feat: first-class reasoning/thinking + tool translation for GLM-5.x, Kimi-K2.6/2.7-code, DeepSeek-V4, grok-4.5

[jatmn]

Summary

Brings first-class translation coverage to the listed models in Codex Warp's model-family catalogs. All four providers (Z.ai, Moonshot, DeepSeek, xAI) expose only OpenAI /chat/completions with function tools — no Responses endpoint and no native custom tool type — so "first class special tool calls" means correct custom→function coercion plus faithful reconstruction (already implemented), and the gaps were in reasoning/thinking shaping, reasoning-history round-tripping, and context windows.

• grok-4.5 (was entirely missing): added catalog entry — 500k context, reasoning_effort support with none→low mapping (grok-4.5 cannot disable reasoning), web search, image modality.
• DeepSeek-v4-flash / -pro: fixed context window (256k → 1M, both are 1M per DeepSeek docs); enabled preserve_reasoning_content_history so tool-use turns replay reasoning_content (provider returns 400 otherwise); now forwards reasoning_effort alongside thinking.type.
• GLM-5.x: broad transform now forwards reasoning_effort so GLM-5.2 honors its native effort control (GLM-5/5.1 ignore it, harmless).
• Kimi-K2.6 / K2.7-code: verified already correct (thinking.type=enabled, thinking.keep=all, history preserved for code variants, sampling overrides dropped) — no change.

Rust change

Ported the reasoning_effort_none_value capability into this branch (self-contained, no dependency on other branches): added the field to TransformConfig/TransformConfigPatch schema, wired it through TransformConfigPatch::apply_to, and added apply_reasoning_effort_none_value in transform_morph.rs (called from responses_to_chat). This lets models that cannot disable reasoning map none to a valid level.

Docs / tests

• Updated docs/model-family-catalogs.md and README.md coverage tables (grok-4.5 added; DeepSeek 1M + history note; GLM-5.2 effort note).
• Added first_class_reasoning_and_tool_translation_for_target_models test; updated the existing DeepSeek context assertion to 1M.

Validation

• cargo fmt --check clean
• cargo test --locked — 126 passed, 0 failed
• cargo build --locked succeeds (one pre-existing unrelated dead_code warning already on main)
• git diff --check clean

Notes / follow-ups

• Optional live smoke test against real API keys for grok-4.5 none→low behavior; xAI's own docs conflict on whether none hard-errors, so the mapping is a safeguard.
• DeepSeek V4 reasoning_effort valid values are high/max (low/medium are mapped up by the provider); forwarding the Codex level works as-is.

🤖 Generated with opencode