fix(agent-core-v2): drop interrupted thinking-only assistant messages at step settle

[7Sageer]

Related Issue

Resolve #1404

Related prior work on the v1 engine / shared adapter: #1882 (resume-time drop of fully-empty assistant tails), #2172 (broader v1 resume/projection-layer repair), #1738 (adapter emits an explicit empty content string). This PR fixes the same bug class in the v2 engine only, leaving v1 and the shared kosong adapters untouched.

Problem

On the v2 engine, a turn interrupted while the model is still streaming thinking (user cancel, process kill) leaves the open assistant message holding only a thinking fragment whose provider signature never arrived. The fold's settle step sealed that message into history because a non-empty thinking block is not "vacuous". On the next request the OpenAI-compatible serialization moves the thinking into reasoning_content, producing an assistant message with neither content nor tool_calls; strict gateways reject that shape (400 Invalid assistant message: content or tool_calls must be set), so every follow-up request fails and the session is effectively bricked.

Reproduced from a real session export (0.34.0): cancel during thinking → content.part with an unsigned think fragment → next turn's request carries the think-only assistant message → 400, non-retryable.

What changed

• contextMemory fold: at settle time, an open assistant with no tool calls whose content parts are all vacuous or unsigned thinking is dropped instead of sealed. Signed thinking (encrypted) still counts as sendable output and is preserved verbatim, as reasoning providers require.
• The drop runs inside the fold, so it applies deterministically at live dispatch and on WireService.restore replay — sessions already bricked by this shape are repaired the next time their records are folded, with the wire records themselves untouched.
• Tests: loopEventFold.test.ts flips the "thinking block has real content but no signature" case from seal to drop (the signed-thinking seal case stays), and loop.test.ts gains an end-to-end regression: cancel during thinking → next prompt's LLM input contains no thinking-only assistant message.

Deliberately not changed:

• The record write path: appendInterruptedStreamContent still persists the interrupted fragment at cancel time, so the live transcript and the wire.jsonl record keep it; only the settled history that feeds provider projections drops it.
• v1 (packages/agent-core) and the kosong adapters: v1 never lands partial stream parts in history, and the adapter-layer fixes are being discussed in fix(agent-core): recover sessions interrupted during thinking #2172 / fix(kosong): serialize reasoning-only assistant content #1738.

Validation

• New fold + loop regressions pass; full @moonshot-ai/agent-core-v2 suite 4900 passed.
• Typecheck, lint:imports, and oxlint clean on the changed files.
• Verified live against a strict OpenAI-compatible reasoning endpoint: the pre-fix shape (assistant message with only reasoning_content) returns the exact 400 from Empty assistant message after terminal crash causes 400 on session resume (DeepSeek) #1404; the post-fix request (message absent) returns 200.

Checklist

• I have read the CONTRIBUTING document.
• I have linked a related issue, or explained the problem above.
• I have added tests that prove my feature works.
• Ran gen-changesets skill, or this PR needs no changeset.
• Ran gen-docs skill, or this PR needs no doc update.

[changeset-bot]

🦋 Changeset detected

Latest commit: b01e499

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
@moonshot-ai/kimi-code	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[pkg-pr-new]

pnpm dlx https://pkg.pr.new/@moonshot-ai/kimi-code@b01e499

npx https://pkg.pr.new/@moonshot-ai/kimi-code@b01e499

commit: b01e499