Apply 5-reviewer audit fixes to per-provider capability buckets (PR #5711)
Five independent reviewers cross-checked every provider's per-model
sampling-knob exposure against live docs (OpenAI, Anthropic, Gemini,
DeepSeek, Kimi, Mistral, OpenRouter, llama.cpp, vLLM, Ollama).
Applying the high-confidence drift fixes here; speculative items (pro
model effort restrictions, gpt-5.3 cap, OpenAI verbosity / o-series
output cap, Gemini topK / service_tier) are deferred to a follow-up
because they need backend wire changes or unverified doc claims.
Anthropic:
- Move claude-opus-4-6 from the 64k group into the 128k group (live
legacy table shows Opus 4.6 Max output = 128k tokens).
https://platform.claude.com/docs/en/about-claude/models/overview
- Add claude-sonnet-4 to the 64k group (was falling through to 32k
default; live legacy table shows Sonnet 4 Max output = 64k tokens).
- Extend ANTHROPIC_REASONING_MODELS with legacy claude-opus-4-1 /
claude-opus-4 / claude-sonnet-4 at none/low/medium/high (live
legacy table marks Extended thinking = Yes for all three).
OpenAI:
- Split the gpt-5/gpt-5.1/gpt-5.2 reasoning bucket. Per Azure docs
footnote ^7^, "minimal is only supported with the original GPT-5
reasoning models. minimal is not supported with gpt-5.1 or greater".
gpt-5.1 / gpt-5.2 now get none/low/medium/high/xhigh with
supportsOff=true; bare gpt-5 keeps minimal/low/medium/high
supportsOff=false. Ordering puts gpt-5.1 / gpt-5.2 before gpt-5 in
the find() loop so the longer prefix matches first.
https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/reasoning
DeepSeek:
- Hide `seed` and `parallel_tool_calls` in the deepseek capability
bucket. Neither field is in the current /chat/completions schema
(body fields: messages, model, thinking, max_tokens, response_format,
stop, stream, stream_options, temperature, top_p, tools, tool_choice,
logprobs, top_logprobs, user_id). Surfacing them in the UI would be
the silent-drop UX the file header warns against.
https://api-docs.deepseek.com/api/create-chat-completion
Mistral:
- magistral-medium-latest / magistral-small-latest are NATIVE
always-on reasoning models; injecting reasoning_effort returns 422
upstream. Switch both to withEnableThinkingStyle({reasoningAlwaysOn:
true}) instead of the old none/medium/high effort ladder.
- mistral-small-latest / mistral-medium-latest / mistral-vibe-cli-latest
expose the documented three-tier adjustable ladder
(none/low/medium/high), not the truncated none/high pair that was
here before. mistral-medium-latest was not handled at all and now
sits in the same bucket as small.
https://docs.mistral.ai/studio-api/conversations/reasoning
https://mistral.ai/news/magistral
OpenRouter:
- Drop google/gemini-pro-latest from OPENROUTER_MANDATORY_REASONING_
MODELS; the gateway 404s the id today
(https://openrouter.ai/google/gemini-pro-latest). Removing rather
than re-pinning to a versioned id that may rotate again.
Local backends:
- Split LOCAL_LLAMA_CAPABILITIES into LLAMA_CPP_CAPABILITIES (full
chain — for llama_cpp + custom) and VLLM_OLLAMA_CAPABILITIES (OpenAI
subset + top_k/min_p/repetition_penalty/seed, no extended samplers).
vLLM's SamplingParams has no typical_p / top_n_sigma / repeat_last_n
/ dynatemp_* / mirostat* fields, and Ollama's OpenAI translator
(ollama/openai/openai.go FromChatRequest) only copies the OpenAI
subset. Surfacing the eight extra sliders for vllm / ollama was
silent-drop UX.
Tests:
- test_deepseek_payload_omits_seed_and_parallel_tool_calls: read the
TS file as text and assert the bucket has seed:false and
parallelToolCalls:false. Backend has no JS engine; this is the
cheapest way to lock the wire-drop invariant.
- 63/63 sampling_params_routing tests pass; frontend tsc clean.
