fix(agents): recover reasoning-only vLLM OpenAI-compat replies by liuxiaopai-ai · Pull Request #31746 · openclaw/openclaw

liuxiaopai-ai · 2026-03-02T13:59:41Z

Summary

Describe the problem and fix in 2–5 bullets:

Problem: some non-OpenAI openai-completions providers (reported with vLLM + DeepSeek) can return assistant turns where user-visible content is empty while reasoning is present, resulting in empty OpenClaw replies.
Why it matters: users see blank assistant output even though inference succeeded.
What changed: added a guarded fallback in extractAssistantText to recover text from reasoning blocks only for non-OpenAI providers on openai-completions, and only when the turn is not a tool-call/error turn.
What did NOT change (scope boundary): OpenAI/OpenAI Codex provider behavior is unchanged; tool-call turns still do not surface reasoning as normal reply text.

Change Type (select all)

Scope (select all touched areas)

Linked Issue/PR

User-visible / Behavior Changes

For non-OpenAI providers configured with api: openai-completions (for example vLLM), reasoning-only assistant responses no longer get dropped as empty text in the normal reply path.

Security Impact (required)

New permissions/capabilities? (No)
Secrets/tokens handling changed? (No)
New/changed network calls? (No)
Command/tool execution surface changed? (No)
Data access scope changed? (No)
If any Yes, explain risk + mitigation:

Repro + Verification

Environment

OS: macOS (local test run)
Runtime/container: Node 22 + pnpm
Model/provider: unit-test coverage for openai-completions + provider gating
Integration/channel (if any): embedded agent subscribe/extract path
Relevant config (redacted): models.providers.vllm.api = "openai-completions"

Steps

Create an assistant message where user-visible text is empty and only thinking/reasoning content exists.
Run extraction through extractAssistantText.
Compare output for provider=vllm vs provider=openai and for tool-call turns.

Expected

Non-OpenAI OpenAI-compat providers recover text instead of returning empty.
OpenAI provider behavior remains unchanged.
Tool-call turns do not leak reasoning as reply text.

Actual

Matches expected with new tests.

Evidence

Attach at least one:

Failing test/log before + passing after
Trace/log snippets
Screenshot/recording
Perf numbers (if relevant)

Human Verification (required)

What you personally verified (not just CI), and how:

Verified scenarios:
- pnpm test src/agents/pi-embedded-utils.test.ts
- pnpm test src/agents/pi-embedded-subscribe.subscribe-embedded-pi-session.subscribeembeddedpisession.test.ts
- pnpm check
Edge cases checked:
- openai provider does not use fallback
- tool-call turn does not use fallback
- think-tag-only content recovers after promotion to thinking blocks
What you did not verify:
- live vLLM endpoint end-to-end in this branch run

Compatibility / Migration

Backward compatible? (Yes)
Config/env changes? (No)
Migration needed? (No)
If yes, exact upgrade steps:

Failure Recovery (if this breaks)

How to disable/revert this change quickly: revert commit fix(agents): recover reasoning-only openai-compat replies.
Files/config to restore: src/agents/pi-embedded-utils.ts.
Known bad symptoms reviewers should watch for: unexpected reasoning text shown for providers that previously returned empty replies.

Risks and Mitigations

List only real risks for this PR. Add/remove entries as needed. If none, write None.

Risk: non-OpenAI OpenAI-compat providers could surface reasoning text in edge cases where a provider intentionally separates private reasoning.
- Mitigation: fallback is strictly gated (non-openai provider, openai-completions API, no tool calls, non-error turn) and covered by regression tests.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0c2b73ac42

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-03-02T14:05:07Z

src/agents/pi-embedded-utils.ts

+  const fallbackThinkingText =
+    !extracted && shouldFallbackToThinkingText(msg)
+      ? stripThinkingTagsFromText(extractAssistantThinking(msg)).trim()
+      : "";
+  const textForUser = extracted || fallbackThinkingText;


Skip thinking fallback when reasoning output is enabled

The new fallback returns extractAssistantThinking(msg) as user text whenever text blocks are empty, but in the message-end flow we also independently emit reasoning from the same thinking blocks (extractAssistantThinking → formatReasoningMessage in src/agents/pi-embedded-subscribe.handlers.messages.ts). For non-OpenAI openai-completions sessions with reasoningMode: "on" and block replies enabled, a thinking-only assistant turn now produces duplicate output (once as normal answer text and once as a reasoning message), which is a regression from the previous behavior where only reasoning was emitted.

