harn-vm 0.9.13

Async bytecode virtual machine for the Harn programming language
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155

# ---------- SambaNova Cloud — fast RDU OpenAI-compatible inference -----------
# SambaNova serves large open weights (DeepSeek, Llama 4, Qwen) at high
# throughput on an OpenAI chat-completions surface with native tool calls.
# DeepSeek reasoning routes expose an inline thinking trace; Llama 4 routes
# are non-thinking multimodal. Catalog keys are `sambanova/<wire>`.

[[provider.sambanova]]
model_match = "*deepseek-v3.2*"
native_tools = true
preferred_tool_format = "native"
tool_mode_parity = "interchangeable"
tool_mode_parity_notes = "2026-06-20 Harn agent-loop smoke: SambaNova DeepSeek-V3.2 completed both native and JSON tool loops with reasoning disabled. Earlier native false negatives were caused by Harn's parser treating terse final answers as billed no-ops; keep native preferred."
structured_output = "native"
thinking_modes = ["enabled"]
auto_reasoning_overrides = { agent = "off", verify = "off", code = "off" }
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.sambanova]]
model_match = "*deepseek*"
native_tools = true
preferred_tool_format = "native"
structured_output = "native"
thinking_modes = ["enabled"]
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.sambanova]]
model_match = "*llama*"
native_tools = true
preferred_tool_format = "native"
structured_output = "native"
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 = "none"

# 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
# 0/5 fidelity AND json 0/5 fidelity (both mangle backslashes — the `\\` Zig
# multiline prefix collapses to `\`), while heredoc text was 5/5 byte-clean. The
# escape-free text channel is the only reliable carrier for code on this route.
[[provider.sambanova]]
model_match = "*minimax*"
native_tools = true
preferred_tool_format = "text"
tool_mode_parity = "native_unreliable"
tool_mode_parity_notes = "2026-06-24 forced-format sweep (N=5): SambaNova MiniMax-M2.7 corrupted a backslash-heavy body on BOTH native (0/5) and fenced-JSON (0/5); heredoc text carried it byte-clean 5/5."
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.sambanova]]
model_match = "*gemma-4*"
native_tools = true
preferred_tool_format = "native"
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"

# SambaNova-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).
# SambaNova'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) ended with a
# provider/tool-protocol failure on the native channel — the same empty
# `tool_calls` / reasoning-channel-only class reported across the open-source
# engines (vLLM #22578, vLLM #44216, SGLang #8976/#10738, openai/harmony #68,
# HuggingFace openai/gpt-oss-20b discussion #80). 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 protocol failure.
# NOTE: this route is ALSO `serving_precision = "degraded"` (below) — even with
# the text-channel pin it must not back a meter baseline.
[[provider.sambanova]]
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): SambaNova native ended with a provider/tool-protocol failure (Harmony empty tool_calls / reasoning-channel-only class). 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"
# SambaNova serves gpt-oss at DEGRADED quality (quantized). PROVEN during the
# 2026-06 meter-convergence effort: c-feat 0/5 and cpp-storage 0/5 on SambaNova
# vs 3/3 and 1/1 on the OpenRouter full-precision reference. The numbers looked
# like a capability collapse but were provider contamination. The precision
# canary MUST refuse to ledger a meter baseline on this route.
serving_precision = "degraded"

[[provider.sambanova]]
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"