Summary

This PR fixes HTTP 400 errors ("The reasoning_content in the thinking mode must be passed back to the API.") for DeepSeek V3.2+ thinking models (deepseek-v4-pro, deepseek-v4-flash, deepseek-reasoner) accessed via the Kilo Code Gateway (https://api.kilo.ai/api/gateway).

Why this is distinct from #14941, #15228, #15237

Existing open/closed DeepSeek reasoning_content PRs target direct DeepSeek API access (api.deepseek.com) and all assume that setting reasoning_content = "" is sufficient. That assumption does not hold through the Kilo gateway:

1. Kilo strips reasoning / reasoning_content / reasoning_details fields on forward — see Kilo-Org/cloud:apps/web/src/lib/ai-gateway/providers/openrouter/request-helpers.ts, function removeChatCompletionsReasoning, and also injectReasoningIntoContent which falls back to if (reasoning) — a truthy check that drops empty strings. So reasoning_content = "" reaches the upstream DeepSeek provider as absent.

2. DeepSeek V3.2+ thinking mode additionally rejects conversations shaped as tool → user — i.e. any user turn after the last tool result causes the 400 even when reasoning_content is present. I reproduced this directly against api.kilo.ai with every combination of reasoning / reasoning_content / reasoning_details (empty, whitespace, non-empty, native and unified naming). The only shapes that succeed are:

   • No trailing user/assistant after the last tool, or
   • Such trailing text merged into the last tool message's content.

This is the same workaround Roo Code adopted in kilocode-legacy/src/api/transform/r1-format.ts (convertToR1Format(..., { mergeToolResultText: true })). Their comment:

"environment_details text after tool_results would create user messages that cause DeepSeek to drop all previous reasoning_content."

This PR ports that insight into Hermes.

Changes

All in run_agent.py, plus a new test file.

1. _needs_deepseek_thinking_tool_merge() (new)

Returns True when base_url is api.kilo.ai and the model contains "deepseek" but isn't deepseek-chat (which is non-thinking and unaffected).

2. _copy_reasoning_content_for_api() extended

New terminal branch: when the above predicate is true and source_msg has tool_calls but no reasoning_content/reasoning, inject reasoning_content = ".".

Why "." and not "": empirically verified against api.kilo.ai/api/gateway/chat/completions with deepseek-v4-pro:

• reasoning_content = "" → HTTP 400 (falsy, gateway drops it).
• reasoning_content = " " (single space) → HTTP 200.
• reasoning_content = "." → HTTP 200.

Minimum-cost placeholder that survives the gateway's truthy filter.

3. _merge_post_tool_text_into_tool() (new, static)

Walks api_messages, finds the last tool message, and if only user/assistant-text messages (no further tool_calls) follow it, merges their text into the tool's content and truncates the list there. Does not mutate the input; returns a new list. Bails out if another tool-call round follows the last tool result, to preserve structure.

4. Invoked at both api_messages assembly sites

• Main agent loop (after the per-message copy+sanitize pass).
• flush_memories().

Testing

Unit tests (new file tests/run_agent/test_kilo_deepseek_thinking.py)

20 tests covering:

• _needs_deepseek_thinking_tool_merge() matrix: kilo × [v4-pro, v4-flash, deepseek-reasoner, deepseek-chat, non-deepseek] + openrouter-with-deepseek (must be False).
• _merge_post_tool_text_into_tool(): empty list, no tool, no trailing, single trailing user, multiple trailing users, mixed asst-text + user trailing, trailing asst-with-tool-calls (must bail), multi-round with merge only on last, input immutability.
• _copy_reasoning_content_for_api() kilo+deepseek branch: "." injection, preservation of explicit reasoning_content, conversion from reasoning, non-injection on non-kilo, non-injection without tool_calls.

All 20 pass. Existing tests/run_agent/test_provider_parity.py (76 tests) still passes.

End-to-end verification

I replayed a real failing request_dump_*.json captured from a user-session (DeepSeek v4-pro via kilo, mid-agent-loop with 7 trailing user messages after the last tool result — caused by plugin context injection). Before fix: HTTP 400. After fix: HTTP 200 with proper content and reasoning fields returned by the model.

Interaction with #14941, #15228

Compatible. This PR only fires when base_url is api.kilo.ai; #15228 paths run for api.deepseek.com. They do not conflict at runtime. If #15228 merges first, a small rebase of this PR removes any duplicated condition around _copy_reasoning_content_for_api. If this merges first, #15228 can follow up unchanged.

References

• DeepSeek Thinking Mode Docs
• Kilo gateway request-helpers (field stripping)
• Roo Code's mergeToolResultText workaround
• Related issues in other agents: SillyTavern #4857, Continue.dev #8989, Roo-Code #10171.