeek V4 flash, Kimi K2.7, MiniMax.
// Evidence: https://www.alibabacloud.com/help/en/model-studio/token-plan-overview
// https://qwencloud.com/pricing/token-plan (qwen3.8 metadata)
// The Team Edition roster (Singapore), verified identical to the CN Team set on 260909.
// deepseek-v4-pro is restored: it remains callable on the plan gateway (probed 260909,
// listed on /models on both regions) after being dropped as "retired" upstream.
export const ALIBABA_INTL_TOKEN_PLAN_MODELS = [
"qwen3.8-max", "qwen3.8-flash", "qwen3.7-max", "qwen3.7-plus", "qwen3.6-plus", "qwen3.6-flash",
"deepseek-v4-pro", "deepseek-v4-pro-0813", "deepseek-v4-flash", "deepseek-v4-flash-0731", "deepseek-v4.1-flash", "deepseek-v3.2",
"kimi-k2.7-code", "kimi-k2.6", "kimi-k2.5",
"glm-5.2", "glm-5.3", "glm-5.1", "glm-5",
"MiniMax-M2.5",
];
export const ALIBABA_INTL_TOKEN_PLAN_QWEN_MODELS = [
"qwen3.8-max", "qwen3.8-flash", "qwen3.7-max", "qwen3.7-plus", "qwen3.6-plus", "qwen3.6-flash",
];
// 260722 Tencent Cloud Coding Plan. The plan's model set is explicitly dynamic; these are the
// current documented ids and live discovery remains enabled so successful /models responses win.
// Tencent marks every Coding Plan model as text-only input and restricts plan keys to interactive
// coding tools (not custom application backends or non-interactive batch automation).
// Evidence: https://cloud.tencent.cn/document/product/1823/130092
export const TENCENT_CODING_PLAN_MODELS = ["tc-code-latest", "glm-5", "kimi-k2.5", "minimax-m2.5"];
// Volcengine's authenticated /api/v3/models catalog mixes chat models with embedding,
// image, video, and 3D generation resources. Keep the Codex-facing presets scoped to
// models documented for text/agent or Coding Plan use.
//
// Maintenance owner: @lidge-jun. Verified 2026-08-01 against the vendor's own docs —
// endpoints https://docs.volcengine.com/docs/82379/1528783 (Coding Plan) and
// https://docs.volcengine.com/docs/82379/2165245 (Agent Plan); Codex CLI integration
// https://www.volcengine.com/docs/82379/2556056; supported clients
// https://www.volcengine.com/docs/82379/2188957; terms https://www.volcengine.com/docs/6256/64903
// (北京火山引擎科技有限公司). Plan quota is restricted to supported AI coding tools and misuse
// is documented as grounds for suspension — see the `note` on both Plan entries.
// Report a break by opening an issue tagging the owner; the three things that rot first are the
// static catalogs (liveModels:false cannot self-heal), the base URLs, and those Plan terms.
// Full evidence ledger: devlog/_fin/260801_pr611_volcengine_evidence/000_evidence_ledger.md
export const VOLCENGINE_ARK_MODELS = [
"doubao-seed-2-1-pro-260628",
"doubao-seed-2-1-turbo-260628",
"doubao-seed-evolving",
"deepseek-v4-flash-260425",
"deepseek-v3-2-251201",
// No glm-5-3 row: Ark pins date-stamped snapshot ids (glm-5-2-260617) that cannot be
// guessed ahead of the vendor publishing them. Add it once /api/v3/models lists one.
"glm-5-2-260617",
"glm-4-7-251222",
];
export const VOLCENGINE_DOUBAO_THINKING_MODELS = [
"doubao-seed-2-1-pro-260628",
"doubao-seed-2-1-turbo-260628",
"doubao-seed-evolving",
];
export const VOLCENGINE_CODING_PLAN_MODELS = [
"ark-code-latest",
"doubao-seed-2.0-code",
"deepseek-v4-flash",
"glm-5.3",
"glm-5.3-flash",
"glm-5.2",
"kimi-k2.6",
"minimax-m3",
];
export const VOLCENGINE_AGENT_PLAN_MODELS = [
"deepseek-v4-flash",
"glm-5.3",
"glm-5.3-flash",
"glm-5.2",
"kimi-k2.6",
"minimax-m3",
"doubao-seed-2.0-pro",
];
export const VOLCENGINE_PLAN_INPUT_MODALITIES: Record<string, string[]> = {
"kimi-k2.6": ["text", "image"],
"minimax-m3": ["text", "image"],
// Native VLM (docs.z.ai/guides/vlm/glm-5.3-flash), so it is declared here and left
// out of the text-only list below.
