# Provider Tool-Calling Verification — 2026-07-30

Phase 7.1 live sweep (ZyquoAgent --verify): every agent-capable model receives the list_files tool schema, must emit a valid tool call (streamed), consume the threaded tool_result, and produce a final answer naming the listed files.

80 models tested · 77 green · 3 failed (all 3 failures are provider-side — see the classification at the bottom)

This table merges the full pass-1 sweep with the per-provider re-runs performed after the two client fixes below (each fixed model was re-verified green live):

OpenAI GPT-5.4/5.5/5.6: /chat/completions rejects function tools combined with any reasoning_effort other than "none" — the client now downgrades reasoning_effort to "none" whenever tools are attached for these models (OpenAICompatibleClient.buildBody).
Gemini 3.x (compat endpoint): tool calls carry an extra_content.google.thought_signature that MUST be echoed back verbatim when the call is threaded into history, or Gemini rejects the request (INVALID_ARGUMENT: Function call is missing a thought_signature). ToolCall now carries the opaque signature; the client captures it from streaming fragments / non-streaming entries and re-attaches it on assistant tool_calls.

Provider	Model	Tool call	Threading	Streaming	Latency	Notes
OpenAI	`gpt-5.6-sol`	✅	✅	✅	3.0s	pass 1 ❌ (tools+reasoning_effort rejected); green after client fix
OpenAI	`gpt-5.6-terra`	✅	✅	✅	1.4s	pass 1 ❌ (tools+reasoning_effort rejected); green after client fix
OpenAI	`gpt-5.6-luna`	✅	✅	✅	1.6s	pass 1 ❌ (tools+reasoning_effort rejected); green after client fix
OpenAI	`gpt-5.5`	✅	✅	✅	4.1s	pass 1 ❌ (tools+reasoning_effort rejected); green after client fix
OpenAI	`gpt-5.4`	✅	✅	✅	1.8s	pass 1 ❌ (tools+reasoning_effort rejected); green after client fix
OpenAI	`gpt-5.4-mini`	✅	✅	✅	0.9s	pass 1 ❌ (tools+reasoning_effort rejected); green after client fix
OpenAI	`gpt-5.2`	✅	✅	✅	3.3s
OpenAI	`gpt-5.1`	✅	✅	✅	1.6s
OpenAI	`gpt-5`	✅	✅	✅	2.4s
OpenAI	`gpt-5-mini`	✅	✅	✅	2.1s
OpenAI	`o3`	✅	✅	✅	1.3s
OpenAI	`o4-mini`	✅	✅	✅	1.2s
Anthropic	`claude-opus-5`	✅	✅	✅	3.9s
Anthropic	`claude-sonnet-5`	✅	✅	✅	4.0s
Anthropic	`claude-fable-5`	✅	✅	✅	7.5s
Anthropic	`claude-opus-4-8`	✅	✅	✅	3.2s
Anthropic	`claude-opus-4-7`	✅	✅	✅	3.2s
Anthropic	`claude-opus-4-6`	✅	✅	✅	6.1s
Anthropic	`claude-sonnet-4-6`	✅	✅	✅	4.1s
Anthropic	`claude-haiku-4-5-20251001`	✅	✅	✅	1.6s
xAI	`grok-4.5`	✅	✅	✅	2.5s
xAI	`grok-4.3`	✅	✅	✅	4.5s
xAI	`grok-4.20`	✅	✅	✅	4.9s
xAI	`grok-4.20-non-reasoning`	✅	✅	✅	1.2s
xAI	`grok-code-fast-1`	✅	✅	✅	5.3s
Mistral	`mistral-medium-latest`	✅	✅	✅	1.4s
Mistral	`mistral-large-latest`	✅	✅	✅	3.7s
Mistral	`mistral-small-latest`	✅	✅	✅	0.9s
Google Gemini	`gemini-3.6-flash`	✅	✅	✅	2.1s	pass 1 ❌ (thought_signature missing on threading); green after client fix
Google Gemini	`gemini-3.5-flash`	✅	✅	✅	2.0s	pass 1 ❌ (thought_signature); green after client fix
Google Gemini	`gemini-3.5-flash-lite`	✅	✅	✅	1.0s	pass 1 ❌ (thought_signature); green after client fix
Google Gemini	`gemini-3.1-pro-preview`	✅	✅	✅	3.9s	pass 1 ❌ (thought_signature); green after client fix
Google Gemini	`gemini-2.5-pro`	✅	✅	✅	3.3s
Google Gemini	`gemini-2.5-flash`	✅	✅	✅	1.3s
Google Gemini	`gemini-pro-latest`	✅	✅	✅	4.2s	alias of Gemini 3.x; pass 1 ❌ (thought_signature); green after client fix
Google Gemini	`gemini-flash-latest`	✅	✅	✅	1.5s	alias of Gemini 3.x; pass 1 ❌ (thought_signature); green after client fix
Alibaba Qwen	`qwen3.7-max`	✅	✅	✅	3.0s
Alibaba Qwen	`qwen3.7-plus`	✅	✅	✅	2.6s
Alibaba Qwen	`qwen3.7-flash`	✅	✅	✅	1.6s
Alibaba Qwen	`qwen3.6-plus`	✅	✅	✅	2.2s
Alibaba Qwen	`qwen3.5-plus`	✅	✅	✅	2.0s
Alibaba Qwen	`qwen3-coder-plus`	✅	✅	✅	2.1s
Alibaba Qwen	`qwen3-coder-flash`	✅	✅	✅	2.0s
Alibaba Qwen	`qwen3-coder-next`	✅	✅	✅	2.3s
Alibaba Qwen	`qwen3-coder-480b-a35b-instruct`	✅	✅	✅	2.6s
Alibaba Qwen	`qwen3.5-397b-a17b`	✅	✅	✅	1.9s
Alibaba Qwen	`deepseek-v4-pro`	✅	✅	✅	4.6s
Alibaba Qwen	`glm-5.2`	✅	✅	✅	2.1s
DeepSeek	`deepseek-v4-flash`	✅	✅	✅	2.5s
DeepSeek	`deepseek-v4-pro`	✅	✅	✅	3.2s
Kimi	`kimi-k3`	❌	❌	❌	100.5s	INFRA: HTTP 429 `engine_overloaded_error` on most requests during the sweep window (reproduced with raw curl in every request shape; intermittent 200s prove the key/tier is fine). One harness attempt DID emit a correct `list_files` call, so the wire format is right — Moonshot capacity issue; re-verify when the engine recovers
Kimi	`kimi-k2.7-code`	✅	✅	✅	2.3s
Kimi	`kimi-k2.7-code-highspeed`	✅	✅	✅	1.8s
Kimi	`kimi-k2.6`	✅	✅	✅	5.1s
Kimi	`kimi-k2.5`	✅	✅	✅	4.8s
Together AI	`moonshotai/Kimi-K3`	✅	✅	✅	19.0s	slow first token (reasoning)
Together AI	`moonshotai/Kimi-K2.7-Code`	✅	✅	✅	1.6s
Together AI	`deepseek-ai/DeepSeek-V4-Pro`	✅	✅	✅	1.6s
Together AI	`zai-org/GLM-5.2`	✅	✅	✅	1.2s
Together AI	`Qwen/Qwen3.7-Max`	✅	✅	✅	4.6s
Together AI	`openai/gpt-oss-120b`	✅	✅	✅	2.2s
Together AI	`nvidia/nemotron-3-ultra-550b-a55b`	✅	✅	✅	1.2s
Together AI	`MiniMaxAI/MiniMax-M3`	✅	✅	✅	6.1s
DeepInfra	`anthropic/claude-fable-5`	✅	✅	✅	6.3s
DeepInfra	`anthropic/claude-opus-5`	✅	✅	✅	5.6s
DeepInfra	`anthropic/claude-sonnet-5`	✅	✅	✅	7.1s
DeepInfra	`anthropic/claude-opus-4-8`	✅	✅	✅	4.1s
DeepInfra	`anthropic/claude-haiku-4-5`	✅	✅	✅	2.3s
DeepInfra	`google/gemini-3.1-pro`	✅	❌	❌	4.0s	SERVER-SIDE: step 1 works, but DeepInfra never returns Gemini's `thought_signature` and strips `extra_content` from requests (verified with raw probes, incl. Google's documented bypass placeholder) → Google rejects every threaded tool result with INVALID_ARGUMENT. Multi-turn tool use is impossible through DeepInfra — REMOVAL CANDIDATE (the same model is green via the native Gemini provider)
DeepInfra	`google/gemini-3.5-flash`	✅	❌	❌	2.3s	SERVER-SIDE: same as `google/gemini-3.1-pro` — thought_signature dropped by DeepInfra; REMOVAL CANDIDATE
DeepInfra	`deepseek-ai/DeepSeek-V4-Pro`	✅	✅	✅	4.1s
DeepInfra	`deepseek-ai/DeepSeek-V4-Flash`	✅	✅	✅	3.3s
DeepInfra	`moonshotai/Kimi-K2.7-Code`	✅	✅	✅	2.2s
DeepInfra	`zai-org/GLM-5.2`	✅	✅	✅	3.1s
DeepInfra	`Qwen/Qwen3.7-Max`	✅	✅	✅	4.7s
DeepInfra	`Qwen/Qwen3-Coder-480B-A35B-Instruct-Turbo`	✅	✅	✅	0.8s
DeepInfra	`openai/gpt-oss-120b`	✅	✅	✅	2.6s
DeepInfra	`MiniMaxAI/MiniMax-M3`	✅	✅	✅	3.0s
Cerebras	`gpt-oss-120b`	✅	✅	✅	1.4s
Cerebras	`gemma-4-31b`	✅	✅	✅	0.6s

# Failure classification

Model	Class	Evidence / disposition
`deepinfra` `google/gemini-3.1-pro`	Broken server-side	DeepInfra never emits `thought_signature` and drops `extra_content` on requests; Gemini 3 requires it for threaded tool calls. No client-side workaround exists (Google's documented placeholder is also stripped). Recommend removing from the agent-capable set — the identical models are green via the native Gemini provider.
`deepinfra` `google/gemini-3.5-flash`	Broken server-side	Same as above. Recommend removing from the agent-capable set.
`kimi` `kimi-k3`	Infra (provider capacity)	HTTP 429 `engine_overloaded_error` on most requests during the sweep window, reproduced with raw curl in every request shape; intermittent 200s prove the key/tier is fine and one harness attempt did emit a correct `list_files` call. Not a wire-format or capability failure — keep in the set, re-verify when Moonshot capacity recovers.

No model id in the agent-capable set was rejected as nonexistent/deprecated, and no model failed tool calling because of model capability — every ❌ above is provider-side.