fix(core): omit truncated reasoning-only history on recovery - #61439

Issue for this PR

Fixes #60292

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

After a reasoning-only response hits the output limit, the recovery request can replay an assistant message with empty content. Strict OpenAI-compatible providers reject it with 400.

Omit that message from model history only when its finish reason is length and all meaningful parts are reasoning. Keep the original reasoning in stored history and send the existing recovery prompt. Responses containing text or tool calls, and normally completed reasoning, keep their existing handling.

This is separate from #56942's empty-content guard: nonempty reasoning survives that guard. Porting its production filtering logic to the same current baseline still reproduces the 400; this comparison does not claim to test the whole older PR branch.

How did you verify your code works?

macOS ARM64, Bun 1.4.4, frozen install; base ce8ca3084eac.

• The issue's local HTTP fixture, adapted to run the source CLI in standalone mode, returns 400 / exit 1 before this change and succeeds / exits 0 after it. No real model calls.
• The updated runner assertion fails before the fix. Runner/message suites: 273 passed. Text and reasoning output-limit cases repeated ten times each under full-suite load: 20 passed. The test also checks that reasoning remains in stored history.
• bun run check: lint has no errors (warnings remain), all 35 typecheck tasks passed. git diff --check passed.
• Full Core: 7,185 passed / 35 skipped / 4 failed across 311 files. The full suite did not pass.

Failure details:

The three affected files on the clean same-SHA baseline produced 643 passed / 11 skipped / 3 failed, with the same three Shell/Bash failures. After restoring the patch, the cache case and two recovery cases passed 3/3. No assertions, skips or timeouts in the failing tests were changed. Reruns do not replace the full-run result.

To run the focused suite: cd packages/core && bun run script/test.ts test/session-runner.test.ts test/session-runner-message.test.ts.

Linux, Windows, real Kimi service behavior and upstream CI have not been verified locally.

Visual

<svg xmlns="http://www.w3.org/2000/svg" width="240" height="72" viewBox="0 0 240 72" fill="none">
  <style>@keyframes ready{0%,100%{opacity:1}50%{opacity:.35}}.ready{animation:ready 2.4s ease-in-out infinite}@media(prefers-reduced-motion:reduce){.ready{animation:none}}</style>
  <g stroke="currentColor" stroke-width="2">
    <rect x="8" y="14" width="76" height="44" rx="7" opacity=".5"/>
    <path d="M22 27h48M22 36h35M22 45h24" opacity=".6"/>
    <path d="M96 36h36m-8-8 8 8-8 8"/>
    <g class="ready"><rect x="146" y="14" width="86" height="44" rx="7"/><path d="m162 36 8 8 16-18M196 29h22M196 38h16"/></g>
  </g>
</svg>

Screenshots / recordings

No UI change. The before/after check uses the real CLI against a local simulated provider.

Checklist

• I have tested my changes locally within the scope above.
• I have not included unrelated changes in this PR.

Added by the opencode bot (AI-generated).

Related: #46881

Upstream sync on 2026-11-05

Head 203be51d5ec52ab4e5eb7fb59fa8883f9181178b merges dev at 8a12d6b4207a2bef758a20fa922a189fe9dea78d and preserves the PR's existing behavior. The session runner and message conversion suites passed 274 tests. On macOS ARM64 with Bun 1.4.4 and frozen dependencies, bun run check passed lint with no errors and all 35 typecheck tasks. The normal pre-push hook passed, and the fork branch SHA was read back after the fast-forward push.

Earlier runtime and visual evidence above remains tied to its stated commits. These local results do not establish hosted CI approval or release validation.