fix(openai-transport): handle Mistral reasoning_content as non-string delta content by rarest · Pull Request #67292 · openclaw/openclaw

rarest · 2026-04-15T16:32:19Z

Fix #67192 — Mistral reasoning (/think) crashes channel with [object Object]

Problem

When using mistral/mistral-small-latest with reasoning enabled (/think high), the channel crashes and outputs NO_REPLY[object Object][object Object]... which permanently corrupts the session.

Root Cause

In src/agents/openai-transport-stream.ts line 1121:

if (choice.delta.content) {
  currentBlock.text += choice.delta.content;

Mistral returns choice.delta.content as a JSON object, not a string. The truthy check passes, and += coerces the object to [object Object].

Fix

Add a typeof guard so only actual string content is treated as text:

if (typeof choice.delta.content === "string") {

Non-string content falls through to getCompletionsReasoningDelta(), which already handles reasoning_content.

Fix #67168 — logging.file config is read but not applied

Problem

Setting logging.file in openclaw.json has no effect — logs always write to /tmp/openclaw/ regardless of configuration.

Root Cause

In src/gateway/config-reload-plan.ts, logging is classified as kind: "none" in BASE_RELOAD_RULES_TAIL. When logging.file changes:

applySnapshot(nextConfig) updates currentConfig in memory
buildGatewayReloadPlan() puts "logging.file" in noopPaths
applyHotReload() runs with empty hotReasons — no subsystem is updated
setRuntimeConfigSnapshot() is never called
resolveSettings() → readLoggingConfig() → loadConfig() → loadPinnedRuntimeConfig() returns the stale startup snapshot

Fix

In applyHotReload(), detect when noopPaths contains a logging.* key and call setRuntimeConfigSnapshot(nextConfig):

const loggingChanged = plan.noopPaths.some(
  (p) => p === "logging" || p.startsWith("logging.")
);
if (loggingChanged) {
  setRuntimeConfigSnapshot(nextConfig);
}

Testing

Mistral Small with /think → no [object Object] output
Change logging.file while gateway runs → logs switch to new path

Mistral's reasoning API returns content blocks as JSON objects (e.g. {type: "thinking", thinking: "..."}) instead of plain strings. When choice.delta.content is a truthy object, the existing if (choice.delta.content) check passes, then currentBlock.text += choice.delta.content coerces the object to '[object Object]', corrupting the chat output. Fix: use typeof guard so only actual string content is treated as text. Non-string content (objects, arrays) falls through to the getCompletionsReasoningDelta() handler, which already parses reasoning_content from the delta structure. Fixes openclaw#67192

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e0fd928945

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 · 2026-04-15T16:35:44Z

      continue;
    }
-    if (choice.delta.content) {
+    if (typeof choice.delta.content === "string") {


Avoid short-circuiting tool call parsing on empty content delta

This condition now treats content: "" as a text delta, and the subsequent continue prevents the same chunk’s tool_calls from being processed. In OpenAI-compatible streams that emit empty-string content alongside initial tool-call tokens, the tool call is dropped entirely, which changes a tool-use turn into a plain stop/text turn and can break tool execution flow.

Useful? React with 👍 / 👎.

greptile-apps · 2026-04-15T16:41:35Z

Greptile Summary

This is a one-line targeted fix replacing the truthy check if (choice.delta.content) with if (typeof choice.delta.content === "string") in the OpenAI-compatible completions stream handler. Mistral's streaming API returns thinking blocks as a JSON object in delta.content rather than a string; the old guard passed the truthy check and += coerced the object to [object Object], permanently corrupting session output. The fix correctly routes non-string content to the existing getCompletionsReasoningDelta() path, which reads reasoning_content from the delta and emits proper thinking events.

Confidence Score: 5/5

Safe to merge — minimal one-line fix with no logic regressions; only finding is a P2 suggestion to add a regression test.

The change is surgical and correct: it tightens a type guard to prevent object-to-string coercion, and the existing fallthrough path already handles the Mistral reasoning delta format. All findings are P2 (missing test).

No files require special attention beyond the optional test addition in src/agents/openai-transport-stream.test.ts.

Prompt To Fix All With AI

This is a comment left during a code review.
Path: src/agents/openai-transport-stream.ts
Line: 1121

Comment:
**Missing regression test for Mistral object-content chunks**

No test covers the fixed case: a chunk where `choice.delta.content` is a plain object (e.g. `{type: "thinking", thinking: "..."}`) alongside a `reasoning_content` string field. Without it there's no automated guard against a future refactor re-introducing the coercion. A small addition to `openai-transport-stream.test.ts` covering a Mistral-style stream where one chunk carries an object `content` + `reasoning_content` delta and a later chunk carries a string `content` would fully close the loop.

How can I resolve this? If you propose a fix, please make it concise.

_{Reviews (1): Last reviewed commit: "fix(openai-transport): guard delta.conte..." | Re-trigger Greptile}

greptile-apps · 2026-04-15T16:41:40Z

      continue;
    }
-    if (choice.delta.content) {
+    if (typeof choice.delta.content === "string") {


Missing regression test for Mistral object-content chunks

No test covers the fixed case: a chunk where choice.delta.content is a plain object (e.g. {type: "thinking", thinking: "..."}) alongside a reasoning_content string field. Without it there's no automated guard against a future refactor re-introducing the coercion. A small addition to openai-transport-stream.test.ts covering a Mistral-style stream where one chunk carries an object content + reasoning_content delta and a later chunk carries a string content would fully close the loop.

Prompt To Fix With AI

This is a comment left during a code review. Path: src/agents/openai-transport-stream.ts Line: 1121 Comment: **Missing regression test for Mistral object-content chunks** No test covers the fixed case: a chunk where `choice.delta.content` is a plain object (e.g. `{type: "thinking", thinking: "..."}`) alongside a `reasoning_content` string field. Without it there's no automated guard against a future refactor re-introducing the coercion. A small addition to `openai-transport-stream.test.ts` covering a Mistral-style stream where one chunk carries an object `content` + `reasoning_content` delta and a later chunk carries a string `content` would fully close the loop. How can I resolve this? If you propose a fix, please make it concise.

The logging.file config key is classified as kind:"none" in BASE_RELOAD_RULES_TAIL, meaning config file changes for it take the noop hot-reload path in applyHotReload(). However, the runtime config snapshot (loadPinnedRuntimeConfig cache) was NOT updated in this path, so resolveSettings() -> readLoggingConfig() -> loadConfig() kept returning the stale STARTUP snapshot and logs continued writing to the pre-change path. Fix: when applyHotReload() sees noopPaths containing a logging.* key, call setRuntimeConfigSnapshot(nextConfig) to refresh the in-process snapshot. The next resolveSettings() call will then read the updated logging.file from disk. Fixes openclaw#67168

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c82f0d0dcc

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 · 2026-04-15T16:52:06Z

+    // logging.file value instead of the stale startup snapshot.
+    const loggingChanged = plan.noopPaths.some((p) => p === "logging" || p.startsWith("logging."));
+    if (loggingChanged) {
+      setRuntimeConfigSnapshot(nextConfig);


Preserve source snapshot when updating runtime config

Calling setRuntimeConfigSnapshot(nextConfig) here drops the paired source snapshot, even though the same hot-reload path has already activated secrets with both runtime+source snapshots. After this line runs, writeConfigFile sees hadBothSnapshots=false and skips source projection, so later config writes derived from runtime state can persist resolved values directly (including SecretRef-backed secrets) instead of preserving source refs. This makes a logging-only reload capable of causing config corruption/secret leakage in subsequent writes.

Useful? React with 👍 / 👎.

prtags · 2026-04-23T09:45:58Z

Related work from PRtags group capable-bison-62uf

Title: Open PR duplicate: [Bug]: Mistral reasoning (/think) crashes channel with [object object] due to missing stream wr…

Number	Title
#67203	fix(mistral): handle content blocks array in reasoning stream
#67292*	fix(openai-transport): handle Mistral reasoning_content as non-string delta content

* This PR

clawsweeper · 2026-04-27T04:49:04Z

Closing this as duplicate or superseded after Codex automated review.

This mixed-scope PR should close as superseded. Current main already addresses the referenced logging.file bug through the logging config path, while the Mistral stream bug remains real but is better tracked by focused parser-side PRs, especially #68418 and related #67203. The submitted branch also has unresolved functional and security/config-write review risks.

Best possible solution:

Close this mixed-scope PR. Keep the shipped logging.file implementation on current main, and route the remaining Mistral parser fix through the focused canonical work in #68418 or the related #67203. A separate logging hot-reload enhancement should only be considered with a source-snapshot-preserving design.

What I checked:

Current main still has the Mistral content coercion path: The OpenAI completions stream still checks truthy choice.delta.content, appends or queues that value as text, and continues before reasoning/tool-call handling. A truthy object or array can still reach the text path on current main. (src/agents/openai-transport-stream.ts:1335, e962381dbf1e)
Reasoning fallback is string-only on current main: getCompletionsReasoningDeltas only accepts non-empty string values for reasoning_content, reasoning, and reasoning_text, so object/array reasoning payloads are not handled there. (src/agents/openai-transport-stream.ts:1481, e962381dbf1e)
Mistral reasoning remains a supported path: The bundled Mistral catalog marks mistral-small-latest as reasoning-capable, and the Mistral compat layer maps OpenClaw thinking levels to Mistral reasoning_effort for that model. (extensions/mistral/model-definitions.ts:62, e962381dbf1e)
Logging half is superseded on main: readLoggingConfig now reads logging settings directly from the active config path resolved by resolveConfigPath, rather than relying on the stale runtime snapshot path described in this PR. (src/logging/config.ts:25, e962381dbf1e)
Logging fix has test and changelog coverage: A regression test writes a config containing logging.file and asserts the logger writes to that configured path. The Unreleased changelog explicitly lists the logging.file fallback fix as fixing [Bug]: logging.file config is read but not applied — logs always write to default /tmp/openclaw/ #67168. (src/logging/logger-redaction-behavior.test.ts:66, e962381dbf1e)
Related PRs cover the remaining Mistral work: The provided PRtags discussion links fix(mistral): handle content blocks array in reasoning stream #67203 as duplicate related work for the Mistral reasoning crash, and fix(agents): unpack typed-block delta.content arrays in openai-completions stream #68418 is an open parser-side candidate that handles typed-block delta.content arrays with regression-test coverage. Those are better canonical paths than this mixed branch.

So I’m closing this here and keeping the remaining discussion on the canonical linked item.

Codex Review notes: model gpt-5.5, reasoning high; reviewed against e962381dbf1e.

openclaw-barnacle Bot added agents Agent runtime and tooling size: XS labels Apr 15, 2026

chatgpt-codex-connector Bot reviewed Apr 15, 2026

View reviewed changes

greptile-apps Bot reviewed Apr 15, 2026

View reviewed changes

openclaw-barnacle Bot added the gateway Gateway runtime label Apr 15, 2026

chatgpt-codex-connector Bot reviewed Apr 15, 2026

View reviewed changes

RaphaAI mentioned this pull request Apr 18, 2026

fix(minimax): promote thinking-only finals to text on openai-completions #68670

Closed

2 tasks

prtags Bot mentioned this pull request Apr 23, 2026

fix(mistral): handle content blocks array in reasoning stream #67203

Closed

clawsweeper Bot mentioned this pull request Apr 27, 2026

[Bug]: Mistral reasoning (/think) crashes channel with [object object] due to missing stream wrapper #67192

Closed

clawsweeper Bot closed this Apr 27, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

fix(openai-transport): handle Mistral reasoning_content as non-string delta content#67292

fix(openai-transport): handle Mistral reasoning_content as non-string delta content#67292
rarest wants to merge 2 commits intoopenclaw:mainfrom
rarest:fix/mistral-reasoning-stream

rarest commented Apr 15, 2026 •

edited

Loading

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

chatgpt-codex-connector Bot Apr 15, 2026

Uh oh!

greptile-apps Bot commented Apr 15, 2026

Uh oh!

greptile-apps Bot Apr 15, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

chatgpt-codex-connector Bot Apr 15, 2026

Uh oh!

prtags Bot commented Apr 23, 2026

Uh oh!

clawsweeper Bot commented Apr 27, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

rarest commented Apr 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Fix #67192 — Mistral reasoning (/think) crashes channel with [object Object]

Problem

Root Cause

Fix

Fix #67168 — logging.file config is read but not applied

Problem

Root Cause

Fix

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Apr 15, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps Bot commented Apr 15, 2026

Greptile Summary

Confidence Score: 5/5

Uh oh!

greptile-apps Bot Apr 15, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Apr 15, 2026

Choose a reason for hiding this comment

Uh oh!

prtags Bot commented Apr 23, 2026

Uh oh!

clawsweeper Bot commented Apr 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

rarest commented Apr 15, 2026 •

edited

Loading

clawsweeper Bot commented Apr 27, 2026 •

edited

Loading