This commit is contained in:
Daniel Han
parent
22111744a4
commit
0234bef047
2 changed files with 129 additions and 33 deletions
|
|
@ -1129,3 +1129,32 @@ def test_anthropic_4_7_sampling_removed_regex_matches_expected_ids():
|
|||
assert RX.match(mid), f"{mid!r} should match 4.7 sampling-removed regex"
|
||||
for mid in should_not_match:
|
||||
assert not RX.match(mid), f"{mid!r} should NOT match 4.7 regex"
|
||||
|
||||
|
||||
def test_deepseek_payload_omits_seed_and_parallel_tool_calls():
|
||||
"""DeepSeek's published /chat/completions schema lists
|
||||
messages/model/thinking/max_tokens/response_format/stop/stream/
|
||||
temperature/top_p/tools/tool_choice/logprobs/top_logprobs/user_id
|
||||
only. `seed` and `parallel_tool_calls` are not in the schema; the
|
||||
capability bucket hides them so the chat-adapter never sends them.
|
||||
Source:
|
||||
https://api-docs.deepseek.com/api/create-chat-completion
|
||||
"""
|
||||
# Frontend capability flags are the source of truth. Re-derive them
|
||||
# by reading the TS file as text (the backend has no JS engine) and
|
||||
# confirm the deepseek bucket has seed:false + parallelToolCalls:false.
|
||||
from pathlib import Path
|
||||
|
||||
src = (
|
||||
Path(__file__).resolve().parents[2]
|
||||
/ "frontend"
|
||||
/ "src"
|
||||
/ "features"
|
||||
/ "chat"
|
||||
/ "provider-capabilities.ts"
|
||||
).read_text(encoding = "utf-8")
|
||||
deepseek_idx = src.index(" deepseek: {")
|
||||
end = src.index("},", deepseek_idx)
|
||||
bucket = src[deepseek_idx:end]
|
||||
assert "seed: false" in bucket, bucket
|
||||
assert "parallelToolCalls: false" in bucket, bucket
|
||||
|
|
|
|||
|
|
@ -234,20 +234,25 @@ const EXTERNAL_MAX_OUTPUT_TOKENS_BY_MODEL: Array<{
|
|||
{ providerType: "openai", prefixes: ["gpt-5.5-pro", "gpt-5.5"], cap: 128000 },
|
||||
{ providerType: "openai", prefixes: ["gpt-5.4-pro", "gpt-5.4"], cap: 65536 },
|
||||
{ providerType: "openai", prefixes: ["gpt-5.3"], cap: 16384 },
|
||||
// Anthropic
|
||||
// Anthropic — overview table at
|
||||
// platform.claude.com/docs/en/about-claude/models/overview. Opus 4.7
|
||||
// and Opus 4.6 BOTH ship 128k Max output (the legacy-table row for
|
||||
// 4.6 reads "128k tokens"); Sonnet 4.6 / Sonnet 4.5 / Sonnet 4 / Opus
|
||||
// 4.5 / Haiku 4.5 ship 64k; Opus 4.1 / Opus 4 ship 32k (covered by
|
||||
// the 32k default below).
|
||||
{
|
||||
providerType: "anthropic",
|
||||
prefixes: ["claude-opus-4-7"],
|
||||
prefixes: ["claude-opus-4-7", "claude-opus-4-6"],
|
||||
cap: 128000,
|
||||
},
|
||||
{
|
||||
providerType: "anthropic",
|
||||
prefixes: [
|
||||
"claude-opus-4-6",
|
||||
"claude-sonnet-4-6",
|
||||
"claude-opus-4-5",
|
||||
"claude-sonnet-4-5",
|
||||
"claude-haiku-4-5",
|
||||
"claude-sonnet-4",
|
||||
],
|
||||
cap: 64000,
|
||||
},
|
||||
|
|
@ -548,11 +553,12 @@ const OPENAI_COMPAT_BASE: ProviderCapabilities = {
|
|||
topA: false,
|
||||
};
|
||||
|
||||
// Local llama.cpp-style backends (own llama-server, vLLM with extended
|
||||
// sampler support, Ollama). Exposes the full llama.cpp sampler chain
|
||||
// (typical_p / top_n_sigma / mirostat / dynatemp / repeat_last_n) but
|
||||
// not OpenRouter's gateway-specific top_a.
