fix: prevent memory leaks in runner that cause OOM over time

[archon-agent]

Summary

Three fixes for memory leaks in src/runner.ts that cause the bun process to grow unboundedly over time, eventually triggering OOM kills on long-running daemon deployments.

1. Capped stream collection (10 MB limit)

• Replaced unbounded new Response(proc.stdout).text() with a collectStream() helper that caps output at 10 MB
• Prevents runaway or very large Claude responses from consuming all available RAM

2. Timer cleanup after process exit

• The setTimeout for process timeout was never cleared on normal completion
• Added clearTimeout in both success and error paths to prevent timer reference leaks

3. Queue chain reference reset

• task.catch(() => {}) passes through fulfilled values, keeping references to all previous task results in the promise chain
• Changed to task.then(() => {}, () => {}) which discards result references, allowing GC to reclaim them
• Applied to both globalQueue and threadQueues

Context

In production (24/7 daemon serving Telegram bot), these leaks caused the bun process to grow from ~43 MB to 23 GB over several days, eventually triggering an OOM kill and crashing the VM.

Test plan

• Verified fixes on a production VM200 instance running claudeclaw as a systemd daemon
• After applying fixes, daemon memory stayed stable at ~43 MB over extended operation
• No existing tests to run, but changes are minimal and isolated to resource management

🤖 Generated with Claude Code