OpenCode threads lose background shell results after the 30-minute idle release #18880

What happened

OpenCode threads lose the result of a background shell command that runs longer than 30 minutes.

In an OpenCode thread, I asked the agent to run sleep 2400; echo finished-2400 as a background command, end its turn, and report the output when the command finished. The reply never appeared in the T3 thread. The same prompt with sleep 120 or sleep 600 works fine. Claude Code and Codex threads don't have this problem.

Diagnosis

The provider session's idle release doesn't wait for background shell commands in the OpenCode 2 adapter. It only waits for background subagents.

• ProviderSessionManager releases a provider session DEFAULT_IDLE_TIMEOUT_MS (30 min) after its turn goes idle. The only exception is when the adapter's hasPendingBackgroundWork returns true, which pins the session for up to DEFAULT_MAX_IDLE_PIN_MS (4 h). This is in apps/server/src/orchestration-v2/ProviderSessionManager.ts:54-55 and releaseIfStillIdle (~L1048).
• The OpenCode adapter implements this as [...threads.values()].some(owesWork) (packages/provider-opencode/src/server/v2/adapter.ts:3771). owesWork (L1039) counts wakes, queued reports, busy subagent sessions and state.calls entries with background: true. state.calls only holds subagent tool calls. That is the [Bug]: session reaper counts opencode background agents activity as idle time on remote connection (headless) #13514 fix.
• A background shell tool is tracked in turn.tools. OpenCode returns session.tool.success immediately with "Command moved to the background (shell ID: …)". The adapter then emits the tool as completed and calls turn.tools.delete(id) (L2307). Nothing records the running shell job, so owesWork is false and the session is released 30 minutes after the turn's idle.
• OpenCode itself keeps the job running. When the job ends, it adds a synthetic message (metadata.source: "shell", state: "completed") and starts a follow-up turn on the session. That follow-up happens after T3 has already deleted its MCP registration and released the session, so the reply never reaches the thread.
• Claude and Codex are unaffected because their hasPendingBackgroundWork covers background shells. Claude checks the pendingBackgroundTasksByNativeThread roster (ClaudeAdapterV2.ts:8844). Codex checks runningCommandItemsByTurn (CodexAdapterV2.ts:6423).

Possible fix: track background shell jobs in the OpenCode adapter. They are identified by input.background === true on session.tool.called, or by the shell ID in the success content/metadata. Count them in owesWork and clear them when the synthetic shell-completion message (metadata.shellID / jobID) arrives or the follow-up turn runs. This matches how background subagents are handled.

The code is unchanged on main as of 2026-10-13: owesWork is identical, and the last commits to adapter.ts are #14752 and #14744 on 2026-10-11.

Steps to reproduce

1. Start an OpenCode (v2) thread. Any model works; I used opencode-go/claude-haiku-5-5.
2. Send: "Run sleep 2400; echo finished-2400 as a background command, then end your turn. When it finishes, tell me what it printed and the time."
3. The agent starts the shell in the background and ends its turn. Leave the thread alone.
4. After 30 minutes, T3 deletes the thread's OpenCode MCP registration and releases the provider session.
5. At 40 minutes, OpenCode wakes the agent and it replies, as recorded in opencode.db. The reply never appears in the T3 thread.

Control: with sleep 120 or sleep 600, the reply arrives normally, because the wake happens before the 30-minute release.

Version

0.0.46-nightly.20261012.3035 (511d20c) on the remote environment where this happened. Same code on main and 0.0.46-nightly.20261013.3042.

Environment

Remote T3 environment: Ubuntu 26.04.1 LTS, Node 24.21.0, opencode v2.0.27. Client: T3 desktop on Windows 11.

Evidence

# T3 server trace on the remote env: the MCP session is deleted exactly 30 min after each thread's last OpenCode `idle`
2026-10-13T00:31:22Z DELETE /api/experimental/mcp/t3-code-a375a814-…   # sleep 120:  last idle 00:01:22, reply delivered
2026-10-13T00:40:20Z DELETE /api/experimental/mcp/t3-code-09c09d91-…   # sleep 600:  last idle 00:10:20, reply delivered
2026-10-13T00:30:37Z DELETE /api/experimental/mcp/t3-code-34b83bc9-…   # sleep 2400: idle after first turn 00:00:37
2026-10-13T00:30:37Z DELETE http://127.0.0.1:3773/mcp?codemode=false   (204)

# opencode.db, session for the 2400 s thread (session_message, timestamps UTC)
00:00:34  assistant  tool shell {"command":"sleep 2400; echo finished-2400","background":true}
                     -> "Command moved to the background (shell ID: sh_…)"
00:00:37  idle       {"outcome":"succeeded"}
00:40:34  synthetic  {"metadata":{"source":"shell","state":"completed","exit":0}, "text":"<shell …>finished-2400</shell>"}
00:40:37  assistant  "The background command finished and printed: finished-2400 … 13:40:37 NZDT"
00:40:39  idle       {"outcome":"succeeded"}

Related issues

#13514 (closed) fixed the same idle release for OpenCode background subagents by adding state.calls to owesWork. Background shell commands go through a different code path (turn.tools) and were not covered. #11730 (closed) dealt with background delegations showing as Settled. Neither covers background shells.

Fix applied or workaround

None applied. The 30-minute idle timeout isn't user-configurable; only tests override idleTimeoutMs. Workaround: in OpenCode threads, keep background commands under 30 minutes, or use a Claude/Codex thread for longer background jobs.

Filed by

Claude Code (claude-opus-5-5) via t3 triage