Make reasoning summaries opt-in via a flag (with per-model override) for self-hosted reasoning models

Summary

For users running self-hosted reasoning models (DeepSeek R1, Qwen QwQ, Qwen-3 thinking, etc.) on Ollama / LM Studio / llama.cpp / vLLM, codex's bundled catalog has no entry for the slug, so supports_reasoning_summaries defaults to false and the reasoning stream never reaches the UI. The only existing escape hatch is the global model_supports_reasoning_summaries knob in ~/.codex/config.toml, which is all-or-nothing for users who run more than one local model.

The result is that running a real reasoning model locally feels strictly worse than running a non-reasoning model locally — the reasoning UI just doesn't show up, even though the model is emitting reasoning content.

Proposed solution

Two new provider-scoped fields on model_providers.<name>:

1. local_provider = true — marks a provider as serving locally-hosted models. When set, codex applies a conservative reasoning-name heuristic that auto-enables supports_reasoning_summaries for unknown models whose slug tokens match a known reasoning-model designator (r1, r2, r3, qwq, thinking, or reasoning, matched on alphanumeric segment boundaries — not substrings, so warp1 doesn't trip r1).

Built-in ollama and lmstudio providers default to local_provider = true. Every other provider (including OpenAI, Azure, Bedrock) defaults to false. User-defined providers opt in explicitly.

1. model_overrides.<slug>.supports_reasoning_summaries — explicit per-model escape hatch for cases the heuristic doesn't catch (e.g. qwen3-30b-a3b-thinking-2507 is fine, but myorg/some-reasoning-thing isn't), or to force-disable the heuristic for a specific model on a local provider. Bidirectional — both true and false are honored.

Example config.toml:

[model_providers.my-vllm]
name = "my vllm cluster"
base_url = "http://localhost:8000/v1"
wire_api = "responses"
local_provider = true                       # heuristic enabled

[model_providers.my-vllm.model_overrides."my-custom-reasoning-model"]
supports_reasoning_summaries = true         # explicit override

Precedence (lowest to highest)

1. Model catalog default (existing behavior preserved).
2. Heuristic auto-enable on local_provider = true for matching slugs.
3. Global model_supports_reasoning_summaries = true (only enables; Some(false) still ignored to preserve existing semantics).
4. Per-model model_overrides.<slug>.supports_reasoning_summaries (always wins).

Why a heuristic at all (and why this specific one)

A blanket auto-enable on local_provider = true would be over-reaching — strict OpenAI-compatible servers fronting non-reasoning models could 4xx on the unknown reasoning.summary request param. The heuristic limits the auto-enable to slugs that are unambiguously reasoning models. The allow-list is small and self-explanatory:

• r1 / r2 / r3 — DeepSeek R-series family
• qwq — Qwen QwQ family
• thinking — Qwen-3 reasoning variants
• reasoning — generic catch-all

Models that don't match still work fine, they just don't auto-enable. The model_overrides escape hatch covers them.

What I considered and rejected

• Auto-enable for every model on a local provider: too aggressive, would break strict servers running non-reasoning models.
• Per-model overrides only (no heuristic): solves correctness but doesn't actually fix the UX — users still have to write one override block per model.
• A new global model_supports_reasoning_summaries_per_provider = true knob: doesn't scale to the multi-model-per-provider case.

The combined heuristic + override design lets the 90% case work with one local_provider = true line, while keeping the 10% case correct via an explicit override.

deny_unknown_fields

ModelProviderOverride uses #[serde(deny_unknown_fields)] so typos like supports_reasoning_summary (missing trailing ies) are caught at config-load time instead of silently no-oping at runtime.

Reference implementation

A working implementation with tests lives on the team-wcv fork:

• Branch: https://github.com/team-wcv/codex/tree/feat/local-provider-reasoning-summaries
• Diff: https://github.com/team-wcv/codex/compare/main...feat/local-provider-reasoning-summaries

Touches:

• codex-rs/model-provider-info/src/lib.rs — new local_provider: bool and model_overrides: Option<HashMap<…, ModelProviderOverride>> fields. Built-in ollama/lmstudio default to local_provider = true; openai/bedrock to false.
• codex-rs/models-manager/src/config.rs — adds active_provider_is_local: bool and model_provider_overrides: HashMap<String, ActiveModelOverride> to ModelsManagerConfig.
• codex-rs/models-manager/src/model_info.rs — new looks_like_reasoning_model(slug) helper plus the four-layer precedence logic in with_config_overrides.
• codex-rs/core/src/config/mod.rs — to_models_manager_config resolves the active provider's overrides into the manager config.
• codex-rs/core/config.schema.json — regenerated via just write-config-schema.
• Tests: 9 new in model_info_tests.rs covering precedence, gating, exact-slug keying, and bidirectional override; 6 new in model_provider_info_tests.rs covering built-in defaults, TOML round-tripping, and deny_unknown_fields.

cargo test -p codex-models-manager -p codex-model-provider-info -p codex-config -p codex-core --lib passes ~1900 tests. cargo clippy --tests -- -D warnings and cargo fmt clean.

Open to alternatives

This is the shape that felt cleanest after weighing user UX against scope, but I'm open to alternatives — e.g. heuristic-only with no per-model override, override-only with no heuristic, or rejecting both and treating local reasoning models as something users should keep configuring manually. Happy to rework the patch shape if there's a direction you'd prefer.