harn-vm 0.10.102

# ---------- DeepInfra — open-weight OpenAI-compatible host -------------------
# DeepInfra serves open weights (DeepSeek, Qwen, Llama, Kimi, GPT-OSS) on a
# standard OpenAI chat-completions surface with native tool calls. Reasoning
# families expose an inline thinking trace; DeepSeek routes honor prompt
# caching. Catalog keys are `deepinfra/<hf-id>`, so patterns match the
# family substring.

[[provider.deepinfra]]
model_match = "*deepseek*"
native_tools = true
preferred_tool_format = "native"
structured_output = "native"
thinking_modes = ["enabled"]
prompt_caching = true
text_tool_wire_format_supported = true
prefers_xml_scaffolding = false
prefers_markdown_scaffolding = true
structured_output_mode = "native_json"
supports_assistant_prefill = false
prefers_role_developer = false
prefers_xml_tools = false
thinking_block_style = "inline"

# DeepInfra's GLM-5.2 deployment specifically: pinned to the fenced-JSON text
# channel because its NATIVE channel duplicates tool calls. 2026-08-15 probe
# against `zai-org/GLM-5.2` with tool_choice=required returned 38 tool calls for
# a single-city prompt — 38 distinct `index`/`id` entries all carrying the same
# `{"city": "Paris"}` — deterministic across 4/4 runs, in both sync and
# streaming mode. tool_choice=auto returned a single clean call.
#
# This is a HOST+MODEL defect, not the GLM family's: on the same DeepInfra
# account, GLM-5.1, GLM-4.7, DeepSeek-V3.2, DeepSeek-V4-Pro and
# DeepSeek-V4-Pro-0813 all returned exactly one call under required, and
# GLM-5.2 served by zai-direct, OpenRouter, Fireworks, NVIDIA and Together
# returned one call too. So the pin lives here rather than in the retired
# cross-host `glm-5` family row (see capability_audit.rs).
#
# A duplicate storm is worse than a leak: each phantom call is separately
# dispatched, so an agent loop would run the same tool 38 times.
# Re-probe before widening this back to native.
[[provider.deepinfra]]
model_match = "*glm-5.2*"
native_tools = true
preferred_tool_format = "json"
tool_mode_parity = "native_unreliable"
tool_mode_parity_notes = "2026-08-15 live probe: native channel emits 38 duplicate tool_calls for one intent under tool_choice=required (deterministic 4/4, sync + streaming); tool_choice=auto is clean. Host-specific to DeepInfra's GLM-5.2 deployment — GLM-5.1/GLM-4.7/DeepSeek on the same host and GLM-5.2 on five other hosts each return a single call. Fenced JSON text-channel tools dispatch correctly."
structured_output = "native"
thinking_modes = ["enabled"]
prompt_caching = true
text_tool_wire_format_supported = true
prefers_xml_scaffolding = false
prefers_markdown_scaffolding = true
structured_output_mode = "native_json"
supports_assistant_prefill = false
prefers_role_developer = false
prefers_xml_tools = false
thinking_block_style = "inline"

# Remaining DeepInfra GLM-5 routes (GLM-5.1, GLM-5). 2026-08-15 probe: GLM-5.1
# returned a single clean `message.tool_calls` under tool_choice=required, with
# no markup leak and no duplication. GLM-4.x is not matched by this pattern and
# keeps whatever it resolved to before — DeepInfra GLM-4.7 also probed clean,
# but adding a `*glm-4*` row here would change a route this change did not set
# out to touch.
[[provider.deepinfra]]
model_match = "*glm-5*"
native_tools = true
preferred_tool_format = "native"
tool_mode_parity_notes = "2026-08-15 cross-host native re-probe: DeepInfra GLM-5.1 returned a single clean `message.tool_calls` under tool_choice=required. The GLM-5.2 duplication is handled by the more specific rule above."
structured_output = "native"
thinking_modes = ["enabled"]
prompt_caching = true
text_tool_wire_format_supported = true
prefers_xml_scaffolding = false
prefers_markdown_scaffolding = true
structured_output_mode = "native_json"
supports_assistant_prefill = false
prefers_role_developer = false
prefers_xml_tools = false
thinking_block_style = "inline"

[[provider.deepinfra]]
model_match = "*qwen3.7*"
native_tools = true
preferred_tool_format = "native"
structured_output = "native"
thinking_modes = ["enabled"]
prompt_caching = true
text_tool_wire_format_supported = true
prefers_xml_scaffolding = false
prefers_markdown_scaffolding = true
structured_output_mode = "native_json"
supports_assistant_prefill = false
prefers_role_developer = false
prefers_xml_tools = false
thinking_block_style = "inline"

