PR #17210: fix(anthropic_adapter): preserve Kimi thinking blocks and strip signatures to fix intermittent 400s

• Repository: NousResearch/hermes-agent
• Author: Leihb
• State: open | merged: False
• Link: https://github.com/NousResearch/hermes-agent/pull/17210

Description (problem / solution / changelog)

Problem

After upgrading from v0.10.0 to v0.11.0, the CLI no longer displays the Reasoning box when using the kimi-coding provider with kimi-k2.6 model. Additionally, re-enabling thinking causes frequent intermittent HTTP 400 errors during multi-turn tool-calling conversations (closes #17201).

Root Cause 1: Reasoning Box Missing

v0.11.0 introduced two changes that together broke reasoning display for Kimi /coding:

1. api_mode switched to anthropic_messages: _detect_api_mode_for_url() started routing api.kimi.com/coding to the Anthropic transport.
2. thinking parameter was disabled for kimi-coding: To avoid HTTP 400 errors, build_anthropic_kwargs() skipped the thinking parameter entirely for kimi-coding endpoints — which meant Kimi stopped returning thinking blocks.

Fix (commit 70d1ed61): Remove the kimi-coding skip in build_anthropic_kwargs(), re-enabling thinking.

Root Cause 2: Intermittent HTTP 400 on Multi-Turn Tool Calls

Kimi's /coding endpoint returns thinking blocks with real non-empty signatures (long base64-like strings). These are Kimi's own opaque tokens, not Anthropic signatures.

The cleanup code in convert_messages_to_anthropic used a truthiness check:

if b.get("signature") or b.get("data"):
    # Anthropic-signed block — strip
    continue
new_content.append(b)

Since Kimi's signatures are truthy (non-empty strings), this stripped ALL thinking blocks from Kimi assistant messages. On replay, Kimi's history validation requires thinking blocks on tool-call messages — finding none, it returns HTTP 400.

Why It's Intermittent

The 400 only triggers on the 2nd+ API call in a multi-turn conversation with tool calls, because that's when prior assistant messages with (now-stripped) thinking blocks get replayed. First-turn requests always succeed.

Fix

Preserve all Kimi thinking blocks but strip only the signature and data fields. Emit the minimal {"type": "thinking", "thinking": "..."} shape that Kimi accepts for replayed history:

if _is_kimi:
    new_content = []
    for b in m["content"]:
        if not isinstance(b, dict) or b.get("type") not in _THINKING_TYPES:
            new_content.append(b)
            continue
        thinking_text = b.get("thinking", "")
        new_content.append({"type": b["type"], "thinking": thinking_text})
    m["content"] = new_content or [{"type": "text", "text": "(empty)"}]

Verification

• Live 3-turn tool-call conversation against api.kimi.com/coding with kimi-k2.6 — all turns succeeded, no 400 errors
• All 26 related unit tests pass (test_kimi_coding_anthropic_thinking.py + thinking/reasoning tests in test_anthropic_adapter.py)

Changes

• agent/anthropic_adapter.py:
  • Re-enable thinking for kimi-coding (commit 1)
  • Preserve Kimi thinking blocks, strip only signature/data fields (commits 2-3)
• tests/agent/test_kimi_coding_anthropic_thinking.py: Updated test expectations

Changed files

• agent/anthropic_adapter.py (modified, +18/-13)
• tests/agent/test_kimi_coding_anthropic_thinking.py (modified, +7/-8)