[Bug] Long conversations hang - session/message endpoint regression v1.14.x

[Canahmetozguven]

Bug Description

Long conversations hang indefinitely. After ~20-30 messages, the session/message endpoint stops responding or returns malformed data, causing ACP clients (OpenChamber, OpenWork) to hang.

Environment

• OpenCode version: 1.14.46 (latest tested, also fails on 1.14.48)
• OS: Arch Linux (cachyos)
• Bun version: 1.3.13

Steps to reproduce

1. Use OpenCode via ACP client (e.g. OpenChamber) for an extended conversation
2. After ~20-30 messages, the response stops
3. Client shows spinning/loading indefinitely or silent failure
4. Direct TUI use of the same session works fine

Root cause

• Silvenga (user on [Bug] Long conversations get stuck indefinitely / no response. openchamber/openchamber#1202) investigated the code directly and confirmed: "The session/{session}/message endpoint is handled by OpenCode — OpenCode has had a few changes to that endpoint"
• The regression appears to have started around v1.2.26: "Replace Bun shell with direct spawn calls in core flows" (opencode extremely slow for versions after 1.2.25 #27310)
• v1.14.46 included a partial fix for numeric HTTP API query parameters for session endpoints, but it did NOT fully resolve the issue
• Multiple users on Attachements do not work on WSL #1202 (openchamber repo) confirm it still fails on 1.14.46 and 1.14.48

What might help (from changelog analysis)

v1.14.46 had these session/message fixes:

• Fixed numeric HTTP API query parameters in the generated OpenAPI spec and SDK for session and file endpoints
• Tolerated legacy stored numeric values in sessions, diffs, and retry events
• Fixed old sessions with negative token counts causing message loads to fail

v1.14.34 had:

• Fix session views and nested session items not getting stuck with stale state
• Improve session compaction so long threads keep more useful context

None of these fully fixed the hang for ACP clients.

Expected behavior

ACP clients should be able to load messages from long conversations reliably.

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• SSE /event endpoint does not forward events to clients in v1.14.48 (regression from v1.14.40) #26866: SSE /event endpoint does not forward events to ACP clients in v1.14.48 (same symptom: client hangs waiting for events that never arrive)
• prompt_async silently discards requests in v1.14.44 (SSE path broken) #26635: prompt_async silently discards requests in v1.14.44 — SSE path broken, clients receive no events after server.connected (same ACP integration breakage)

If your issue is specifically the SSE event delivery failure rather than the session/{id}/message endpoint itself, one of the above may already track this regression. Consider leaving a comment there with your findings.