[weiddeng]

Stacked on #130; I'll retarget it to main once #130 merges. Refs #129.

Benchmarks compare agents at the same sampling settings. S-Code always samples at temperature 0. It starts with 8,192 output tokens per model call and doubles that after a cut-off response, up to 32,768. This change adds settings to match another agent, for example the Terminus-2 baseline's temperature of 0.6 and 4,096 output tokens.

Changes

• New settings: model.temperature (S_CODE_MODEL_TEMPERATURE, 0 to 2) and model.max_output_tokens (S_CODE_MODEL_MAX_OUTPUT_TOKENS, 1 to 1,000,000).
• An explicit output cap stays fixed instead of doubling.
• The Harbor adapter passes both through as the temperature and max_output_tokens agent kwargs, and records them in provenance.

Backward compatibility

Nothing changes unless these are set: temperature stays 0, and the output budget grows as before. Saved agent runs from before this change still load, because the new fixed-cap flag defaults to off.

Testing

• cargo fmt and cargo clippy -D warnings are clean.
• Tests pass for config (27), agent-core (61), evals and the daemon binary. New tests cover the settings' bounds and check that a fixed cap never grows.
• Adapter tests pass (18), including new tests for the kwargs.

[ylzhang2447]

Looks good. The bounds match between the Rust config (0..=2, 1..=1_000_000) and the Harbor adapter, defaults are preserved (temperature 0, 8192 with doubling), and tests cover the config validation, the fixed-cap no-doubling path, and the adapter kwargs.

One optional point: I didn't see a test that a configured value actually reaches the request. with_model_sampling -> execute_turn wires temperature/max_output_tokens via unwrap_or; a small test that a set value lands in the AgentRunRequest would close the config-to-request gap. Minor, since it's a direct unwrap_or.

[weiddeng]

Added in 71e08c8: configured_sampling_settings_reach_the_model_request runs a turn with and without the settings and checks the request the model provider receives: temperature 0 and 8,192 tokens by default, 0.6 and 4,096 when configured. Could you approve if it looks good?

[ylzhang2447]

The new configured_sampling_settings_reach_the_model_request test covers exactly the config-to-request path I meant. LGTM.