"glm-5.3-flash": ["text", "image"],
};
// Every other Plan model is text-only. Declaring this explicitly keeps the vision
// sidecar from advertising image input for models that cannot accept it — the same
// treatment tencent-coding-plan gives its (entirely text-only) plan catalog.
export const VOLCENGINE_PLAN_TEXT_ONLY_MODELS = [
"ark-code-latest",
"doubao-seed-2.0-code",
"deepseek-v4-flash",
"glm-5.3",
"glm-5.2",
"doubao-seed-2.0-pro",
];
export const ALIBABA_INTL_TOKEN_PLAN_INPUT_MODALITIES: Record<string, string[]> = {
...ALIBABA_TOKEN_PLAN_INPUT_MODALITIES,
"qwen3.6-plus": ["text", "image"],
"deepseek-v4-flash": ["text"],
"deepseek-v3.2": ["text"],
"kimi-k2.7-code": ["text", "image"],
"kimi-k2.6": ["text", "image"],
"kimi-k2.5": ["text", "image"],
"glm-5.1": ["text"],
"glm-5": ["text"],
"MiniMax-M2.5": ["text"],
};
// Shared Token Plan metadata (260909 gateway probes; output ceilings are max_tokens
// boundary probes: accept at N, reject at N+1).
export const QWEN38_FAMILY = ["qwen3.8-max", "qwen3.8-flash"];
export const ALIBABA_TOKEN_PLAN_CONTEXT_WINDOWS: Record<string, number> = {
"qwen3.8-max": 1_000_000, "qwen3.8-flash": 1_000_000, "qwen3.7-max": 1_000_000, "qwen3.7-plus": 1_000_000,
"qwen3.6-plus": 1_000_000, "qwen3.6-flash": 1_000_000,
"deepseek-v4-pro": 1_000_000, "deepseek-v4-pro-0813": 1_000_000, "deepseek-v4-flash": 1_000_000,
"deepseek-v4-flash-0731": 1_000_000, "deepseek-v4.1-flash": 1_000_000, "deepseek-v3.2": 131_072,
"kimi-k2.7-code": 262_144, "kimi-k2.6": 262_144, "kimi-k2.5": 262_144,
"glm-5.2": 1_000_000, "glm-5.3": 1_000_000, "glm-5.1": 202_752, "glm-5": 202_752,
"MiniMax-M2.5": 196_608,
};
export const ALIBABA_TOKEN_PLAN_MAX_OUTPUT_TOKENS: Record<string, number> = {
"qwen3.8-max": 131_072, "qwen3.8-flash": 131_072, "qwen3.7-max": 131_072, "qwen3.7-plus": 131_072,
"qwen3.6-plus": 65_536, "qwen3.6-flash": 65_536,
"deepseek-v4-pro": 393_216, "deepseek-v4-pro-0813": 393_216, "deepseek-v4-flash": 393_216,
"deepseek-v4-flash-0731": 393_216, "deepseek-v4.1-flash": 393_216, "deepseek-v3.2": 65_536,
"kimi-k2.7-code": 262_144, "kimi-k2.6": 262_144, "kimi-k2.5": 98_304,
"glm-5.2": 131_072, "glm-5.3": 131_072, "glm-5.1": 128_000, "glm-5": 16_384,
"MiniMax-M2.5": 32_768,
};
export const ALIBABA_TOKEN_PLAN_NO_VISION = [
"qwen3.7-max", "deepseek-v4-pro", "deepseek-v4-pro-0813", "deepseek-v4-flash",
"deepseek-v4-flash-0731", "deepseek-v3.2", "glm-5.2", "glm-5.3", "glm-5.1", "glm-5", "MiniMax-M2.5",
];
export const ALIBABA_TOKEN_PLAN_PRESERVE_REASONING = [
"qwen3.8-max", "qwen3.8-flash", "qwen3.7-max", "qwen3.7-plus", "qwen3.6-plus", "qwen3.6-flash",
"deepseek-v4-pro", "deepseek-v4-pro-0813", "deepseek-v4-flash", "deepseek-v4-flash-0731",
"deepseek-v4.1-flash", "glm-5.2", "glm-5.3",
];
// 260717 Kimi K3: the subscription endpoint uses one upstream id (`k3`) for both
// entitlement tiers. Bare `k3` advertises the Moderato 256K ceiling; the local `[1m]`
// alias advertises Allegretto's 1M ceiling and is stripped before the upstream request.
// The separately billed Moonshot API uses `kimi-k3`.
// 260921: `k3-256k` is the same K3 served under the explicit ceiling id (verified live
// 260921: same 988-token scaffold and identity answer as bare `k3` on the same input).
// Evidence: https://www.kimi.com/code/docs/en/kimi-code/models.html
// https://www.kimi.com/code/docs/en/kimi-code/error-reference.html
export const KIMI_K3_STANDARD_CONTEXT_WINDOW = 262_144;
export const KIMI_K3_1M_CONTEXT_WINDOW = 1_048_576;
export const KIMI_CODING_K3_MODELS = ["k3", "k3[1m]", "k3-256k"];
// 260921 Kimi K2.8: `kimi-for-coding` is the stable subscription alias Moonshot re-points
// at each coding release. Live GET /coding/v1/models lists only kimi-for-coding[-highspeed],
// k3, k3-256k — the k2.x ids are retired from the subscription endpoint. Since K2.8 Preview
// the alias serves an adjustable low/high/max thinking ladder (same wire map as k3) and a
// 1M context ceiling. Verified live 260921: 350K-token request accepted; upstream rejects
// with "model token limit: 1048576" beyond that.
// Evidence: https://www.kimi.com/code/docs/en/kimi-code/models.html
export const KIMI_CODING_K28_MODELS = ["kimi-for-coding"];
export const KIMI_CODING_LIVE_MODELS = [...KIMI_CODING_K3_MODELS, ...KIMI_CODING_K28_MODELS];
export const KIMI_CODING_ADJUSTABLE_THINKING_MODELS = [...KIMI_CODING_LIVE_MODELS];
export const KIMI_LEGACY_API_MODELS = ["kimi-k2.7-code", "kimi-k2.7-code-highspeed", "kimi-k2.6", "kimi-k2.5"];
export const KIMI_API_MODELS = ["kimi-k3", ...KIMI_LEGACY_API_MODELS];
// Every kimi coding preset record - picker, context windows, locked-parameter lists -
// derives from the live ids only. seeding a retired id in a metadata list would re-arm
// the model-rename migration on every boot (#5066): the list holds the retired id but
// not the live alias, so the residue guard cannot skip it. The retired ids survive only
// in KIMI_LEGACY_API_MODELS (moonshot platform API records); model-rename-migration
// repairs saved rows still naming them.
export const KIMI_THINKING_MODELS = KIMI_CODING_LIVE_MODELS;
export const KIMI_CODING_NO_REASONING_MODELS = KIMI_CODING_LIVE_MODELS.filter(id => !KIMI_CODING_ADJUSTABLE_THINKING_MODELS.includes(id));
export const KIMI_API_NO_REASONING_MODELS = KIMI_API_MODELS.filter(id => id !== "kimi-k3");
export const KIMI_CODING_K3_REASONING_EFFORTS = ["low", "high", "max"];
export const KIMI_CODING_K3_REASONING_EFFORT_MAP: Record<string, string> = {
none: "none",
low: "low",
medium: "high",
high: "high",
xhigh: "max",
max: "max",
};
export const KIMI_CODING_REASONING_EFFORTS = Object.fromEntries(
KIMI_CODING_LIVE_MODELS.map(id => [id, KIMI_CODING_ADJUSTABLE_THINKING_MODELS.includes(id) ? KIMI_CODING_K3_REASONING_EFFORTS : []]),
);
export const KIMI_CODING_DEFAULT_REASONING_EFFORTS = Object.fromEntries(
KIMI_CODING_ADJUSTABLE_THINKING_MODELS.map(id => [id, "max"]),
);
export const KIMI_CODING_REASONING_EFFORT_MAPS = Object.fromEntries(
KIMI_CODING_ADJUSTABLE_THINKING_MODELS.map(id => [id, KIMI_CODING_K3_REASONING_EFFORT_MAP]),
);
export const KIMI_API_REASONING_EFFORTS = Object.fromEntries(
KIMI_API_MODELS.map(id => [id, id === "kimi-k3" ? ["max"] : []]),
);
export const KIMI_LOCKED_PARAMETER_MODELS = KIMI_CODING_LIVE_MODELS;
export const KIMI_AUTO_TOOL_CHOICE_ONLY_MODELS = ["kimi-for-coding"];
export const KIMI_API_MODEL_CONTEXT_WINDOWS: Record<string, number> = Object.fromEntries(
KIMI_API_MODELS.map(id => [id, id === "kimi-k3" ? KIMI_K3_1M_CONTEXT_WINDOW : 262_144]),
);
export const KIMI_API_MODEL_INPUT_MODALITIES = { "kimi-k3": ["text", "image"] };
// 260715 NVIDIA NIM kimi family (issue #126): documented served ids on integrate
// chat/completions per docs.api.nvidia.com/nim/reference/llm-apis; live /v1/models
// currently lists only kimi-k2.6 but the list is dynamic, so carry the documented family.
