Skip to content

Mobile: large Claude Code session (≥ ~9 MB JSONL) fails with "the shell ended this stream (overflow)" #28556

Closed

Description

@sjpark-jpg

Summary

On the mobile (phone) build, opening a Claude Code agent session whose transcript JSONL is larger than roughly 9 MB fails with "the shell ended this stream (overflow)". The same session opens fine on the desktop app. We hit this today with Orca 1.4.224 on macOS (host) + the phone companion.

Steps to reproduce

  1. Run a long Claude Code session in an Orca worktree until ~/.claude/projects/<project>/<session>.jsonl is ≥ ~9 MB (large tool outputs accumulate quickly; ours reached 12.8 MB in half a day).
  2. Open that session from the phone app.
  3. The stream aborts with "the shell ended this stream (overflow)"; the session never renders.

Observed

  • Break point ≈ 9 MB (12:38 KST today); a 1 MB session on the same host opens normally.
  • Desktop renders the same session (slowly).

Related

Request

Paginate / tail-load large Claude Code transcripts on mobile (e.g. render the last N turns and lazy-load earlier ones), or expose a configurable overflow cap with a clear error instead of aborting the stream. Partial-cache persistence from #26175 would also help here.

Environment

Orca 1.4.224, macOS 27.0 (host), Claude Code 2.1.x, phone companion app.

Activity

  1. AmethystLiang commented on Oct 14, 2026

    @AmethystLiang
    Contributor

    Thanks for reporting. Looking into it

  2. linear-code commented on Oct 14, 2026

    @linear-code

    from brennan:

    Thanks for reporting this. Jinwoo's fix in PR #28203 keeps mobile chat updates within the phone's size limit so long conversations can load and earlier messages can be paged in.

    The fix is already included in the 1.4.225 release being prepared now, as commit b5ec0640aa3. It is a desktop-side fix and works with mobile 0.0.55, so please update the desktop when that release becomes available.

    We're marking this resolved for the included fix. If the same conversation still fails after updating, please share the desktop and mobile versions and we'll reopen it to investigate.

  3. xinzhoulvxinda commented on Oct 16, 2026

    @xinzhoulvxinda
    Contributor

    Additional evidence for the 1.4.224 failure, including Codex native Chat on a Linux desktop host. I see the maintainer's note that #28203 fixes this in 1.4.225; this is corroboration of the older-version bug, not a claim of a regression in the fixed release.

    The phone reports Could not load conversation: the shell ended this stream (overflow) for an existing long conversation. The desktop can still open it. We need to read earlier messages and continue sending in that same conversation.

    Measured frame mismatch

    • In the 1.4.224 source, the phone bridge caps each JSON message at 655,360 bytes (640 KiB). BridgeHostSubscriptions ends an oversized event stream with overflow.
    • The desktop hydration path uses a roughly 2 MiB history-page budget and sends its initial snapshot as one subscription event. Its 200-item own-agent window can also include subagent items.
    • A read-only agentSession.history tail request (limit: 200) returned 237 items / 1,418,929 bytes. Wrapping that page in a basic snapshot/bridge envelope is 1,419,107 bytes, already over the phone cap before optional command/queue/task fields.
    • A synthetic test using the real journal, desktop subscribers and unmodified mobile bridge reproduces failure on original 1.4.224 for first hydration, stale-epoch reset and a large live catch-up.

    This points to individual serialized frame size, rather than a universal transcript-file-size threshold or an authentication problem.

    Local validation of smaller stream pages

    A local 1.4.224 workaround reuses the existing pagination, with a smaller budget for hydration/reset/live catch-up/stop-note refresh. In an isolated consistent copy of the affected journal, all 1,373 items survived backward paging (6 history pages), forward catch-up (25 batches), client reduction and existing history-RPC reply chunk/reassembly. Initial snapshot was 450,332 bytes, maximum subscription bridge frame 450,353 bytes. There were no substituted oversized-item placeholders, omitted items or backward duplicates. A synthetic user-message append was delivered as a live batch without contacting a model.

    These are transport/state regression tests and offline replay, not installed-build phone UI acceptance. Real queue was empty; provider command catalogs/background-task fields were not replayed. Our smaller-history workaround therefore does not claim to cover arbitrary large attached metadata; the upstream whole-frame/queue handling in #28203 is preferable.

    The official latest release is now 1.4.227. We have not installed/tested that release on the affected phone yet, and are preserving the local patch only as a version-pinned fallback. No private conversation contents or credentials are included here.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions