fix(llm): invert reasoning default — unknown models skip think/final tags

[henrypark133]

Summary

• Invert reasoning default: New requires_think_final_tags() with empty allowlist — unknown/alias models (including NEAR AI auto) get the safe direct-answer prompt instead of <think>/<final> injection
• Fix field name mismatch: Add #[serde(alias = "reasoning")] so vLLM/SGLang's reasoning field is accepted (was only reading reasoning_content)
• Fix direct-answer prompt wording: Removed misleading "native reasoning" reference; prompt now reads "Respond directly with your final answer. Do not wrap your response in any special tags." — applicable to all model types
• Remove model alias resolution: Dropped active_model update from response.model — requires_think_final_tags() returns false for both alias and resolved names, so resolution is unnecessary

Change Type

• Bug fix
• New feature
• Refactor
• Documentation
• CI/Infrastructure
• Security
• Dependencies

Linked Issue

None

Validation

• cargo fmt --all -- --check
• cargo clippy --all --benches --tests --examples --all-features -- -D warnings
• cargo build
• Relevant tests pass: unit tests for requires_think_final_tags(), system prompt formatting, reasoning serde alias
• cargo test --features integration if database-backed or integration behavior changed
• Manual testing: confirmed via direct curl testing against NEAR AI staging API; cargo test --lib — 4174 passed, 0 failed

Security Impact

None

Database Impact

None

Blast Radius

Touches src/llm/reasoning.rs, src/llm/reasoning_models.rs, src/llm/nearai_chat.rs. Could affect prompt formatting for all models routed through the NEAR AI provider, but the change is conservative — unknown/alias models now default to the simpler direct-answer prompt rather than injecting <think>/<final> tags.

Rollback Plan

Revert commits on this branch. No schema or config changes; rollback is a straight revert.

Review Follow-Through

• Model alias resolution (active_model from response.model) was prototyped and removed — requires_think_final_tags() returns false for both "auto" and any resolved name, making the resolution unnecessary.
• Direct-answer system prompt wording was updated to be model-agnostic per reviewer feedback.

---

Review track: B (feature/maintainer-requested refactor)

[Copilot]

Pull request overview

This PR hardens the LLM “reasoning” pipeline by making the safe direct-answer prompt the default (skipping <think>/<final> injection for unknown/alias models), improving compatibility with native-thinking models and NEAR AI’s "auto" alias behavior.

Changes:

• Invert the prompt-injection default via requires_think_final_tags() (empty allowlist) so unknown/alias models use the direct-answer format.
• Accept vLLM/SGLang’s reasoning field via #[serde(alias = "reasoning")] when deserializing NEAR AI chat responses.
• Update NEAR AI provider active_model from response.model so post-alias capability checks use the resolved model name.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
src/llm/reasoning.rs	Switches system-prompt formatting default to direct-answer and updates related tests.
src/llm/reasoning_models.rs	Introduces requires_think_final_tags() allowlist-based decision (currently empty) and adds unit tests.
src/llm/nearai_chat.rs	Parses response.model, resolves model aliases into active_model, and adds serde alias/tests for reasoning.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Same issue as in complete(): comparing response.model to self.config.model will keep re-setting active_model on every request when the configured model is an alias (e.g. "auto"). Compare against the current active_model instead so this is only done when the resolved name changes.

Suggested change

	if let Some(ref resolved) = response.model
	&& *resolved != self.config.model
	let active_model = self.active_model_name();
	if let Some(ref resolved) = response.model
	&& *resolved != active_model

[henrypark133]

Same — removed in 112ab13.

[Copilot]

Removed in 112ab13 — same resolution as the complete() instance.

[gemini-code-assist]

Code Review

This pull request enhances model alias resolution and reasoning field handling for the NEAR AI provider. It introduces a "safe default" for prompt injection, only applying / tags to explicitly allowlisted models to avoid breaking native-thinking LLMs. Feedback suggests optimizing the model resolution logic to prevent redundant write locks and ensuring that per-request model overrides do not inadvertently update the provider's global state.

I am having trouble creating individual review comments. Click here to see my feedback.

src/llm/nearai_chat.rs (491-495)

The current check against self.config.model leads to redundant write lock acquisitions on every request when using an alias like "auto", because the resolved model name will always differ from the original configuration. Comparing against self.active_model_name() avoids this overhead. Furthermore, adding a check for req.model.is_none() prevents per-request model overrides from unexpectedly updating the provider's global active_model state.

        if req.model.is_none()
            && let Some(ref resolved) = response.model
            && resolved != &self.active_model_name()
        {
            let _ = self.set_model(resolved);
        }

src/llm/nearai_chat.rs (577-581)

The current check against self.config.model leads to redundant write lock acquisitions on every request when using an alias like "auto", because the resolved model name will always differ from the original configuration. Comparing against self.active_model_name() avoids this overhead. Furthermore, adding a check for req.model.is_none() prevents per-request model overrides from unexpectedly updating the provider's global active_model state.

        if req.model.is_none()
            && let Some(ref resolved) = response.model
            && resolved != &self.active_model_name()
        {
            let _ = self.set_model(resolved);
        }

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[serrrfirat]

Review finding from the current PR diff:

1. Medium - the prompt-behavior change is global, but the new coverage only proves local string/deserialization behavior.

Reasoning::build_system_prompt_with_tools() now sends every model down the direct-answer path unless requires_think_final_tags() returns true, and that allowlist is currently empty. In practice this is broader than the NEAR AI "auto" / unknown-model fix: it removes the old <think>/<final> prompt path for every provider/model, with no runtime escape hatch if a non-native model still depends on that framing.

The tests added here only verify prompt-string construction and the reasoning serde alias in nearai_chat.rs. There still isn’t an end-to-end regression test showing that a regular non-native model preserves correct final-answer / tool-call behavior under the new shared default.

Suggested fix: either keep a scoped override/allowlist path for any known tag-dependent models, or add an end-to-end respond_with_tools() regression test that exercises a normal non-native model under the new default so this global behavior change is actually covered.