export const NVIDIA_NIM_KIMI_THINKING_MODELS = [
"moonshotai/kimi-k2.6", "moonshotai/kimi-k2.5", "moonshotai/kimi-k2-thinking",
];
export const NVIDIA_NIM_KIMI_MODELS = [
...NVIDIA_NIM_KIMI_THINKING_MODELS,
"moonshotai/kimi-k2-instruct", "moonshotai/kimi-k2-instruct-0905",
];
/**
* 260804 issue #956: NIM publishes no input-modality metadata on `/v1/models`, so the
* registry is the only source of truth for which models can see images.
*
* Two lists, both verified per-model against NVIDIA documentation on 2026-08-04
* (build.nvidia.com model pages and docs.api.nvidia.com/nim/reference/*). Evidence and
* the per-id audit: devlog/_fin/260804_stack7_service_vision/011_nim_id_audit.md.
*
* Read `noVisionModels` carefully — it lists models that CANNOT see images, which is
* what routes them through the proxy's vision sidecar (src/vision/index.ts) and makes the
* catalog advertise image input for them. Membership is wrong in BOTH directions:
* - a text-only model missing from it keeps issue #956 (images blocked or rejected);
* - a vision model wrongly IN it gets its image silently replaced by another model's
* text description — no error, worse answers, extra cost.
*
* A new NIM id must be classified DELIBERATELY against its NVIDIA page, never assumed
* from its name: `google/gemma-4-31b-it` carries no vision marker yet accepts images,
* `-vl` also appears on embedding/reranking models, and `google/codegemma-7b` is
* text-only while `google/codegemma-1.1-7b` has no current page at all. An unclassified
* id is intentionally left alone rather than defaulted, because NIM serves non-chat
* endpoints (embeddings, rerankers, guards, OCR) that reach the same code path.
*/
export const NVIDIA_NIM_VISION_MODELS = [
"meta/llama-3.2-11b-vision-instruct", "meta/llama-3.2-90b-vision-instruct",
"nvidia/llama-3.1-nemotron-nano-vl-8b-v1", "nvidia/nemotron-nano-12b-v2-vl",
"nvidia/nemotron-3-nano-omni-30b-a3b-reasoning", "nvidia/cosmos3-nano-reasoner",
"nvidia/ising-calibration-1.5-31b", "nvidia/ising-calibration-1-35b-a3b",
"google/gemma-4-31b-it", "google/diffusiongemma-26b-a4b-it",
"minimaxai/minimax-m3", "moonshotai/kimi-k2.6", "moonshotai/kimi-k2.5",
"stepfun-ai/step-3.7-flash", "thinkingmachines/inkling",
"mistralai/mistral-medium-3.5-128b",
"z-ai/glm-5.3-flash",
];
/**
* The catalog advertises image input only for `noVisionModels` members, so a natively
* vision-capable model would otherwise be published as text-only and the Codex app would
* block attachments before the native path ever runs.
*/
export const NVIDIA_NIM_VISION_INPUT_MODALITIES: Record<string, string[]> = Object.fromEntries(
NVIDIA_NIM_VISION_MODELS.map(id => [id, ["text", "image"]]),
);
/**
* Text-only NIM chat models — 26 ids, each carrying an explicit `Input Modalities: Text`
* (or equivalent) on its NVIDIA page. PR #964 proposed ~64; six of those are natively
* image-capable and live in NVIDIA_NIM_VISION_MODELS above, and 32 more had no current
* NVIDIA page and were dropped rather than assumed.
*
* kimi-k2-thinking and kimi-k2-instruct are text-only while k2.5/k2.6 are not — vision
* and reasoning are independent axes, so all four stay in NVIDIA_NIM_KIMI_MODELS for
* reasoning suppression regardless of which list they appear in here.
*/
export const NVIDIA_NIM_NO_VISION_MODELS = [
"deepseek-ai/deepseek-v4-flash",
"google/codegemma-7b",
"meta/llama-3.1-70b-instruct", "meta/llama-3.1-8b-instruct",
"meta/llama-3.2-1b-instruct", "meta/llama-3.2-3b-instruct",
"meta/llama-3.3-70b-instruct", "meta/llama2-70b",
"mistralai/mistral-7b-instruct-v0.3", "mistralai/mistral-nemotron",
"moonshotai/kimi-k2-thinking", "moonshotai/kimi-k2-instruct",
"nvidia/llama-3.1-nemotron-nano-8b-v1", "nvidia/llama-3.1-nemotron-ultra-253b-v1",
"nvidia/llama-3.3-nemotron-super-49b-v1", "nvidia/llama-3.3-nemotron-super-49b-v1.5",
"nvidia/nemotron-3-nano-30b-a3b", "nvidia/nemotron-3-super-120b-a12b",
"nvidia/nemotron-3-ultra-550b-a55b", "nvidia/nemotron-mini-4b-instruct",
"nvidia/nvidia-nemotron-nano-9b-v2",
"openai/gpt-oss-120b", "openai/gpt-oss-20b",
// z-ai/glm-5.3-flash belongs in NVIDIA_NIM_VISION_MODELS, not here: Z.AI documents
// it under docs.z.ai/guides/vlm/. The header above says an id must be classified
// deliberately rather than assumed from its name, and inheriting glm-5.3's
// text-only verdict because of the shared prefix is exactly that mistake.
"poolside/laguna-xs-2.1", "z-ai/glm-5.3", "z-ai/glm-5.2",
];
export const KIMI_CODING_MODEL_CONTEXT_WINDOWS: Record<string, number> = Object.fromEntries(
KIMI_CODING_LIVE_MODELS.map(id => [id, (id === "k3[1m]" || KIMI_CODING_K28_MODELS.includes(id)) ? KIMI_K3_1M_CONTEXT_WINDOW : KIMI_K3_STANDARD_CONTEXT_WINDOW]),
);
export const KIMI_CODING_MODEL_INPUT_MODALITIES = Object.fromEntries(
KIMI_CODING_ADJUSTABLE_THINKING_MODELS.map(id => [id, ["text", "image"]]),
);
export const NEURALWATT_REASONING_HISTORY_MODELS = [
"glm-5.3", "glm-5.3-short", "glm-5.3-flash",
"glm-5.2", "glm-5.2-short",
"kimi-k2.6", "kimi-k2.7-code",
"qwen3.5-397b", "qwen3.6-35b",
];
// 260728 Baseten Model APIs: `/v1/models` owns the live lineup, while these hints
// describe only capabilities that Baseten documents per slug. Unlisted live models
// intentionally inherit the empty provider ladder instead of being advertised with
// opencodex's generic reasoning defaults. Audio is omitted because the current proxy
// request model does not carry OpenAI `audio_url` parts.
// Evidence: https://docs.baseten.co/inference/model-apis/reasoning
// https://docs.baseten.co/inference/model-apis/vision
export const BASETEN_FULL_REASONING_EFFORTS = ["low", "medium", "high", "xhigh", "max"];
export const BASETEN_MODEL_REASONING_EFFORTS: Record<string, string[]> = {
"thinkingmachines/inkling": BASETEN_FULL_REASONING_EFFORTS,
"openai/gpt-oss-120b": BASETEN_FULL_REASONING_EFFORTS,
"moonshotai/Kimi-K3": ["low", "high", "max"],
// 260814: GLM-5.3 honours low/high/max upstream, unlike 5.2's high/max on Baseten.
