Salvages #13897 by @HiddenPuppy onto current main, with critical follow-up fixes.

Problem

Kimi's /coding endpoint speaks Anthropic Messages protocol but enables thinking server-side. When replaying conversation history, it requires every assistant tool-call message to carry a thinking block — even when we don't send thinking.enabled in the request kwargs (PR #13826 correctly suppressed that). Without this, sessions break after tool calls with:

HTTP 400: thinking is enabled but reasoning_content is missing in assistant tool call message at index N

PR #13975 fixed the chat_completions path in run_agent.py (_copy_reasoning_content_for_api), but the anthropic_messages path (convert_messages_to_anthropic in anthropic_adapter.py) was still missing the reasoning_content → thinking block conversion.

Issue confirmed real: Brand-new gateway sessions (e.g. 20260422_011800_eea181) fail at message index 2 — the very first replay after an initial tool call. Not stale-session artifacts.

Cherry-picked from #13897

• convert_messages_to_anthropic(): When an assistant message has reasoning_content, convert it to a {"type": "thinking", "thinking": ...} block in the Anthropic message format.

Follow-up fixes (found during self-review)

1. Thinking block stripped by signature management (CRITICAL): The thinking block synthesised from reasoning_content was immediately stripped by the third-party signature-stripping code — Kimi is classified as _is_third_party_anthropic_endpoint. Added a Kimi-specific carve-out: preserve unsigned thinking blocks (which we synthesised from reasoning_content) while stripping Anthropic-signed blocks Kimi can't validate.

2. Empty-string reasoning_content silently dropped: The truthiness check (if reasoning_content and ...) evaluates to False for "". But _copy_reasoning_content_for_api() deliberately injects "" as a tier-3 fallback for Kimi tool-call messages with no reasoning. Changed to isinstance(reasoning_content, str) so empty-string reasoning still produces a thinking block.

3. Wrong block ordering: Thinking block was appended AFTER tool_use blocks. Anthropic protocol requires thinking → text → tool_use ordering. Changed to blocks.insert(0, ...).

4. Duplicate thinking content on native Anthropic (CRITICAL): The reasoning_content → thinking block insertion ran for ALL endpoints, not just Kimi. On native Anthropic where reasoning_details already contributes signed thinking blocks, this created a second unsigned thinking block with the same text. The unsigned block would be downgraded to a spurious text block on the last assistant message — inflating context and potentially confusing the model. Added a guard: only insert the reasoning_content thinking block when no thinking blocks already exist from reasoning_details.

Scope verification

• _is_kimi_coding_endpoint() is confirmed a strict subset of _is_third_party_anthropic_endpoint() — the if/elif ordering is correct
• Non-Kimi third-party endpoints (MiniMax, etc.): reasoning_content thinking block gets inserted then stripped → zero behavior change
• Native Anthropic with reasoning_details: duplicate guard prevents insertion → zero behavior change
• Native Anthropic without reasoning_details (edge case — provider switch mid-session): unsigned thinking block inserted → downgraded to text on last assistant. Acceptable.

Files changed

• agent/anthropic_adapter.py — reasoning_content → thinking block conversion + Kimi carve-out in signature stripping + duplicate guard
• scripts/release.py — AUTHOR_MAP entries for @HiddenPuppy

Validation

• 8/8 existing kimi thinking tests pass
• 8/8 kimi build_api_kwargs tests pass
• 23/23 total reasoning-related tests pass
• 8 E2E validations: non-empty, empty-string, signed stripping, unsigned survival, non-Kimi third-party, no-reasoning_content, native Anthropic no-duplication, native Anthropic no-details edge case

Closes #13848. Closes #13897.