[
  {"type":"text","text":""},
  {"type":"thinking","thinking":"<model reasoning about calling web_search>","thinkingSignature":""}
]

Relevant fields from model.completed / trace.artifacts:


stopReason: "stop" (not error, not toolUse)
assistantTexts: []
toolMetas: []
didSendViaMessagingTool: false
aborted: false, timedOut: false, idleTimedOut: false, promptErrorSource: null
usage: 141 output tokens (session 1), 83 output tokens (session 2) — all consumed by the thinking block
finalStatus: "success" ← surfaced as success despite zero deliverable content

Notably, the corresponding session .jsonl files were never written — only the .trajectory.jsonl files exist for these two runs. A later success on the same prompt (session 3a5760f4-..., 2026-05-15T14:21:12Z) did write a .jsonl.


Two distinct issues here

Model side: Despite reasoningLevel: "off" in OpenClaw config, the model emitted only thinking content. Either OpenClaw isn't translating "off" into the correct Bedrock Converse API parameter, or openai.gpt-oss-120b-1:0 on Bedrock doesn't honor that flag. (Cannot determine which without the outbound request payload.)

OpenClaw side (this issue's primary ask): Regardless of model behavior, the runtime should detect "turn has empty visible text AND no tool call" and:

Retry, or
Surface a specific actionable error (e.g., "model returned no visible content"), or
Fall back to a configured secondary model
…rather than marking the run success and letting the channel emit the generic fallback.


Suggested fix surface area

Wherever finalStatus is computed for a model turn (likely in the runtime that produces model.completed / trace.artifacts), add a check: if assistantTexts is empty AND toolMetas is empty AND stopReason === "stop", treat the turn as a failed/empty completion and route to retry/fallback/error instead of success.
Consider also: ensure .jsonl session file is flushed even when the turn produces no visible content, so debugging is easier.

### Steps to reproduce

Send above telegram message. The subsequent error is received every time.

### Expected behavior

The web_search tool should be called and the results remitted via the agent.

### Actual behavior

Please reference summary

### OpenClaw version

Unknown (have since upgraded)

### Operating system

AWS Linux 2023

### Install method

npm

### Model

amazon‑bedrock / openai.gpt‑oss‑120b‑1:0

### Provider / routing chain

Telegram -> OpenClaw

### Additional provider/model setup details

_No response_

### Logs, screenshots, and evidence

```shell