|
||||
const LOCAL_LLAMA_CAPABILITIES: ProviderCapabilities = {
|
||||
// Unsloth's first-party llama-server runtime (provider type `llama_cpp`)
|
||||
// plus the permissive `custom` preset. Exposes the full llama.cpp
|
||||
// sampler chain (typical_p / top_n_sigma / mirostat / dynatemp /
|
||||
// repeat_last_n) per the upstream server README. Not used for vLLM or
|
||||
// Ollama — see VLLM_OLLAMA_CAPABILITIES below.
|
||||
const LLAMA_CPP_CAPABILITIES: ProviderCapabilities = {
|
||||
temperature: true,
|
||||
topP: true,
|
||||
topK: true,
|
||||
|
|
@ -575,6 +581,26 @@ const LOCAL_LLAMA_CAPABILITIES: ProviderCapabilities = {
|
|||
topA: false,
|
||||
};
|
||||
|
||||
// vLLM and Ollama's OpenAI-compat endpoints accept the OpenAI subset
|
||||
// plus top_k / min_p / repetition_penalty / seed, but neither forwards
|
||||
// the llama.cpp-only extended samplers (typical_p, top_n_sigma,
|
||||
// repeat_last_n, dynatemp_*, mirostat*). vLLM's SamplingParams has no
|
||||
// fields for them (vllm/sampling_params.py) and Ollama's OAI
|
||||
// translator (ollama/openai/openai.go FromChatRequest) only copies the
|
||||
// OpenAI subset. Surfacing the eight extra sliders here would be the
|
||||
// silent-drop UX the file header warns against, so we hide them.
|
||||
const VLLM_OLLAMA_CAPABILITIES: ProviderCapabilities = {
|
||||
...LLAMA_CPP_CAPABILITIES,
|
||||
typicalP: false,
|
||||
topNSigma: false,
|
||||
repeatLastN: false,
|
||||
dynatempRange: false,
|
||||
dynatempExponent: false,
|
||||
mirostat: false,
|
||||
mirostatTau: false,
|
||||
mirostatEta: false,
|
||||
};
|
||||
|
||||
// OpenRouter is a router-of-routers: the gateway accepts a wider set
|
||||
// of OpenAI-style sampling fields than any single upstream supports
|
||||
// and silently drops what the chosen route does not, per
|
||||
|
|
@ -791,15 +817,18 @@ const PROVIDER_CAPABILITIES: Record<string, ProviderCapabilities> = {
|
|||
mirostatEta: false,
|
||||
topA: false,
|
||||
},
|
||||
// DeepSeek deprecated presence/frequency penalty in their current docs.
|
||||
// Chat-class defaults (deepseek-chat / deepseek-v4-flash non-thinking):
|
||||
// accept temperature, top_p, seed, stop. Reasoning class
|
||||
// (deepseek-reasoner / deepseek-v4-flash thinking-mode) ignores
|
||||
// temperature, top_p, presence_penalty, frequency_penalty entirely and
|
||||
// 400s on logprobs — see
|
||||
// DeepSeek deprecated presence/frequency penalty and never published
|
||||
// `seed` or `parallel_tool_calls` in the current chat-completion
|
||||
// schema — see https://api-docs.deepseek.com/api/create-chat-completion
|
||||
// (body fields: messages, model, thinking, max_tokens, response_format,
|
||||
// stop, stream, stream_options, temperature, top_p, tools, tool_choice,
|
||||
// logprobs, top_logprobs, user_id). Chat-class (deepseek-chat /
|
||||
// deepseek-v4-flash non-thinking) accepts temperature, top_p, stop;
|
||||
// reasoning class (deepseek-reasoner / deepseek-v4-flash thinking-mode)
|
||||
// additionally ignores temperature, top_p, presence_penalty,
|
||||
// frequency_penalty per
|
||||
// https://api-docs.deepseek.com/guides/reasoning_model. Per-model
|
||||
// resolution in getProviderCapabilities downshifts reasoner ids onto
|
||||
// DEEPSEEK_REASONING_CAPABILITIES.
