fix(opencode): kill server process group + configurable IDLE_TIMEOUT_MS

[glifocat]

Type of Change

• Feature skill - adds a channel or integration (source code changes + SKILL.md)
• Utility skill - adds a standalone tool (code files in .claude/skills/<name>/, no source changes)
• Operational/container skill - adds a workflow or agent skill (SKILL.md only, no source changes)
• Fix - bug fix or security fix to source code
• Simplification - reduces or simplifies source code
• Documentation - docs, README, or CONTRIBUTING changes only

Description

Closes #2148. Closes #2149.

Two related bugs in the OpenCode provider that fire together when a local backend (Ollama, llama.cpp) is slower than the hardcoded 90 s event timeout. Bundled into a single PR because they share container/agent-runner/src/providers/opencode.ts and a small helper.

#2148 — proc.kill('SIGKILL') leaks the underlying binary, holding port 4096

spawn('opencode', ...) runs the npm opencode-ai wrapper script that execs the platform binary opencode-linux-*/bin/opencode — which is the actual port listener on 127.0.0.1:4096. SIGKILL on the wrapper PID either races with the exec or the listener has already detached; the binary survives and the port stays bound. Next spawnOpencodeServer call fails with Failed to start server on port 4096 / EADDRINUSE.

Fix: spawn detached and signal the whole process group via a new killProcessTree(proc) helper that calls process.kill(-pid, 'SIGKILL') (with a fallback to plain proc.kill('SIGKILL') if the negative-PID call throws — covers the case where the spawn never made it into a process group).

Both call sites updated:

• startup-timeout cleanup in spawnOpencodeServer
• destroySharedRuntime

#2149 — Configurable idle timeout

IDLE_TIMEOUT_MS = 90_000 was hardcoded. Used as a between-events watchdog, but on a freshly-prompted session it acts as a TTFT ceiling — fine for cloud APIs (sub-second TTFT), too tight for local 30B+ inference on cold start.

Fix: read OPENCODE_IDLE_TIMEOUT_MS from env, default to 300_000 (5 min). Generous for cloud, just enough for slow local. Per-group override via container.json env, e.g. "OPENCODE_IDLE_TIMEOUT_MS": "600000" — no rebuild needed since src/ is bind-mounted.

Tests

No behavior-changing additions. Manually verified:

• Process group kill: docker exec <container> pgrep -af opencode no longer shows orphan [opencode] <defunct> after a forced timeout; 127.0.0.1:4096 is free immediately.
• Configurable timeout: env override applied; default 300 s confirmed when var unset.

Compounding behavior

Without #2148 fixed, every timeout from the 90 s ceiling (or any idle ceiling) leaks a process and renders the agent container unusable until restarted. Fixing one without the other is half a fix — that's why they're filed together.

For Skills

• SKILL.md contains instructions, not inline code (code goes in separate files)
• SKILL.md is under 500 lines
• I tested this skill on a fresh clone

Not a skill PR — section N/A.

[gavrielc]

Thanks!