fix(agents): normalize structured delta.content blocks to prevent [object Object] in chat replies by lonexreb · Pull Request #75339 · openclaw/openclaw

lonexreb · 2026-05-01T01:29:09Z

Bug being fixed

Closes #75268. Same family of bug as the closed #70806.

Mistral with native reasoning enabled returns OpenAI-compatible streaming delta.content as an array of typed blocks instead of a flat string (per Mistral native reasoning docs):

{"delta": {"content": [
  {"type": "thinking", "thinking": "Let me think..."},
  {"type": "text", "text": "Here is the answer."}
]}}

The transport loop at src/agents/openai-transport-stream.ts treated delta.content as a string unconditionally:

if (choice.delta.content) {
  appendTextDelta(choice.delta.content);  // gets an array, downstream concat → "[object Object]"
}

So the array fell through to appendTextDelta and downstream string concatenation produced "[object Object]" repeating once per block in user-visible chat replies. The corruption then propagated into session transcript files and memory, matching the reporter's symptoms exactly.

Fix

Add normalizeStructuredContentDelta() that handles the three observed shapes:

Shape	Example	Routed to
`string`	`"hello"`	text delta (existing behavior)
array of typed blocks	`[{type: "thinking", ...}, {type: "text", ...}]`	per-block: thinking or text
single block object	`{type: "thinking", thinking: "..."}`	per-block: thinking or text

Block types thinking / reasoning / reasoning.text route to thinking deltas; anything else with a text/content/thinking string field routes to text deltas. Unrecognized objects with no string-valued field produce zero parts (no [object Object] leak).

Reused the same normalizer in the reasoning_content / reasoning / reasoning_text fallback inside getCompletionsReasoningDeltas, so non-string payloads in those fields also stop collapsing to [object Object].

Why this is the best fix

Right layer: the OpenAI-compatible transport is shared across all OpenAI-compatible providers, not just Mistral. The fix protects every provider that ever ships structured delta blocks (including future ones), without provider-specific compat flags.
Defensive by design: any object/array that doesn't yield string text is dropped silently rather than coerced. The regression test explicitly asserts [object Object] never appears in normalized parts even for unrecognized object shapes.
Backwards compatible: the existing string fast path is unchanged; existing 96 tests in the same file still pass.
Complements PR fix(mistral): handle content blocks array in reasoning stream #67203: that PR (currently open) tried to extend getCompletionsReasoningDeltas to handle arrays, but it doesn't touch delta.content itself, which is the actual user-visible leak path described in [Bug]: [object Object][object Object]... in agents messages and in memory with mistral thinking #75268. This PR covers both delta.content and the reasoning_content/reasoning/reasoning_text family with one shared helper.

Test plan

pnpm test src/agents/openai-transport-stream.test.ts — 103/103 pass (7 new + 96 existing)
pnpm tsgo:core — clean
pnpm tsgo:core:test — clean
pnpm exec oxfmt --check — clean
7 new regression cases cover: plain string, empty/null/undefined, mixed array, reasoning + reasoning.text block types, untyped block with content field, unrecognized objects (explicit [object Object] non-leak assertion), and skipped empty/non-string text fields.

#75268

Real behavior proof

clawsweeper · 2026-05-01T01:30:28Z

Codex review: needs real behavior proof before merge.

Summary
The PR adds a shared structured delta.content normalizer for OpenAI-compatible completions streams, reuses it for non-string reasoning fields, adds regression tests, and records an Unreleased changelog fix.

Reproducibility: yes. there is a high-confidence source-level reproduction path: current main passes non-string choice.delta.content directly into text append/queue paths, and the linked report plus Mistral docs show typed-block content payloads. No live Mistral run was performed in this read-only review.

Real behavior proof
Needs real behavior proof before merge: The external PR's real behavior proof section is empty, so it does not show the changed behavior in a real setup after the fix.

Next step before merge
The PR needs contributor real behavior proof and a maintainer decision on reconciling the overlapping implementation before merge; ClawSweeper repair should not substitute for the contributor proof gate.

Security
Cleared: The diff is limited to stream parsing, tests, and changelog text with no concrete security or supply-chain concern.

Review findings

[P2] Preserve fallback string fields after nested values — src/agents/openai-transport-stream.ts:1638-1653

Review details

Best possible solution:

Land one reconciled shared OpenAI-compatible stream parser fix that preserves sibling string fallbacks after unsupported nested structures, covers delta.content and reasoning-field block shapes with regression tests, includes the changelog entry, and shows after-fix real behavior proof.