|
||||
// resolution in getProviderCapabilities downshifts reasoner ids.
|
||||
deepseek: {
|
||||
temperature: true,
|
||||
topP: true,
|
||||
|
|
@ -808,10 +837,10 @@ const PROVIDER_CAPABILITIES: Record<string, ProviderCapabilities> = {
|
|||
repetitionPenalty: false,
|
||||
presencePenalty: false,
|
||||
frequencyPenalty: false,
|
||||
seed: true,
|
||||
seed: false,
|
||||
stop: true,
|
||||
serviceTier: false,
|
||||
parallelToolCalls: true,
|
||||
parallelToolCalls: false,
|
||||
typicalP: false,
|
||||
topNSigma: false,
|
||||
repeatLastN: false,
|
||||
|
|
@ -830,12 +859,15 @@ const PROVIDER_CAPABILITIES: Record<string, ProviderCapabilities> = {
|
|||
// OpenRouter API docs do not list them; they would be silently
|
||||
// dropped on most underlying models.
|
||||
openrouter: OPENROUTER_CAPABILITIES,
|
||||
// Local OpenAI-compatible connections terminate at llama-server-
|
||||
// style backends — full llama.cpp sampler chain available.
|
||||
custom: LOCAL_LLAMA_CAPABILITIES,
|
||||
vllm: LOCAL_LLAMA_CAPABILITIES,
|
||||
ollama: LOCAL_LLAMA_CAPABILITIES,
|
||||
llama_cpp: LOCAL_LLAMA_CAPABILITIES,
|
||||
// `llama_cpp` and the permissive `custom` preset terminate at the
|
||||
// first-party llama-server runtime, so the full sampler chain is
|
||||
// available. vLLM and Ollama only surface the OpenAI subset
|
||||
// (+ top_k/min_p/rep_penalty/seed) — the 8 extended llama.cpp
|
||||
// samplers are hidden to avoid the silent-drop UX.
|
||||
custom: LLAMA_CPP_CAPABILITIES,
|
||||
llama_cpp: LLAMA_CPP_CAPABILITIES,
|
||||
vllm: VLLM_OLLAMA_CAPABILITIES,
|
||||
ollama: VLLM_OLLAMA_CAPABILITIES,
|
||||
};
|
||||
|
||||
const DEFAULT_EXTERNAL_CAPABILITIES = OPENAI_COMPAT_BASE;
|
||||
|
|
@ -876,8 +908,11 @@ export function getProviderCapabilities(
|
|||
}
|
||||
|
||||
const DEFAULT_EFFORT_LEVELS = ["low", "medium", "high"] as const;
|
||||
// OpenRouter ids that have NO non-reasoning mode. `google/gemini-pro-latest`
|
||||
// used to live here but the gateway 404s the id today
|
||||
// (https://openrouter.ai/google/gemini-pro-latest); drop it rather than
|
||||
// re-pin to a versioned id that may rotate again.
|
||||
const OPENROUTER_MANDATORY_REASONING_MODELS = new Set([
|
||||
"google/gemini-pro-latest",
|
||||
"baidu/cobuddy:free",
|
||||
"inclusionai/ring-2.6-1t:free",
|
||||
"deepseek/deepseek-r1",
|
||||
|
|
@ -908,6 +943,9 @@ const NO_REASONING_CAPS: ReasoningCaps = {
|
|||
reasoningEffortLevels: DEFAULT_EFFORT_LEVELS,
|
||||
};
|
||||
|
||||
// Order matters: longest/most-specific prefixes first so the find() loop
|
||||
// in resolveAnthropicReasoningEffortCapabilities lands the right bucket
|
||||
// before the bare-family fallback ("claude-opus-4") sweeps an id.
