Bug type

Regression (worked before, now fails)

Summary

Bug Report: kimi-coding/k2p5 crashes on turn 2+ since v2026.3.7 (preserveSignatures regression)

Summary

kimi-coding with model k2p5 crashes on every multi-turn session since v2026.3.7.
It worked correctly on v2026.3.1 and earlier.

Root cause

The following change in src/agents/transcript-policy.ts between v3.1 and v3.7:

- preserveSignatures: false,
+ preserveSignatures: isAnthropic,

kimi-coding uses modelApi: "anthropic-messages", so isAnthropic = true.
In v3.1, thinking block signatures were always stripped before re-sending context.
In v3.7, kimi-coding now receives its own thinkingSignature blobs back in context — which it cannot handle, causing two distinct crashes:

Crash 1 (@anthropic-ai/sdk lib/MessageStream.js):
_accumulateMessage calls partialParse(jsonBuf) with no try/catch. kimi appends trailing bytes after the closing } in input_json_delta events, causing JSON.parse to throw Unexpected non-whitespace character after JSON at position N.

Crash 2 (@anthropic-ai/sdk core/streaming.js):
yield JSON.parse(sse.data) in fromSSEResponse() — catch block logs and rethrows with no recovery. Same trailing-byte defect, but at the outer SSE event level before MessageStream is reached.

Suggested fix

// transcript-policy.ts
preserveSignatures: isCopilotClaude,  // not all isAnthropic — kimi-coding can't handle re-sent signatures

Or equivalently, extend dropThinkingBlocks to include kimi-coding:

const dropThinkingBlocks = isCopilotClaude || provider === "kimi-coding";

(dropping blocks makes preserveSignatures irrelevant for kimi-coding regardless)

Versions

• OpenClaw: v2026.3.7 (broken) vs v2026.3.1 (working)
• @anthropic-ai/sdk: 0.73.0 (same in both versions — SDK is not the root cause)
• Model: kimi-coding / k2p5

Commit range

git diff 2a8ac974e 42a1394c5 -- src/agents/transcript-policy.ts

Workaround applied locally

Three runtime patches applied to node_modules in the running container — see Upgrade_procedure.md section 3.3 for full re-apply procedure:

1. patches/openclaw/patch-messagestream-kimi-trailingjson.js — wraps partialParse(jsonBuf) in try/catch in MessageStream.js
2. patches/openclaw/patch-json-parse-kimi-trailingjson.js — prefix extraction in json-parse.js before partialParse
3. patches/openclaw/patch-streaming-kimi-trailingjson.js — tolerant fallback in both yield JSON.parse(sse.data) catch blocks in streaming.js
4. dropThinkingBlocks dist bundle patch — const dropThinkingBlocks = isCopilotClaude || provider === "kimi-coding";

GitHub issue

File at: https://github.com/openclaw/openclaw/issues/new

Suggested title:

kimi-coding/k2p5 crashes on turn 2+ since v2026.3.7 (preserveSignatures regression)

Steps to reproduce

1. Configure a model using provider kimi-coding, model ID k2p5
2. Run any agent session (cron, manual, or spawned subagent) that requires more than one LLM turn — e.g. a task that calls a tool and then continues
3. Observe the session crash on turn 2

Minimal example: a cron prompt that calls web_fetch once and then continues reasoning. The crash occurs as soon as the assistant's turn-1 response (which contains a thinking block with a thinkingSignature) is re-sent as context in turn 2.

Expected behavior

The session continues across multiple turns as it did on v2026.3.1. Thinking blocks from prior turns are stripped or handled gracefully before being re-sent to the kimi API.

Actual behavior

The session crashes on turn 2 with:

Unexpected non-whitespace character after JSON at position N

where N equals the exact byte length of the last tool call's argument JSON. The error originates in @anthropic-ai/sdk@0.73.0 in two separate call sites:

• lib/MessageStream.js — partialParse(jsonBuf) called with no surrounding try/catch in _accumulateMessage
• core/streaming.js — yield JSON.parse(sse.data) in fromSSEResponse(), catch block rethrows with no recovery

The session is marked isError=true and terminates with stopReason: "error".

OpenClaw version

v2026.3.7

Operating system

Linux (Docker container, Ubuntu-based) on macOS host

Install method

No response

Logs, screenshots, and evidence

Impact and severity

No response

Additional information

No response