# probed 2026-06-24 (docs/eval/provider-tool-mode-sweep-2026-06-24.md, N=5,
# forced-format single-tool authoring of a backslash-heavy Zig body): native
# 1/5 fidelity (bills empty completions), json 2/5 (flaky parse), text 5/5
# byte-clean. The provider-native channel cannot carry backslash-heavy code, so
# steer to the escape-free heredoc text channel.
[[provider.deepinfra]]
model_match = "*qwen3.6*"
native_tools = true
preferred_tool_format = "text"
tool_mode_parity = "native_unreliable"
tool_mode_parity_notes = "2026-06-24 forced-format sweep (N=5): DeepInfra Qwen3.6-35B-A3B native bills empty completions (1/5) and fenced-JSON is flaky (2/5); heredoc text carried a backslash-heavy Zig body byte-clean 5/5."
structured_output = "native"
thinking_modes = ["enabled"]
vision = true
vision_supported = true
text_tool_wire_format_supported = true
prefers_xml_scaffolding = false
prefers_markdown_scaffolding = true
structured_output_mode = "native_json"
supports_assistant_prefill = false
prefers_role_developer = false
prefers_xml_tools = false
thinking_block_style = "inline"

[[provider.deepinfra]]
model_match = "*kimi-*"
native_tools = true
preferred_tool_format = "native"
structured_output = "native"
thinking_modes = ["enabled"]
prompt_caching = true
vision = true
vision_supported = true
video_supported = true
text_tool_wire_format_supported = true
prefers_xml_scaffolding = false
prefers_markdown_scaffolding = true
structured_output_mode = "native_json"
supports_assistant_prefill = false
prefers_role_developer = false
prefers_xml_tools = false
thinking_block_style = "inline"

# DeepInfra-hosted GPT-OSS is the same Harmony model as cerebras/fireworks
# gpt-oss-120b. Placed BEFORE the catch-all `*` rule (first match wins) so it
# does not fall through to the non-reasoning default — without it
# `reasoning_required_for_tools` resolves OFF and the eval loop bills a
# noncommittal because gpt-oss calls tools INSIDE the chain-of-thought channel.
# Reasoning-effort thinking {low, medium, high}, `reasoning_required_for_tools =
# true`. Must NOT carry a Qwen-style `auto_reasoning_overrides = "off"`.
#
# TOOL CHANNEL — PINNED TO TEXT (matches the Fireworks #3505 gpt-oss precedent).
# DeepInfra's gpt-oss native channel is a Harmony footgun: a 2026-06-24 Harn
# agent-loop run (gpt-oss-120b, zig-feat, tool grounding present) saw DeepInfra
# native bill a non-empty completion (completion_tokens=86) with NO dispatchable
# tool call and NO answer — the action was serialized only in the private
# reasoning/commentary channel of the Harmony format, never emitted as a
# provider-native `tool_calls` entry. Repeated ~10x -> the run was unusable. This
# is the same class of defect reported across the open-source engines DeepInfra
# runs: vLLM #22578 (chat-completions tool_calls empty / missing arguments),
# vLLM #44216 (tool_choice="required" ignored, empty tool_calls + finish_reason
# stop), SGLang #8976 (commentary-channel crash) and #10738 (tool parser not
# working), openai/harmony #68 (tool intent stuck in commentary, never
# dispatched), and HuggingFace openai/gpt-oss-20b discussion #80 (function in
# reasoning_content with no tool_call structure). The clean+reliable channel on
# pay-per-token gpt-oss routes is TEXT (heredoc). `tool_mode_parity =
# "native_unreliable"` makes a `native` pin auto-correct to `text` via
# `validate_tool_format` with an explanatory `correction`, so no alias pin or
# `--tool-format native` can silently re-introduce the empty tool stream.
[[provider.deepinfra]]
model_match = "*gpt-oss*"
native_tools = false
preferred_tool_format = "text"
tool_mode_parity = "native_unreliable"
tool_mode_parity_notes = "2026-06-24 Harn agent-loop (gpt-oss-120b, zig-feat, tool grounding present): DeepInfra native billed completion_tokens=86 with no dispatchable tool call or answer (Harmony reasoning-channel-only / upstream contract violation), repeated ~10x -> run unusable. Text/heredoc is the clean pay-per-token channel. See vLLM #22578/#44216, SGLang #8976/#10738, openai/harmony #68."
structured_output = "native"
thinking_modes = ["effort"]
reasoning_effort_supported = true
reasoning_effort_levels = ["low", "medium", "high"]
reasoning_required_for_tools = true
text_tool_wire_format_supported = true
prefers_xml_scaffolding = false
prefers_markdown_scaffolding = true
structured_output_mode = "native_json"
supports_assistant_prefill = false
prefers_role_developer = false
prefers_xml_tools = false
thinking_block_style = "reasoning_summary"

[[provider.deepinfra]]
model_match = "*"
native_tools = true
preferred_tool_format = "native"
structured_output = "native"
text_tool_wire_format_supported = true
prefers_xml_scaffolding = false
prefers_markdown_scaffolding = true
structured_output_mode = "native_json"
supports_assistant_prefill = false
prefers_role_developer = false
prefers_xml_tools = false
thinking_block_style = "none"