|
||||
const ANTHROPIC_REASONING_MODELS = [
|
||||
{
|
||||
prefixes: ["claude-opus-4-7"],
|
||||
|
|
@ -922,6 +960,13 @@ const ANTHROPIC_REASONING_MODELS = [
|
|||
// Backend maps semantic levels to manual budget_tokens.
|
||||
levels: ["none", "low", "medium", "high"],
|
||||
},
|
||||
{
|
||||
// Legacy 4.x models. Live overview lists "Extended thinking = Yes"
|
||||
// for opus-4-1, sonnet-4, opus-4 (the latter two retire 2026-06-15
|
||||
// but the registry still surfaces them).
|
||||
prefixes: ["claude-opus-4-1", "claude-opus-4", "claude-sonnet-4"],
|
||||
levels: ["none", "low", "medium", "high"],
|
||||
},
|
||||
] as const;
|
||||
|
||||
function matchesModelPrefix(
|
||||
|
|
@ -968,7 +1013,16 @@ const OPENAI_REASONING_MODELS = [
|
|||
levels: ["none", "low", "medium", "high", "xhigh"],
|
||||
},
|
||||
{
|
||||
prefixes: ["gpt-5", "gpt-5.1", "gpt-5.2"],
|
||||
// Original gpt-5: minimal is supported, but per Azure footnote ^7^
|
||||
// "minimal is only supported with the original GPT-5 reasoning
|
||||
// models. minimal is not supported with gpt-5.1 or greater".
|
||||
// Listed before the gpt-5.1/5.2 entry so the longer match wins.
|
||||
prefixes: ["gpt-5.1", "gpt-5.2"],
|
||||
supportsOff: true,
|
||||
levels: ["none", "low", "medium", "high", "xhigh"],
|
||||
},
|
||||
{
|
||||
prefixes: ["gpt-5"],
|
||||
supportsOff: false,
|
||||
levels: ["minimal", "low", "medium", "high"],
|
||||
},
|
||||
|
|
@ -1036,19 +1090,32 @@ function resolveKimiReasoningCapabilities(modelId: string): ExternalReasoningCap
|
|||
}
|
||||
|
||||
function resolveMistralReasoningCapabilities(modelId: string): ExternalReasoningCapabilities {
|
||||
if (modelId === "magistral-medium-latest") {
|
||||
return withReasoningEffortStyle({
|
||||
// Native always-on reasoning family: magistral-* per
|
||||
// https://mistral.ai/news/magistral and
|
||||
// https://docs.mistral.ai/studio-api/conversations/reasoning .
|
||||
// "Always reasons; no parameter needed" — injecting reasoning_effort
|
||||
// returns 422 upstream. Treat like an OpenAI o-series always-on.
|
||||
if (
|
||||
modelId === "magistral-medium-latest" ||
|
||||
modelId === "magistral-small-latest"
|
||||
) {
|
||||
return withEnableThinkingStyle({
|
||||
supportsReasoning: true,
|
||||
supportsReasoningOff: false,
|
||||
// Native reasoning model: present baseline as Medium in the UI.
|
||||
reasoningEffortLevels: ["medium", "high"] as const,
|
||||
reasoningAlwaysOn: true,
|
||||
});
|
||||
}
|
||||
if (modelId === "mistral-small-latest" || modelId === "mistral-vibe-cli-latest") {
|
||||
// Adjustable reasoning family: three documented levels low/medium/high
|
||||
// plus the "none" off-switch (Mistral Studio conversations doc). The
|
||||
// earlier two-level ["none","high"] ladder was wrong.
|
||||
if (
|
||||
modelId === "mistral-small-latest" ||
|
||||
modelId === "mistral-medium-latest" ||
|
||||
modelId === "mistral-vibe-cli-latest"
|
||||
) {
|
||||
return withReasoningEffortStyle({
|
||||
supportsReasoning: true,
|
||||
supportsReasoningOff: true,
|
||||
reasoningEffortLevels: ["none", "high"] as const,
|
||||
reasoningEffortLevels: ["none", "low", "medium", "high"] as const,
|
||||
});
|
||||
}
|
||||
return withEnableThinkingStyle();
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue