PR: feat(config): per-model extra_body and thinking overrides

[TheGoddessInari]

Summary
Multi-model providers like NVIDIA NIM host models with incompatible reasoning wires under a single base URL. Today extra_body and thinking are provider-scoped: every model on the provider receives the same body parameters. This breaks providers where Model A needs chat_template_kwargs.thinking_mode while Model B needs chat_template_kwargs.enable_thinking and Model C needs neither.
Splitting a multi-model provider into N single-model entries works around the limitation but duplicates base URL, API key, and headers — configuration that is identical across models.
This adds two fields to ProviderModelOverride:

• ExtraBody map[string]any — per-model extra top-level JSON request body fields, merged into (and overriding) the provider-level extra_body for that model only.
• Thinking string — per-model thinking toggle ("enabled" / "disabled"), overriding the provider-level thinking field for that model only. The merge happens in applyModelOverride(), which already handles every other per-model override. Model-level keys win over provider-level keys; omitted keys inherit the provider value unchanged.
  Files changed
  File Change
  internal/config/config.go Add ExtraBody and Thinking fields to ProviderModelOverride; extend applyModelOverride() with merge + override logic internal/config/render.go Include new fields in renderModelOverride() and renderAnyMap() TOML output; update modelOverrideEmpty() internal/config/render_test.go Round-trip test with nested maps; legacy struct compatibility test; merge semantics test Config example (NVIDIA NIM)
  [[providers]]
  name = "nvidia"
  kind = "openai"
  base_url = "https://integrate.api.nvidia.com/v1"
  models = [
  "minimaxai/minimax-m3",
  "nvidia/nemotron-3-super-120b-a12b",
  "z-ai/glm-5.2",
  "google/gemma-4-31b-it",
  "deepseek-ai/deepseek-v4-flash",
  ]
  api_key_env = "NVIDIA_API_KEY"

[providers.model_overrides]
"minimaxai/minimax-m3" = {
context_window = 262000,
max_output_tokens = 16384,
extra_body = { chat_template_kwargs = { thinking_mode = "enabled" } },
}
"nvidia/nemotron-3-super-120b-a12b" = {
context_window = 1000000,
max_output_tokens = 32768,
extra_body = { chat_template_kwargs = { enable_thinking = true } },
}
"z-ai/glm-5.2" = {
context_window = 1000000,
max_output_tokens = 16384,
thinking = "enabled",
extra_body = { chat_template_kwargs = { enable_thinking = true, clear_thinking = false } },
}
Merge semantics
Model overrides are applied after the provider entry is fully constructed.
For extra_body: the provider-level map is shallow-cloned, then model-level
keys are written on top. A model override with extra_body = {} clears all
provider-level keys for that model. A model without extra_body in its override
inherits the provider value unchanged.
For thinking: a non-empty string replaces the provider-level value. An empty
or omitted string inherits the provider value.
Provider: thinking = "enabled", extra_body = { min_p = 0.01, stream = true }
Override: thinking = "disabled", extra_body = { chat_template_kwargs = { enable_thinking = true } }
Result: thinking = "disabled", extra_body = { min_p = 0.01, stream = true, chat_template_kwargs = { enable_thinking = true } }
What this does NOT do

• No new provider kind or wire protocol.
• No new vendor detection heuristics.
• No changes to buildRequest() — the merged extra_body and resolved
  thinking flow through the existing OpenAI-compatible request serialization.
• No impact on system prompt, tool schemas, or prefix construction.
  Backward compatibility
  BurntSushi/toml ignores unknown fields when decoding. Older Reasonix releases
  reading a config with the new fields will silently skip them — no parse error.
  The existing TestRenderTOMLModelOverrides legacy compatibility test
  (render_test.go:1127) already validates this pattern for context_window /
  max_output_tokens; the new fields follow the same convention.
  Cache-impact: low
  Only touches request body serialization for providers that use model overrides.
  The prefix (system prompt + tools) is unaffected. Guard: verify that a
  non-overridden model's marshalled request body is byte-identical before and
  after the change.
  Testing

1. go test ./internal/config/ -run TestRenderModelOverrideExtraBodyAndThinking —
   round-trip TOML encode/decode with nested maps, plus legacy struct
   compatibility
2. go test ./internal/config/ -run TestApplyModelOverrideMergesExtraBody —
   merge semantics: model wins over provider for conflicting keys, provider
   values inherit for missing keys, thinking override works
3. go test ./... — full suite regression

Issues

Verification

Documentation impact

Documentation-impact: TODO

For changes to user-visible CLI, Desktop, configuration, provider, permission,
or tool behavior, use one of:

Documentation-impact: updated - added model_overrides section with per-model extra_body/thinking docs to CONFIG_PATHS.md and GUIDE.md
:

Cache impact

Cache-impact: none — changes only apply when a specific model is selected via model_overrides; provider request serialization for non-overridden models is byte-identical before and after.
Cache-guard: existing TestRenderTOMLModelOverrides round-trip test + new TestRenderModelOverrideExtraBodyAndThinking cover the changed surface.
System-prompt-review: N/A

For cache-sensitive changes, fill these lines before requesting review:

• Cache-impact: none, low, medium, or high, plus the reason.
• Cache-guard: the focused guard test/command added or run, or why an existing guard covers the change.
• System-prompt-review: required reviewer/approval note when provider-visible system prompt, memory prefix, output style, or skill index behavior changes.