Useful? React with 👍 / 👎.

greptile-apps · 2026-03-02T14:05:22Z

Greptile Summary

Fixes empty assistant responses for non-OpenAI openai-completions providers (vLLM, DeepSeek) when inference succeeds but returns reasoning-only content with empty text blocks.

Key changes:

Added shouldFallbackToThinkingText() guard that only activates for non-OpenAI providers using openai-completions API
Modified extractAssistantText() to recover text from reasoning blocks when normal extraction returns empty
Excludes tool-call turns and error turns from fallback to prevent reasoning leakage
OpenAI and OpenAI Codex provider behavior remains unchanged

Test coverage:

Verified vLLM provider recovers reasoning text when content is empty
Confirmed OpenAI provider still returns empty (no behavioral change)
Ensured tool-call turns don't leak reasoning
Validated think-tag promotion recovery for compat providers

The implementation is well-gated with defensive checks for edge cases (case-insensitive provider comparison, non-string provider values, multiple thinking blocks). Minor caveat: no live vLLM endpoint verification, relying on unit test coverage.

Confidence Score: 4/5

Safe to merge with minimal risk - well-scoped bug fix with comprehensive test coverage
Score reflects thorough implementation with careful gating (API check, provider exclusions, stop reason checks, tool call exclusions), comprehensive test coverage of edge cases, and minimal scope limited to non-OpenAI providers. Not 5/5 due to lack of live vLLM endpoint verification, but unit tests provide strong confidence in correctness
No files require special attention - all changes are well-tested and focused

_{Last reviewed commit: 0c2b73a}

steipete · 2026-03-02T23:44:30Z

Thanks for the PR! Multiple PRs address issue #31598. Keeping #31703 as the earliest submission. Closing to reduce noise. This is an AI-assisted triage review. If we got this wrong, feel free to reopen — happy to revisit.

fix(agents): recover reasoning-only openai-compat replies

0c2b73a

openclaw-barnacle bot added agents Agent runtime and tooling size: S experienced-contributor labels Mar 2, 2026

chatgpt-codex-connector bot reviewed Mar 2, 2026

View reviewed changes

steipete closed this Mar 2, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

fix(agents): recover reasoning-only vLLM OpenAI-compat replies#31746

fix(agents): recover reasoning-only vLLM OpenAI-compat replies#31746
liuxiaopai-ai wants to merge 1 commit intoopenclaw:mainfrom
liuxiaopai-ai:codex/vllm-empty-response-31598

liuxiaopai-ai commented Mar 2, 2026

Uh oh!

chatgpt-codex-connector bot left a comment

Uh oh!

chatgpt-codex-connector bot Mar 2, 2026

Uh oh!

greptile-apps bot commented Mar 2, 2026

Uh oh!

steipete commented Mar 2, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

liuxiaopai-ai commented Mar 2, 2026

Summary

Change Type (select all)

Scope (select all touched areas)

Linked Issue/PR

User-visible / Behavior Changes

Security Impact (required)

Repro + Verification

Environment

Steps

Expected

Actual

Evidence

Human Verification (required)

Compatibility / Migration

Failure Recovery (if this breaks)

Risks and Mitigations

Uh oh!

chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector bot Mar 2, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot commented Mar 2, 2026

Greptile Summary

Confidence Score: 4/5

Uh oh!

steipete commented Mar 2, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants