Skip to content

fix(core): compact once after payload size rejections - #61306

Open
ekeane wants to merge 1 commit into
anomalyco:devfrom
ekeane:payload-recovery
Open

fix(core): compact once after payload size rejections#61306
ekeane wants to merge 1 commit into
anomalyco:devfrom
ekeane:payload-recovery

Conversation

@ekeane

@ekeane ekeane commented Nov 2, 2026 •
edited by opencodejr Bot
Loading

Copy link
Copy Markdown

Issue for this PR

Closes #53107. Related: #53108 (earlier PR for the same issue).

Type of change

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

What does this PR do?

HTTP 413 payload rejections currently leave long sessions resending the same oversized history. Route these failures through the existing one-shot compaction recovery before assistant output starts. Recovery summaries omit inline media without unnecessarily reducing the token budget; original messages and images remain stored. Automatic-compaction opt-out and the existing recovery limit still apply.

This is an independent implementation against current dev.

How did you verify your code works?

  • Root bun run check and core bun run typecheck:tsc pass.
  • Core runner, step, compaction and native-compaction suites: 294 tests pass. Provider-error suite: 46 tests pass.
  • Six new cases cover raw/streamed rejections, stored images, disabled compaction, repeated rejection, failed summaries and partial output. All pass across 10 repetitions; the two recovery cases fail on unchanged dev.
  • Full core suite: 7,192 pass, 35 skip, 3 fail. The three shell cases also fail on unchanged dev with macOS /bin/bash 3.2.

To reproduce the focused coverage: from packages/core, run bun run script/test.ts test/session-runner.test.ts test/session-step.test.ts test/session-compaction.test.ts test/session-native-compaction.test.ts.

Visual

<svg xmlns="http://www.w3.org/2000/svg" width="240" height="72" viewBox="0 0 240 72" role="img" aria-label="A rejected request compacts once and resumes">
  <style>
    @keyframes pulse { 0%,100% { opacity:.45 } 50% { opacity:1 } }
    .compact { animation:pulse 3s infinite }
    .resume { animation:pulse 3s 1s infinite }
    @media (prefers-reduced-motion:reduce) { .compact,.resume { animation:none } }
  </style>
  <g fill="none" stroke="currentColor" opacity=".7">
    <rect x="3" y="16" width="62" height="40" rx="8"/>
    <path d="M69 36h15m-5-5 5 5-5 5M154 36h15m-5-5 5 5-5 5"/>
  </g>
  <g fill="currentColor" font-family="system-ui,sans-serif" font-size="11" text-anchor="middle">
    <text x="34" y="40">413</text>
    <g class="compact"><rect x="89" y="16" width="62" height="40" rx="8" fill="none" stroke="currentColor"/><text x="120" y="40">Compact</text></g>
    <g class="resume"><rect x="175" y="16" width="62" height="40" rx="8" fill="none" stroke="currentColor"/><text x="206" y="40">Resume</text></g>
  </g>
</svg>

Screenshots / recordings

No UI changes; verified with deterministic runner fixtures.

Checklist

  • I have tested my changes locally
  • I have not included unrelated changes in this PR

Added by the opencode bot (AI-generated).

Related: #48095, #52042

opencodejr[bot]
opencodejr Bot reviewed Nov 2, 2026
View reviewed changes

@opencodejr opencodejr Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This looks right to me. I ran the #53107 reproduction from source against a mock gateway that returns 413 for request bodies over 200 KB. On the base, the turn with the image and every --continue after it fail, and the gateway only ever sees requests that still contain the image. With this PR, the 413 leads to one summary request without the image, then a rebuilt request without the image, and both that turn and the next one succeed. In the context-overflow control case, the request rebuilt after compaction also no longer resends the image. The runner, step and compaction tests (289) and typecheck:tsc pass.

Public API: unchanged. The recoverOverflow signature and the new payload value for the Trigger reason are internal to packages/core.

One behaviour to be aware of, not blocking: when the image that triggers the rejection is attached to the current turn, the model then answers that turn with the image left out (the summary says [image/png omitted]) and the user isn't told. That's probably the right trade-off compared with a session that stays stuck, but a visible notice may be worth a follow-up.

This overlaps closely with #53108, which takes the same approach for the same issue and now has merge conflicts with dev.

Related: #48095, #52042

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@opencodejr opencodejr[bot] opencodejr[bot] left review comments

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

fix(core): recover sessions after HTTP 413 payload rejections

1 participant

@ekeane