harn-vm 0.8.134

Async bytecode virtual machine for the Harn programming language
Documentation
# ── Alias tool-calling probe state ───────────────────────────────────────────
# Per-alias overrides recording the last-observed native vs. text vs.
# streaming tool-call probe outcome and the desired fallback. Hosts may
# update these via providers.toml overlays as they re-probe a model.

[alias_tool_calling.ollama-gemma4]
native = "unknown"
text = "unknown"
streaming_native = "unknown"
fallback_mode = "disabled"
failure_reason = "requires_tool_probe"

[alias_tool_calling.ollama-gemma4-12b]
native = "unknown"
text = "unknown"
streaming_native = "unknown"
fallback_mode = "disabled"
failure_reason = "requires_tool_probe"

[alias_tool_calling."llamacpp-qwen3.6-q4"]
native = "pass"
text = "unknown"
streaming_native = "pass"
fallback_mode = "native"
last_probe_at = "2026-06-05"

[alias_tool_calling."local-qwen3.6"]
native = "pass"
text = "unknown"
streaming_native = "pass"
fallback_mode = "native"
last_probe_at = "2026-06-05"

[alias_tool_calling."local-qwen3.6-gguf"]
native = "pass"
text = "unknown"
streaming_native = "pass"
fallback_mode = "native"
last_probe_at = "2026-06-05"

# MLX route now serves the Qwen3.6-35B-A3B MoE via mlx_lm.server (burin #2717).
# native is UNPROVEN on mlx_lm.server (the runtime profile requires a tool_probe
# before native is trusted; the validated `native` precedent is the llama.cpp
# sibling, a different server). Keep native = "unknown" so the readiness/probe
# path verifies served identity + a one-tool probe before selecting native.
[alias_tool_calling."mlx-qwen3.6"]
native = "unknown"
text = "unknown"
streaming_native = "unknown"
fallback_mode = "native"
failure_reason = "requires_served_identity_and_tool_probe"

[alias_tool_calling."mlx-qwen3.6-27b"]
native = "unknown"
text = "unknown"
streaming_native = "unknown"
fallback_mode = "native"
failure_reason = "requires_served_identity_and_tool_probe"