harn-vm 0.10.28

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

# ---------- Moonshot AI — direct Kimi K2 (OpenAI-compatible /v1) -------------
# Moonshot's first-party API serves the Kimi K2 family with native tool calls,
# an inline thinking mode, prompt caching, and multimodal input. K2.7 Code
# additionally accepts video. Catalog keys are `moonshot/<wire>` (e.g.
# `moonshot/kimi-k2.6`), so the patterns match the `*kimi*` substring to cover
# both the prefixed catalog key and the bare wire id.
#
# TOOL-CHANNEL CAVEAT (credentialed probe 2026-07-18, harn 0.10.23 binary,
# direct Moonshot /v1, N=2 per case). The whole Kimi family reasons, and that
# interacts with tool calling two ways:
#   1. FORCED tool_choice (`required` or a named function) returns HTTP 400
#      "tool_choice '...' is incompatible with thinking enabled". This is the
#      likely cause of the "empty completion, end_turn, 0 tokens, $0" agent-loop
#      drops reported on Kimi (a 400 bills nothing) — NOT a native-emission
#      drop. Keep `allowed_tool_choice_modes` to `["auto", "none"]` for every
#      Kimi route; never advertise `required`.
#   2. With `tool_choice: auto` the native channel is reliable — k2.6 and k3
#      both carried a backslash/quote/unicode large-string argument byte-exact
#      2/2. The TEXT channel is model-dependent: k3 parsed byte-exact, but k2.6
#      burned its whole budget in the reasoning channel (finish_reason=length,
#      empty content) and never emitted the `<tool_call>` — so do NOT flip the
#      general (k2.5/k2.6) Kimi rule to json/text; native+auto is its reliable
#      path.

[[provider.moonshot]]
model_match = "*kimi-k3*"
# K3 credentialed probe (2026-07-18, harn 0.10.23 binary, direct Moonshot
# /v1, N=2): with `tool_choice: auto` the native channel carried the
# echo_marker call's backslash/quote/unicode large-string argument byte-exact
# (2/2), and the Harn text-tool channel emitted a parseable `<tool_call>`
# block byte-exact (2/2) with no reasoning-channel over-run — native and text
# are interchangeable here. FORCED tool_choice is the footgun: K3 always
# reasons (`reasoning_effort: max`), and Moonshot rejects `tool_choice:
# "required"` / a named function with HTTP 400 "tool_choice '...' is
# incompatible with thinking enabled". So `required` is dropped from
# allowed_tool_choice_modes below; only `auto`/`none` are real modes.
native_tools = true
preferred_tool_format = "native"
tool_mode_parity = "interchangeable"
tool_mode_parity_notes = "2026-07-18 credentialed probe (harn 0.10.23, direct Moonshot /v1, tool_choice=auto, N=2): native carried the large backslash/quote/unicode string argument byte-exact 2/2, and the Harn text-tool channel parsed byte-exact 2/2 with no reasoning over-run. Forced tool_choice (required/specified) returns HTTP 400 'incompatible with thinking enabled' and is not an available mode."
structured_output = "native"
# K3 always reasons through the OpenAI-compatible `reasoning_effort` field;
# it accepts only the `max` effort below, rather than a generic enabled mode.
thinking_modes = ["effort"]
vision = true
vision_supported = true
prompt_caching = true
# Forced tool_choice (`required` / a named function) returns HTTP 400
# "incompatible with thinking enabled" on K3's mandatory reasoning (measured
# 2026-07-18); only `auto`/`none` are real modes.
allowed_tool_choice_modes = ["auto", "none"]
requires_tool_result_adjacency = true
# K3 always reasons, accepts only `reasoning_effort: max`, requires
# `max_completion_tokens`, and rejects fixed sampling/penalty knobs.
requires_completion_tokens = true
reasoning_effort_supported = true
reasoning_effort_levels = ["max"]
reasoning_none_supported = false
reasoning_disable_supported = false
reasoning_history_wire_field = "reasoning_content"
temperature_supported = false
top_p_supported = false
frequency_penalty_supported = false
presence_penalty_supported = false
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.moonshot]]
model_match = "*kimi-k2.7-code*"
native_tools = true
preferred_tool_format = "json"
tool_mode_parity = "native_unreliable"
tool_mode_parity_notes = "2026-06-20 Harn agent-loop smoke after parser fix: forced native/off emitted no dispatchable tool_calls and claimed the tool was unavailable. Harn JSON tools completed the loop; text tools also pass after fixing text-mode history projection."
structured_output = "native"
thinking_modes = ["enabled"]
vision = true
vision_supported = true
video_supported = true
prompt_caching = true
allowed_tool_choice_modes = ["auto", "none"]
requires_tool_result_adjacency = true
temperature_supported = false
top_p_supported = false
frequency_penalty_supported = false
presence_penalty_supported = false
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.moonshot]]
model_match = "*kimi*"
native_tools = true
preferred_tool_format = "native"
structured_output = "native"
thinking_modes = ["enabled"]
vision = true
vision_supported = true
prompt_caching = true
requires_tool_result_adjacency = true
allowed_tool_choice_modes = ["auto", "none"]
# Every Kimi model pins these sampling fields to a single server-side value and
# rejects any other with HTTP 400 ("invalid temperature: only 1 is allowed for
# this model" / "invalid top_p: only 0.95 is allowed"); confirmed by live probe
# on kimi-k2.5, k2.6, k2.7-code, and k2.7-code-highspeed (2026-07-14). Gating
# them here strips the fields so the server default (the sole allowed value)
# applies. The more specific k2.7-code rule sets these too; rules are
# first-match, so this general rule is what covers k2.5/k2.6. The non-kimi
# moonshot-v1-* models accept both and are correctly not matched here.
temperature_supported = false
top_p_supported = false
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"