standalone L1 extraction can fail with DeepSeek due to thinking/tool interference

[sirenexcelsior]

Summary

In standalone/gateway mode, L1 extraction can fail against DeepSeek-compatible chat APIs even when the model itself is capable of returning valid JSON.

I reproduced this with DeepSeek chat-style endpoints and saw L1 extraction produce either:

• empty output (output=0 chars)
• non-JSON assistant text
• The reasoning_content in the thinking mode must be passed back to the API.

This leaves L0 recorded successfully while L1 stays empty for the same session.

Reproduction

Environment:

• mode: standalone gateway
• LLM: DeepSeek-compatible OpenAI API
• model: deepseek-chat
• TDAI L1 extraction path (StandaloneLLMRunner with enableTools=false)

Observed behavior:

1. POST /capture records L0 successfully.
2. L1 runner starts for the same session.
3. L1 returns empty/non-JSON output, or errors with reasoning-related message.
4. No L1 records are stored.

Representative logs:

[memory-tdai][l1-extractor] No JSON array found in extraction response
[memory-tdai][l1-extractor] [l1-debug] NO_JSON taskId=l1-extraction, rawLen=0, cleanedLen=0, rawFull=""
[memory-tdai] [standalone-runner] run() failed after 13911ms: The `reasoning_content` in the thinking mode must be passed back to the API.

What seems to be happening

Two details in the standalone runner make L1 extraction brittle for DeepSeek:

1. When enableTools=false, the runner still exposes a read-only tool (read_file).
   For L1 extraction this is undesirable because the prompt expects pure JSON text output, but the model may still opportunistically attempt tool use.

2. There is no way on origin/main to configure standalone gateway requests with disableThinking, so DeepSeek-compatible backends may enter reasoning/thinking mode and produce output that is not compatible with the current L1 JSON parser.

Suggested fix

I tested a minimal fix locally and it resolved the L1 write path for my DeepSeek setup:

1. In src/adapters/standalone/llm-runner.ts

   • when enableTools=false, do not pass tools at all
   • add optional disableThinking?: boolean to StandaloneLLMConfig
   • when disableThinking=true, inject thinking: { type: "disabled" } into the OpenAI-compatible request body

2. In src/gateway/config.ts

   • parse llm.disableThinking from config/env and pass it into StandaloneLLMConfig

Why I think this is a runner/config issue rather than a model issue

I verified separately that:

• direct calls to the same DeepSeek API can return valid JSON
• the same L1 extraction prompt can return valid JSON outside the live gateway pipeline

So the failure appears to be in the standalone runner/config behavior, not in deepseek-chat itself.

Expected behavior

For L1 extraction in standalone mode:

• pure text tasks should not expose tools
• DeepSeek-compatible backends should be able to run with reasoning disabled via config
• successful POST /capture should allow L1 records to be generated for the same session when the extracted output is valid

[MoonMao42]

我也遇到这个问题了,这个本质上llm模式下没有进行思考链的优化,不知道为啥不用现成的litellm,要不就应该做好每个模型的适配.

[Maxwell-Code07]

Thanks for the thorough report! We've confirmed this is a valid bug. Will fix it ASAP.