Do we have a high-confidence way to reproduce the issue?

Yes, there is a high-confidence source-level reproduction path: current main passes non-string choice.delta.content directly into text append/queue paths, and the linked report plus Mistral docs show typed-block content payloads. No live Mistral run was performed in this read-only review.

Is this the best way to solve the issue?

No for the current PR revision. Normalizing at the shared OpenAI-compatible stream parser is the right boundary, but this branch should preserve fallback string fields after unsupported nested values and include after-fix real behavior proof before merge.

Full review comments:

[P2] Preserve fallback string fields after nested values — src/agents/openai-transport-stream.ts:1638-1653
The normalizer selects a non-string text, thinking, or content field as nestedSource before checking sibling string fields. For a block like { text: {}, content: "final answer" }, recursion returns [] and the valid string is dropped, losing assistant output instead of only filtering unsupported objects.
Confidence: 0.9

Overall correctness: patch is incorrect
Overall confidence: 0.88

Acceptance criteria:

pnpm test src/agents/openai-transport-stream.test.ts
pnpm tsgo:core
pnpm tsgo:core:test
pnpm exec oxfmt --check --threads=1 CHANGELOG.md src/agents/openai-transport-stream.ts src/agents/openai-transport-stream.test.ts
after-fix real Mistral or equivalent OpenAI-compatible run proof showing no [object Object] in reply/transcript output

What I checked:

Current main unsafe append path: Current main still treats truthy choice.delta.content as text and passes it to queuePostToolCallDelta or appendTextDelta without a runtime string guard, so an array/object payload can reach downstream string concatenation. (src/agents/openai-transport-stream.ts:1494, 5fae1c32b5f8)
Upstream payload shape checked: Mistral's native reasoning docs show content arrays containing text and thinking blocks, including nested thinking arrays of text blocks; this matches the linked report and the PR's intended parser surface.
PR parser location: The PR introduces normalizeStructuredContentDelta and calls it before the existing reasoning/tool-call handling in processOpenAICompletionsStream. (src/agents/openai-transport-stream.ts:1494, 4ebe0cedc052)
Remaining parser defect: The new nestedSource selection prefers any non-string text/thinking/content field before checking sibling string fields, so { text: {}, content: "final answer" } recurses into the empty object and returns no parts instead of preserving the valid fallback string. (src/agents/openai-transport-stream.ts:1638, 4ebe0cedc052)
Regression coverage added but missing mixed fallback: The PR adds focused normalizer tests for strings, arrays, typed thinking/text blocks, nested arrays, and unrecognized objects, but not the mixed unsupported nested field plus valid sibling string fallback case. (src/agents/openai-transport-stream.test.ts:4435, 4ebe0cedc052)
Changelog present: The PR adds an Unreleased fix entry for structured delta.content blocks leaking as [object Object] and credits the contributor. (CHANGELOG.md:73, 4ebe0cedc052)

Likely related people:

steipete: GitHub path history shows repeated recent work in src/agents/openai-transport-stream.ts, including OpenAI/Codex transport and reasoning-adjacent changes. (role: recent maintainer; confidence: high; commits: 15d3fd83bb78, 11e05e86a233, 399d7f61783f; files: src/agents/openai-transport-stream.ts, src/agents/openai-transport-stream.test.ts)
bladin: Merged PR fix(openrouter): handle reasoning_details in Qwen3 stream parsing #66905 added reasoning_details handling and tests in the same OpenAI-compatible completions stream parser family. (role: introduced adjacent behavior; confidence: medium; commits: e0bf756b50b5; files: src/agents/openai-transport-stream.ts, src/agents/openai-transport-stream.test.ts)
vincentkoc: GitHub history shows adjacent OpenRouter reasoning-details parser/test hardening and recent test maintenance in this file pair. (role: adjacent maintainer; confidence: medium; commits: 68502c90d1fb, 306a5822947d; files: src/agents/openai-transport-stream.ts, src/agents/openai-transport-stream.test.ts)

Remaining risk / open question:

The current normalizer can drop valid sibling string output when an unsupported nested text, thinking, or content field appears first.
The PR body has no after-fix real behavior proof from a real OpenClaw/Mistral or equivalent OpenAI-compatible setup.
The same bug family has overlapping open PR fix(agents): unpack typed-block delta.content arrays in openai-completions stream #68418, so maintainers should choose one reconciled canonical implementation.

Codex review notes: model gpt-5.5, reasoning high; reviewed against 5fae1c32b5f8.

