Summary

OpenClaw 2026.4.14 (and prior 2026.4.x) hardcodes google-generative-ai to tagged reasoning output mode in dist/provider-utils-CWYcvxhN.js:

const BUILTIN_REASONING_OUTPUT_MODES = { "google-generative-ai": "tagged" };

tagged mode injects this directive into the system prompt (via dist/system-prompt-mHpoeHEN.js:358-367):

ALL internal reasoning MUST be inside <think>...</think>.
Format every reply as <think>...</think> then <final>...</final>, with no other text.

But Gemini 2.5-pro / 2.5-flash also receive thinkingConfig: { includeThoughts: true, thinkingBudget|thinkingLevel: ... } from resolveGoogleThinkingConfig in dist/anthropic-vertex-stream-YyrYtvts.js:5385, which enables native thought parts in the response.

Failure mode

When prompts are long/persona-heavy (PM cron loops, devsecops loops), the model:

1. Emits a native thought:true part (3000+ chars)
2. Then opens a <think> text block per the directive
3. Pivots to a tool call before closing </think> and never emits <final>
4. Post-tool-result assistant turn comes back with content: []

Gateway then logs:

[agent/embedded] incomplete turn detected: stopReason=stop payloads=0 — surfacing error to user

The user sees an error message, but the model has already burned ~140K input tokens for nothing. In our environment this added up to ~$20 USD of wasted spend in a few hours on Gemini 2.5 Pro before we caught it.

Reproduction

• OpenClaw 2026.4.14
• google/gemini-2.5-pro or google/gemini-2.5-flash with thinkingDefault: medium
• Long persona system prompt + a tool call request

We have several reproducible session jsonls if helpful — pattern is consistent: one thinking content block, then a text block beginning <think> (no closing tag, no <final>), then a toolCall, then an empty assistant turn.

Proposed fix

Drop google-generative-ai from BUILTIN_REASONING_OUTPUT_MODES. Gemini already exposes native thinking via thoughtSignature / thought:true parts and the thinkingConfig request field — there is no need for tagged reasoning, and combining both confuses the model.

We've patched the dist file locally ({"google-generative-ai": "native"}) and confirmed:

• Smoke tests pass with clean text output
• No `` text-tag pollution in cron sessions
• Native thinking still works
• No regression in other Gemini behaviour

Happy to send a PR if useful.

Workaround for others hit by this

If you can't patch the dist file:

1. Author a tiny provider plugin for `google-generative-ai` that returns `"native"` from `resolveReasoningOutputMode` (hook config: support clawdis.json path for rebranding #17 per docs/plugins/sdk-provider-plugins.md)
2. Or move primaries off Gemini until fixed

Related code

• `dist/provider-utils-CWYcvxhN.js:4` — `BUILTIN_REASONING_OUTPUT_MODES`
• `dist/system-prompt-mHpoeHEN.js:358-367` — tagged-mode directive
• `dist/anthropic-vertex-stream-YyrYtvts.js:5385` — `resolveGoogleThinkingConfig`
• `dist/pi-embedded-utils-ZJNz7jk_.js:101` — `splitThinkingTaggedText` returns null on unclosed tag (the parser path that produces the empty visible text)