"zai-org/GLM-5.3": ["low", "high", "max"],
"zai-org/GLM-5.3-Fast": ["low", "high", "max"],
"zai-org/GLM-5.2": ["high", "max"],
"zai-org/GLM-5.2-Fast": ["high", "max"],
};
export const BASETEN_MODEL_REASONING_EFFORT_MAP: Record<string, Record<string, string>> = {
"thinkingmachines/inkling": { none: "none", minimal: "minimal" },
"openai/gpt-oss-120b": { none: "none", minimal: "minimal" },
"moonshotai/Kimi-K3": { none: "none" },
"zai-org/GLM-5.3": { none: "none" },
"zai-org/GLM-5.3-Fast": { none: "none" },
"zai-org/GLM-5.2": { none: "none" },
"zai-org/GLM-5.2-Fast": { none: "none" },
};
export const BASETEN_MODEL_DEFAULT_REASONING_EFFORTS: Record<string, string> = {
"thinkingmachines/inkling": "high",
"openai/gpt-oss-120b": "medium",
"moonshotai/Kimi-K3": "max",
};
export const BASETEN_MODEL_INPUT_MODALITIES: Record<string, string[]> = {
"thinkingmachines/inkling": ["text", "image"],
"moonshotai/Kimi-K2.6": ["text", "image"],
"moonshotai/Kimi-K2.7-Code": ["text", "image"],
"moonshotai/Kimi-K3": ["text", "image"],
};
// 260801 DigitalOcean and Scaleway expose OpenAI-shaped `/v1/models` rows with only
// id/object/created/owned_by, while their shared serverless catalogs also contain
// non-chat and endpoint-specific models. Fail closed by intersecting live discovery
// with ids that the providers' current first-party model tables establish for Chat
// Completions. A newly listed id therefore needs a docs-backed registry refresh before
// it can enter the Codex catalog.
// Evidence: https://docs.digitalocean.com/products/inference/details/models/
// https://docs.digitalocean.com/reference/api/reference/serverless-inference/
// https://www.scaleway.com/en/docs/generative-apis/reference-content/supported-models/
export const DIGITALOCEAN_CHAT_COMPLETION_MODELS = [
"arcee-trinity-large-thinking",
"openai-gpt-5.6-sol",
"openai-gpt-5.6-terra",
"openai-gpt-5.6-luna",
// 260923 preemptive: GPT-6 Sol/Luna ahead of DigitalOcean's model list.
"openai-gpt-6-sol",
"openai-gpt-6-luna",
"qwen3-coder-flash",
"qwen3.5-397b-a17b",
"deepseek-4-flash",
"deepseek-3.2",
"gemma-4-31B-it",
"minimax-m2.5",
"kimi-k3",
"kimi-k2.6",
"kimi-k2.5",
"llama3.3-70b-instruct",
"llama-4-maverick",
"mistral-3-14B",
"nemotron-3-ultra-550b",
"nvidia-nemotron-3-super-120b",
"nemotron-3-nano-omni",
"nemotron-nano-12b-v2-vl",
"mimo-v2.5-pro",
"glm-5.3",
"glm-5.3-flash",
"glm-5.2",
"glm-5.1",
"glm-5",
// The API reference uses this native slash id in its Chat Completions example.
"meta-llama/Meta-Llama-3.1-8B-Instruct",
] as const;
export const SCALEWAY_SERVERLESS_CHAT_MODELS = [
"glm-5.3",
"glm-5.3-flash",
"glm-5.2",
// gpt-oss-120b is intentionally omitted: Scaleway requires Responses API for tool calling,
// while this preset routes Codex agent tools through Chat Completions.
"qwen3.6-35b-a3b",
"qwen3.5-397b-a17b",
"qwen3-235b-a22b-instruct-2507",
"qwen3-coder-30b-a3b-instruct",
"gemma-4-26b-a4b-it",
"llama-3.3-70b-instruct",
"mistral-medium-3.5-128b",
"mistral-small-3.2-24b-instruct-2506",
"pixtral-12b-2409",
] as const;
export const SCALEWAY_MODEL_INPUT_MODALITIES: Record<string, string[]> = {
"pixtral-12b-2409": ["text", "image"],
};
export const UMANS_MODELS = [
"umans-coder",
"umans-kimi-k2.7",
"umans-flash",
"umans-glm-5.3",
"umans-glm-5.3-flash",
"umans-glm-5.2",
"umans-glm-5.1",
"umans-qwen3.6-35b-a3b",
];
export const UMANS_REASONING_EFFORTS = ["low", "medium", "high", "xhigh", "max"];
export const UMANS_GLM_REASONING_EFFORTS = ["high", "xhigh", "max"];
// 260814: Z.AI folds GLM-5.3 efforts into low/high/max, so `low` is a real tier here and
// `xhigh` is not distinct from `max` (docs.z.ai/devpack/latest-model).