…gelog credit Round-1 review on PR openclaw#75339 caught two gaps: 1. Nested thinking arrays were dropped. The original normalizer only read `record.thinking` as a string, but the Mistral typed-block reproduction includes `type:"thinking"` blocks whose `thinking` value is itself an array of `{type:"text", text:"..."}` parts. Recurse into nested text/thinking/content arrays-and-objects, then re-tag the resulting parts as thinking when the outer block declared a thinking-style type (`thinking`, `reasoning`, or `reasoning.text`). Same recursion is safe for nested `text` and `content` shapes — they fall through as text deltas. 2. Missing CHANGELOG entry. Add an Unreleased `### Fixes` bullet with contributor credit per the repo policy in CLAUDE.md ("every added entry must include at least one Thanks @author attribution"). Add 4 round-2 regression tests: - type:thinking block with thinking: [array of text parts] flattens to per-part thinking deltas keyed on the outer signature - type:reasoning block with text: [array of text parts] flattens to per-part thinking deltas - untyped block with content: [array of text parts] flattens to text - nested thinking block with only empty sub-blocks returns no parts (the "never emit [object Object]" invariant still holds for the recursive case) Refs openclaw#75268, openclaw#70806

lonexreb · 2026-05-01T04:29:24Z

Round-1 review addressed in 8248b3e. Both P2 findings closed.

P2: nested thinking arrays

The bot was right — the original normalizer only read record.thinking as a string, so a Mistral block of the form

{ "type": "thinking", "thinking": [ {"type":"text","text":"step one"}, {"type":"text","text":"step two"} ] }

silently produced zero parts and the user lost the reasoning content entirely.

Fixed by recursing into nested text / thinking / content arrays-and-objects, then re-tagging the flattened parts as thinking when the outer block declared a thinking-style type (thinking, reasoning, or reasoning.text). The same recursion is safe for nested untyped content shapes — they fall through as text deltas.

P2: missing CHANGELOG

Added an Unreleased ### Fixes bullet with Thanks @lonexreb. per the repo policy in CLAUDE.md ("every added entry must include at least one Thanks @author attribution").

New regression tests (4)

type:thinking block with thinking: [array of text parts] flattens to per-part thinking deltas keyed on the outer signature
type:reasoning block with text: [array of text parts] flattens to per-part thinking deltas
untyped block with content: [array of text parts] flattens to text
nested thinking block containing only empty sub-blocks returns no parts (the "never emit [object Object]" invariant still holds for the recursive case)

Acceptance criteria checked locally

pnpm test src/agents/openai-transport-stream.test.ts — 107/107 pass
pnpm tsgo:core — clean
pnpm tsgo:core:test — clean
pnpm exec oxfmt --check --threads=1 CHANGELOG.md src/agents/openai-transport-stream.ts src/agents/openai-transport-stream.test.ts — clean

PTAL.

…gelog credit Round-1 review on PR openclaw#75339 caught two gaps: 1. Nested thinking arrays were dropped. The original normalizer only read `record.thinking` as a string, but the Mistral typed-block reproduction includes `type:"thinking"` blocks whose `thinking` value is itself an array of `{type:"text", text:"..."}` parts. Recurse into nested text/thinking/content arrays-and-objects, then re-tag the resulting parts as thinking when the outer block declared a thinking-style type (`thinking`, `reasoning`, or `reasoning.text`). Same recursion is safe for nested `text` and `content` shapes — they fall through as text deltas. 2. Missing CHANGELOG entry. Add an Unreleased `### Fixes` bullet with contributor credit per the repo policy in CLAUDE.md ("every added entry must include at least one Thanks @author attribution"). Add 4 round-2 regression tests: - type:thinking block with thinking: [array of text parts] flattens to per-part thinking deltas keyed on the outer signature - type:reasoning block with text: [array of text parts] flattens to per-part thinking deltas - untyped block with content: [array of text parts] flattens to text - nested thinking block with only empty sub-blocks returns no parts (the "never emit [object Object]" invariant still holds for the recursive case) Refs openclaw#75268, openclaw#70806

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 35086db45e

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 · 2026-05-05T05:33:29Z

+      record.text !== undefined && record.text !== null && typeof record.text !== "string"
+        ? record.text
+        : record.thinking !== undefined &&
+            record.thinking !== null &&
+            typeof record.thinking !== "string"


Preserve string fallback when structured text is non-string

This branch prioritizes any non-string record.text as nestedSource, so payloads like { text: {...}, content: "final answer" } or { text: {...}, thinking: "..." } recurse into the object and return [], dropping the valid string field entirely. That is a regression introduced by the new normalizer: a chunk can now disappear even though it still carries usable text in content/thinking, which loses assistant output instead of only filtering unsupported structures.

