fix(agent): enable thinking for Kimi /coding endpoint via Anthropic Messages

[www654cc-pixel]

The kimi-coding provider routes sk-kimi-* keys to https://api.kimi.com/coding and correctly uses anthropic_messages API mode. However, build_anthropic_kwargs() previously skipped the thinking parameter for all Kimi-family endpoints due to concerns that Kimi required reasoning_content on every replayed tool-call message in thinking mode.

I tested directly against the live https://api.kimi.com/coding/v1/messages endpoint:

• Single-turn requests with thinking: {"type": "enabled"} return type: "thinking" blocks
• Multi-turn replay with signed thinking blocks → 200 OK
• Multi-turn replay with unsigned thinking blocks → 200 OK
• Multi-turn replay without any thinking block → 200 OK
• Tool-call conversations (thinking + tool_use + tool_result + follow-up) → 200 OK

Kimi no longer enforces the reasoning_content requirement that motivated the original exclusion. The existing _needs_thinking_reasoning_pad() and _manage_thinking_signatures() machinery already handles the echo-back for Kimi, so no additional changes are needed for multi-turn continuity.

Testing: I applied this fix locally and confirmed that Kimi K2.6 now displays reasoning/thinking output in Hermes agent.

[www654cc-pixel]

Hi @kshitijk4poor @teknium1 — could one of you take a look at this small fix when you have a moment?\n\nIt enables the parameter for the Kimi endpoint in , which was previously skipped for all Kimi-family providers. I validated the change directly against the live endpoint for single-turn, multi-turn replay (signed/unsigned/no thinking blocks), and tool-call conversations — all returned 200 OK.\n\nOnly one file changed (). Happy to address any feedback.

[www654cc-pixel]

Hi @kshitijk4poor @teknium1 — could one of you take a look at this small fix when you have a moment?\n\nIt enables the thinking parameter for the Kimi /coding endpoint in build_anthropic_kwargs(), which was previously skipped for all Kimi-family providers. I validated the change directly against the live https://api.kimi.com/coding/v1/messages endpoint for single-turn, multi-turn replay (signed/unsigned/no thinking blocks), and tool-call conversations — all returned 200 OK.\n\nOnly one file changed (agent/anthropic_adapter.py). Happy to address any feedback.

[www654cc-pixel]

Closing this PR in favor of an issue. The fix is trivial (removes the _is_kimi_coding guard in build_anthropic_kwargs that was incorrectly blocking thinking for all Kimi endpoints) but has been pending review since June 19. See issue for details.