export const UMANS_GLM_53_REASONING_EFFORTS = ["low", "high", "max"];
// `umans-glm-5.3-flash` is NOT here: Z.AI documents glm-5.3-flash under
// docs.z.ai/guides/vlm/, so it takes images natively and does not need the proxy's
// vision sidecar. The seeding pass classified it from the family name and a later
// pass corrected only some of the providers; this is one it missed.
export const UMANS_TEXT_ONLY_MODELS = ["umans-glm-5.3", "umans-glm-5.2", "umans-glm-5.1"];
export const UMANS_MODEL_CONTEXT_WINDOWS: Record<string, number> = {
"umans-coder": 262_144,
"umans-kimi-k2.7": 262_144,
"umans-flash": 262_144,
"umans-glm-5.3": 405_504,
// Mirrors the sibling this provider already carries. Umans has not published a
// separate window for the flash tier; asserting a different number would be a guess.
"umans-glm-5.3-flash": 405_504,
"umans-glm-5.2": 405_504,
"umans-glm-5.1": 202_752,
"umans-qwen3.6-35b-a3b": 262_144,
};
export const UMANS_MODEL_INPUT_MODALITIES: Record<string, string[]> = Object.fromEntries(
UMANS_MODELS.map(id => [id, UMANS_TEXT_ONLY_MODELS.includes(id) ? ["text"] : ["text", "image"]]),
);
export const CLINE_PASS_MODELS = [
"cline-pass/glm-5.3",
"cline-pass/glm-5.3-flash",
"cline-pass/glm-5.2",
"cline-pass/kimi-k3",
"cline-pass/kimi-k2.7-code",
"cline-pass/kimi-k2.6",
"cline-pass/deepseek-v4-flash",
"cline-pass/mimo-v2.6-pro",
"cline-pass/mimo-v2.6-flash",
"cline-pass/mimo-v2.5",
"cline-pass/mimo-v2.5-pro",
"cline-pass/minimax-m3",
"cline-pass/qwen3.8-max",
"cline-pass/qwen3.7-max",
"cline-pass/qwen3.7-plus",
];
export const ORCAROUTER_MODEL_DISCOVERY: ProviderModelDiscoverySpec = {
path: "models",
query: { capability: "chat" },
maxResponseBytes: 512 * 1024,
maxModels: 512,
filter: {
anyOf: [{
path: ["supported_endpoint_types"],
containsAny: ["openai", "openai-response", "anthropic", "gemini"],
caseInsensitive: true,
}],
noneOf: [{
path: ["supported_endpoint_types"],
containsAny: ["image-generation", "openai-video", "jina-rerank"],
caseInsensitive: true,
}],
},
};
// Preserve the previously verified cold-start catalog. Live discovery remains authoritative
// when it succeeds, but a temporary catalog outage must not erase the provider's known-good
// selectors from the picker. `orcarouter/auto` is intentionally retained here even though the
// public catalog did not enumerate it at the latest verification (2026-09-07).
export const ORCAROUTER_MODELS = [
"openai/gpt-5.5",
"anthropic/claude-opus-4.8",
"google/gemini-3.5-flash",
"orcarouter/auto",
];
export const ORCAROUTER_MODEL_REASONING_EFFORTS = {
// Live /models currently exposes ids and modalities, not the accepted reasoning ladder.
"openai/gpt-5.5": ["low", "medium", "high", "xhigh"],
};
export const CLINE_PASS_MODEL_CONTEXT_WINDOWS: Record<string, number> = {
"cline-pass/glm-5.3": 1_048_576,
"cline-pass/glm-5.3-flash": 1_048_576,
"cline-pass/glm-5.2": 1_048_576,
"cline-pass/kimi-k3": 1_048_576,
"cline-pass/kimi-k2.7-code": 262_144,
"cline-pass/kimi-k2.6": 262_144,
"cline-pass/deepseek-v4-flash": 1_048_576,
"cline-pass/mimo-v2.6-pro": 1_048_576,
"cline-pass/mimo-v2.6-flash": 1_048_576,
"cline-pass/mimo-v2.5": 1_050_000,
"cline-pass/mimo-v2.5-pro": 1_050_000,
"cline-pass/minimax-m3": 1_048_576,
"cline-pass/qwen3.7-max": 1_000_000,
"cline-pass/qwen3.7-plus": 1_000_000,
};
export const CLINE_PASS_IMAGE_MODELS = new Set([
"cline-pass/kimi-k3",
"cline-pass/kimi-k2.7-code",
"cline-pass/kimi-k2.6",
"cline-pass/mimo-v2.5",
"cline-pass/minimax-m3",
"cline-pass/qwen3.7-plus",
// Native VLM (docs.z.ai/guides/vlm/), so its images do not go through the proxy's
// sidecar. Adding it here moves it out of CLINE_PASS_TEXT_ONLY_MODELS and flips its
// declared modalities to ["text", "image"] in one edit, because both are derived
// from this set.