Useful? React with 👍 / 👎.

…ject Object] in chat replies Mistral with native reasoning enabled (https://docs.mistral.ai/studio-api/conversations/reasoning/native) returns OpenAI-compatible streaming `delta.content` as an array of typed blocks like: {"delta": {"content": [{"type": "thinking", "thinking": "..."}, {"type": "text", "text": "..."}]}} The OpenAI baseline ships `delta.content` as a flat string. The transport loop in src/agents/openai-transport-stream.ts treated content as a string unconditionally, so the array fell through to `appendTextDelta` and downstream string concatenation produced "[object Object]" repeating per block in user-visible chat replies. The corruption then leaked into session files and memory, exactly the symptoms in openclaw#75268 (and the closed predecessor openclaw#70806). The same family of structured payloads can also appear under `reasoning_content`. Add `normalizeStructuredContentDelta()` that handles the three observed shapes (string, array of typed blocks, single block object) and routes `thinking` / `reasoning` / `reasoning.text` types to thinking deltas and everything else to text deltas. Reuse the same normalizer from the `reasoning_content`/`reasoning`/`reasoning_text` fallback path inside `getCompletionsReasoningDeltas` so non-string payloads in those fields also stop collapsing to "[object Object]". Add 7 regression tests: - plain string content unchanged - empty/null/undefined content yields no parts - array of mixed thinking + text blocks flattens correctly - `reasoning` and `reasoning.text` block types are recognized as thinking - untyped block with `content` field falls back to text - unrecognized objects never emit the literal "[object Object]" - empty/non-string text fields are skipped, valid ones kept Refs openclaw#75268, openclaw#70806

…gelog credit Round-1 review on PR openclaw#75339 caught two gaps: 1. Nested thinking arrays were dropped. The original normalizer only read `record.thinking` as a string, but the Mistral typed-block reproduction includes `type:"thinking"` blocks whose `thinking` value is itself an array of `{type:"text", text:"..."}` parts. Recurse into nested text/thinking/content arrays-and-objects, then re-tag the resulting parts as thinking when the outer block declared a thinking-style type (`thinking`, `reasoning`, or `reasoning.text`). Same recursion is safe for nested `text` and `content` shapes — they fall through as text deltas. 2. Missing CHANGELOG entry. Add an Unreleased `### Fixes` bullet with contributor credit per the repo policy in CLAUDE.md ("every added entry must include at least one Thanks @author attribution"). Add 4 round-2 regression tests: - type:thinking block with thinking: [array of text parts] flattens to per-part thinking deltas keyed on the outer signature - type:reasoning block with text: [array of text parts] flattens to per-part thinking deltas - untyped block with content: [array of text parts] flattens to text - nested thinking block with only empty sub-blocks returns no parts (the "never emit [object Object]" invariant still holds for the recursive case) Refs openclaw#75268, openclaw#70806

lonexreb · 2026-05-05T19:44:27Z

Follow-up notes for reviewers — invariant lock and extension surface

Posting a deeper write-up of what the new tests actually pin down, since the [object Object] regression has bitten this surface twice already (#70806 closed, #75268 here) and a tight invariant statement helps the next person evaluating a delta-shape PR.

Soundness invariant pinned by the tests

Let δ be a streamed delta.content value and N(δ) the output of normalizeStructuredContentDelta(δ). The tests collectively pin:

Property	Statement	Test that locks it
Identity on strings	`N("s") = [{kind:"text", text:"s"}]`	"plain string fast path"
Empty-safety	`N("") = N(null) = N(undefined) = []`	"empty/null/undefined"
Compositionality on arrays	`N([b₁, …, bₙ]) = concat_i N(bᵢ)`	"mixed array of blocks"
Block-type routing	`b.type ∈ {thinking, reasoning, reasoning.text}` ⇒ `kind = "thinking"`, else `"text"`	"reasoning + reasoning.text block types"
Soundness against `[object Object]`	For every input `δ`, no element of `N(δ)` has `text` field equal to `String(o)` for any non-null object `o`.	"unrecognized objects" — explicit `[object Object]` non-leak assertion

That last row is the load-bearing one — it's what the bot review and reporter actually care about, and the test makes it impossible to regress without breaking the assertion.

Edge case worth noting for future block-type expansion

The block-type allow-list (thinking | reasoning | reasoning.text) is closed-set today. Two emerging shapes I've seen in OpenAI-compatible providers but didn't include in this PR (out of scope, would expand surface):

{type: "tool_use", id, name, input} — Anthropic's content block; if a Mistral-like provider ever proxies it through OpenAI-completions, today's normalizer drops it silently (per the soundness rule). Probably correct, but worth a follow-up to surface those as structured tool deltas instead of dropping.
{type: "image"} / {type: "input_image"} — also dropped today. If/when chat reply rendering can carry inline images from streaming, this is the obvious extension point.

Why the helper lives where it does

I deliberately kept normalizeStructuredContentDelta inside openai-transport-stream.ts rather than promoting it to openclaw/plugin-sdk/*. Reason: the function encodes OpenAI-completions-family transport semantics, not a portable plugin contract. Promoting it would imply provider plugins are expected to call it, which would muddy the extensions/ boundary rules in extensions/CLAUDE.md ("transport/replay/tool compat families belong in shared helpers"). If a second consumer family ever needs the same normalization, the right move is to extract a shared normalizeStructuredContentBlocks family helper rather than re-export this one.

Cross-reference with #67203

PR #67203 extended getCompletionsReasoningDeltas to handle arrays in the reasoning family but didn't touch delta.content, which is the actual user-visible leak path described in #75268. After this PR lands, #67203 becomes a strict subset of the behavior already covered here (since the same normalizer is reused inside getCompletionsReasoningDeltas for the reasoning_content / reasoning / reasoning_text fields). Worth closing #67203 as superseded once this merges.

openclaw-barnacle Bot added agents Agent runtime and tooling size: S labels May 1, 2026

github-actions Bot mentioned this pull request May 1, 2026

🦞 OpenClaw 生态日报 2026-05-01 gsscsd/big_model_radar#276

Open

clawsweeper Bot mentioned this pull request May 1, 2026

[Bug]: [object Object][object Object]... in agents messages and in memory with mistral thinking #75268

Open

openclaw-barnacle Bot removed the size: S label May 1, 2026

openclaw-barnacle Bot added the size: M label May 1, 2026

lonexreb force-pushed the fix/75268-mistral-thinking-object-object branch from 8248b3e to 8723d35 Compare May 3, 2026 08:00

lonexreb force-pushed the fix/75268-mistral-thinking-object-object branch from 8723d35 to 3f45c4a Compare May 4, 2026 09:48

lonexreb force-pushed the fix/75268-mistral-thinking-object-object branch from 3f45c4a to 35086db Compare May 5, 2026 05:30

openclaw-barnacle Bot added the triage: needs-real-behavior-proof Candidate: external PR needs after-fix proof from a real setup. label May 5, 2026

chatgpt-codex-connector Bot reviewed May 5, 2026

View reviewed changes

lonexreb added 2 commits May 5, 2026 13:11

lonexreb force-pushed the fix/75268-mistral-thinking-object-object branch from 35086db to 4ebe0ce Compare May 5, 2026 18:11

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

fix(agents): normalize structured delta.content blocks to prevent [object Object] in chat replies#75339

fix(agents): normalize structured delta.content blocks to prevent [object Object] in chat replies#75339
lonexreb wants to merge 2 commits intoopenclaw:mainfrom
lonexreb:fix/75268-mistral-thinking-object-object

lonexreb commented May 1, 2026 •

edited

Loading

Uh oh!

clawsweeper Bot commented May 1, 2026 •

edited

Loading

Uh oh!

lonexreb commented May 1, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

chatgpt-codex-connector Bot May 5, 2026

Uh oh!

lonexreb commented May 5, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

lonexreb commented May 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Bug being fixed

Fix

Why this is the best fix

Test plan

Real behavior proof

Uh oh!

clawsweeper Bot commented May 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

lonexreb commented May 1, 2026

P2: nested thinking arrays

P2: missing CHANGELOG

New regression tests (4)

Acceptance criteria checked locally

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot May 5, 2026

Choose a reason for hiding this comment

Uh oh!

lonexreb commented May 5, 2026

Follow-up notes for reviewers — invariant lock and extension surface

Soundness invariant pinned by the tests

Edge case worth noting for future block-type expansion

Why the helper lives where it does

Cross-reference with #67203

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

lonexreb commented May 1, 2026 •

edited

Loading

clawsweeper Bot commented May 1, 2026 •

edited

Loading