harn-vm 0.10.121

# ---------- 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. A 2026-07-18 credentialed probe found that forced tool
# choice returned HTTP 400 while thinking was enabled. Moonshot added
# `required` support for K3 by 2026-08-26: direct controls with mandatory
# `reasoning_effort: max` returned one native call for both `auto` and
# `required`. Older Kimi rows retain the measured restriction.
#
# 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. A 2026-08-26 direct control confirmed that K3 now
# accepts `required` with mandatory `reasoning_effort: max` and returns a
# native tool call with `finish_reason: tool_calls`.
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. A 2026-08-26 direct control confirmed that K3 accepts tool_choice=required with reasoning_effort=max and returns a native tool call."
structured_output = "native"
# K3 always reasons through the OpenAI-compatible `reasoning_effort` field.
# Moonshot documents `low`, `high`, and `max`; `max` remains the route default.
thinking_modes = ["effort"]
vision = true
vision_supported = true
prompt_caching = true
# Directly measured with K3's mandatory reasoning on 2026-08-26.
allowed_tool_choice_modes = ["auto", "required", "none"]
requires_tool_result_adjacency = true
# K3 always reasons, requires `max_completion_tokens`, and rejects fixed
# sampling and penalty knobs.
requires_completion_tokens = true
reasoning_effort_supported = true
reasoning_effort_levels = ["low", "high", "max"]
reasoning_none_supported = false
reasoning_disable_supported = false
reasoning_history_wire_field = "reasoning_content"
reasoning_round_trip = "echo_same_key"
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"