"cline-pass/glm-5.3-flash",
]);
export const CLINE_PASS_MODALITY_KNOWN_MODELS = CLINE_PASS_MODELS.filter(id => id !== "cline-pass/qwen3.8-max");
export const CLINE_PASS_TEXT_ONLY_MODELS = CLINE_PASS_MODALITY_KNOWN_MODELS.filter(id => !CLINE_PASS_IMAGE_MODELS.has(id));
export const CLINE_PASS_MODEL_INPUT_MODALITIES: Record<string, string[]> = Object.fromEntries(
CLINE_PASS_MODALITY_KNOWN_MODELS.map(id => [id, CLINE_PASS_IMAGE_MODELS.has(id) ? ["text", "image"] : ["text"]]),
);
// Opper seed: bare *pool* names. A pool is every provider Opper serves that model through; Opper
// picks the route per request. Each name is the `.model` of a `pooled: true` entry in the public
// catalogue snapshot supplied by the original provider author
// (https://api.opper.ai/v3/models?limit=2000, captured 2026-09-14); `vendor/model` ids
// (anthropic/claude-sonnet-4-6) pin one route and stay valid, they are just not seeded.
// 260923: `claude-opus-5-5` pool (anthropic, aws eu, vertex, vertex-eu members; all 1M / 128K,
// vision) read from the same catalogue endpoint the day after Anthropic's release.
export const OPPER_MODELS = [
"claude-sonnet-4-6",
"claude-opus-5-5",
"claude-opus-5",
"gpt-5.5",
"gpt-5.4-mini",
"gemini-3.8-flash",
"deepseek-v4-pro",
"kimi-k3",
"mistral-large-2512",
];
// Smallest value across each pool's members in that snapshot, capped at the lab model's own limit
// (kimi-k3 output); live discovery owns which models exist.
export const OPPER_MODEL_CONTEXT_WINDOWS: Record<string, number> = {
"claude-sonnet-4-6": 1_000_000,
"claude-opus-5-5": 1_000_000,
"claude-opus-5": 1_000_000,
"gpt-5.5": 1_050_000,
"gpt-5.4-mini": 400_000,
"gemini-3.8-flash": 1_048_576,
"deepseek-v4-pro": 1_000_000,
"kimi-k3": 1_048_576,
"mistral-large-2512": 256_000,
};
export const OPPER_MODEL_MAX_OUTPUT_TOKENS: Record<string, number> = {
"claude-sonnet-4-6": 64_000,
"claude-opus-5-5": 128_000,
"claude-opus-5": 128_000,
"gpt-5.5": 128_000,
"gpt-5.4-mini": 128_000,
"gemini-3.8-flash": 65_536,
"deepseek-v4-pro": 65_536,
"kimi-k3": 131_072,
"mistral-large-2512": 8_192,
};
// Pools whose members do not all accept image input (deepseek-v4-pro: no member does; kimi-k3: the
// sference route is text-only), so the shared modality set is text.
export const OPPER_TEXT_ONLY_MODELS = ["deepseek-v4-pro", "kimi-k3"];
export const OPPER_MODEL_INPUT_MODALITIES: Record<string, string[]> = Object.fromEntries(
OPPER_MODELS.map(id => [id, OPPER_TEXT_ONLY_MODELS.includes(id) ? ["text"] : ["text", "image"]]),
);
export const STEPFUN_MODELS = [
"step-5-preview",
"step-3.5-flash",
"step-3.7-flash",
];
export const STEPFUN_MODEL_CONTEXT_WINDOWS: Record<string, number> = {
"step-5-preview": 1_000_000,
"step-3.5-flash": 256_000,
"step-3.7-flash": 256_000,
};
export const STEPFUN_MODEL_INPUT_MODALITIES: Record<string, string[]> = {
"step-5-preview": ["text", "image"],
"step-3.5-flash": ["text"],
"step-3.7-flash": ["text", "image"],
};
export const STEPFUN_NO_VISION_MODELS = ["step-3.5-flash"];
export const STEPFUN_REASONING_EFFORTS = ["low", "medium", "high"];