harn-vm 0.7.18

# Harn provider capability matrix.
#
# One `[[provider.<name>]]` array entry per rule; first match wins per
# (provider, model). Place more specific `model_match` patterns before
# wildcards. `version_min = [major, minor]` narrows the match to a model
# ID whose `(major, minor)` version (parsed from the Anthropic / OpenAI
# naming schemes) is greater than or equal to the given tuple. Rules
# whose `version_min` is unparseable for the given model are skipped.
#
# `[provider_family]` declares the sibling providers that inherit rules
# from a canonical family when they have no rule of their own (OpenRouter
# et al. speak the same Responses API and forward `tool_search` /
# `defer_loading` unchanged — they fall through to `[[provider.openai]]`
# by default).
#
# Users override or extend this table per-project via
# `[[capabilities.provider.<name>]]` entries in `harn.toml`. Project
# overrides are checked before the built-in rules for the same provider
# name and are authoritative on overlap.
#
# Supported per-rule fields:
#   model_match     : glob pattern matched against the lowercased model ID.
#   version_min     : [major, minor] lower bound, provider-aware parse.
#   native_tools    : whether the model accepts native tool-call wire shape.
#   defer_loading   : whether `defer_loading: true` is honored on tool defs.
#   tool_search     : list of native tool-search variants, preferred first.
#                     Anthropic = ["bm25", "regex"];
#                     OpenAI    = ["hosted", "client"].
#   max_tools       : cap on tool-definition count the provider will accept.
#                     Used by harn-lint to warn about oversized registries.
#   prompt_caching  : whether the provider honors cache_control blocks.
#   thinking        : whether extended / adaptive thinking is available.

# ---------- Anthropic (Claude) ------------------------------------------------

# Haiku 4.5+ supports server-side tool search.
[[provider.anthropic]]
model_match = "claude-haiku-*"
version_min = [4, 5]
native_tools = true
defer_loading = true
tool_search = ["bm25", "regex"]
max_tools = 10000
prompt_caching = true
thinking = true

# Opus 4.0+ supports tool search.
[[provider.anthropic]]
model_match = "claude-opus-*"
version_min = [4, 0]
native_tools = true
defer_loading = true
tool_search = ["bm25", "regex"]
max_tools = 10000
prompt_caching = true
thinking = true

# Sonnet 4.0+ supports tool search.
[[provider.anthropic]]
model_match = "claude-sonnet-*"
version_min = [4, 0]
native_tools = true
defer_loading = true
tool_search = ["bm25", "regex"]
max_tools = 10000
prompt_caching = true
thinking = true

# OpenRouter-style `anthropic/claude-...` prefixes.
[[provider.anthropic]]
model_match = "anthropic/claude-haiku-*"
version_min = [4, 5]
native_tools = true
defer_loading = true
tool_search = ["bm25", "regex"]
max_tools = 10000
prompt_caching = true
thinking = true

[[provider.anthropic]]
model_match = "anthropic/claude-opus-*"
version_min = [4, 0]
native_tools = true
defer_loading = true
tool_search = ["bm25", "regex"]
max_tools = 10000
prompt_caching = true
thinking = true

[[provider.anthropic]]
model_match = "anthropic/claude-sonnet-*"
version_min = [4, 0]
native_tools = true
defer_loading = true
tool_search = ["bm25", "regex"]
max_tools = 10000
prompt_caching = true
thinking = true

# Catch-all for older Claude models — native tools + prompt caching +
# thinking, but no progressive tool disclosure.
[[provider.anthropic]]
model_match = "claude-*"
native_tools = true
prompt_caching = true
thinking = true

# ---------- OpenAI family -----------------------------------------------------
#
# `provider.openai` rules are inherited by the sibling providers declared
# in `[provider_family]` below (OpenRouter, Together, Groq, DeepSeek,
# Fireworks, HuggingFace, local vLLM/SGLang). Siblings may still add their
# own `[[provider.<name>]]` rules and those win over the openai fallback.

# gpt-5.4+ exposes native `tool_search` on the Responses API.
[[provider.openai]]
model_match = "gpt-*"
version_min = [5, 4]
native_tools = true
defer_loading = true
tool_search = ["hosted", "client"]

# Legacy GPT: native tool calls only.
[[provider.openai]]
model_match = "gpt-*"
native_tools = true

# Reasoning family (o1, o3, o4, ...).
[[provider.openai]]
model_match = "o1*"
native_tools = true

[[provider.openai]]
model_match = "o3*"
native_tools = true

[[provider.openai]]
model_match = "o4*"
native_tools = true

# OpenRouter-style provider-prefixed IDs.
[[provider.openai]]
model_match = "openai/gpt-*"
version_min = [5, 4]
native_tools = true
defer_loading = true
tool_search = ["hosted", "client"]

[[provider.openai]]
model_match = "openai/gpt-*"
native_tools = true

# ---------- Local / Ollama ----------------------------------------------------
#
# Local providers don't advertise native tool_search or prompt caching.
# Native-tools coverage depends on the specific model; leave conservative
# (unset) at the default so users must opt in via a harn.toml override.

# ---------- Mock --------------------------------------------------------------
#
# Mock spoofs either Anthropic or OpenAI shape depending on the model ID.
# Handled specially in the loader (see `capabilities::lookup`): Claude-
# shape model strings route to the `anthropic` rule list first, otherwise
# fall through to the `openai` rule list.

# ---------- provider_family aliases ------------------------------------------

[provider_family]
openrouter = "openai"
together = "openai"
groq = "openai"
deepseek = "openai"
fireworks = "openai"
huggingface = "openai"
local = "openai"