fix(opencode): unbounded memory growth during active usage#16346
fix(opencode): unbounded memory growth during active usage#16346GooseG17 wants to merge 1 commit intoanomalyco:devfrom
Conversation
Addresses multiple reports of OpenCode processes growing to 100+ GB virtual memory, triggering OOM kills and system crashes (anomalyco#13230, anomalyco#11399, anomalyco#7046, anomalyco#9140). Root cause: Bun/JSC reserves virtual memory regions that accumulate without bound during active LLM streaming and tool execution. Combined with several code-level leaks, the process grows at ~2.3 GB/min of virtual address space until the kernel OOM killer intervenes. Fixes: 1. plugin/index.ts: Bus.subscribeAll() was called without storing the unsubscribe function. Each call to init() added a permanent wildcard subscriber, causing O(n) callback fan-out per bus event. 2. session/prompt.ts: The full conversation history (msgs array) was kept alive during processor.process() after model message conversion. Now released immediately after use. Added Bun.gc(true) after each processing turn and at loop exit to force JSC heap compaction. 3. session/processor.ts: Hoisted currentText/reasoningMap outside try block for proper cleanup on both success and error paths. Clear toolcalls record after processing. Added Bun.gc(false) between steps to curb growth during multi-tool-call sequences. Release stream references before retry. 4. lsp/client.ts: diagnostics Map and files object grew with every unique file path and were never cleared. Now cleared on shutdown(). 5. lsp/index.ts: clients array, broken Set, and spawning Map were left populated after teardown, preventing GC. Now cleared.
github-actions
bot
added
needs:compliance
|
Thanks for your contribution! This PR doesn't have a linked issue. All PRs must reference an existing issue. Please:
See CONTRIBUTING.md for details. |
|
The following comment was made by an LLM, it may be inaccurate: I found several related PRs that address similar memory leak issues: Potential Related PRs:
These PRs are related because they target memory leak issues in the same subsystems: Bus subscriptions, LSP client state, and stream/processor cleanup. PR #15646 appears most directly related as it covers LSP, Bus, and stream cleanup—the same components addressed in PR #16346. |
GooseG17
changed the title
github-actions
bot
removed
needs:compliance
|
Thanks for updating your PR! It now meets our contributing guidelines. 👍 |
1 participant
Issue for this PR
Fixes #13230
Also relates to #11399, #9140
Type of change
What does this PR do?
It fixes the memory leaks that caused opencode to increase its virtual memory at about 2.3 GB/min, leading to application, or even system, crashes.
Five leak/retention points contribute to this:
plugin/index.ts —
Bus.subscribeAll()is called without storing the unsubscribe handle. Eachinit()call stacks another permanent wildcard subscriber. Fixed by storing and calling unsubscribe before re-subscribing.session/prompt.ts — The full
msgsarray stays alive duringprocessor.process()after it's no longer needed. Now released immediately after building model messages, andBun.gc(true)is called after each turn to let JSC compact the heap.session/processor.ts —
toolcalls,reasoningMap, andcurrentTexthold references across retries and after processing finishes. Now cleared on completion, error, and retry paths.Bun.gc(false)added between multi-tool steps.lsp/client.ts —
diagnosticsMap andfilesobject grow with every unique file path and are never cleared. Now cleared on shutdown.lsp/index.ts —
clients,broken,spawningcollections are never cleaned up on teardown. Now cleared.The
Bun.gc()calls are wrapped in try/catch so they're no-ops if the API isn't available.How did you verify your code works?
Monitored memory via
/proc/<pid>/status(VmSize, VmData, VmRSS) and/proc/<pid>/mapsline counts during active streaming sessions before and after the fix. Before: +2.3 GB/min virtual growth, mapping count climbing ~1,278/min. After: virtual memory stabilized and RSS stayed flat across multiple extended sessions. The system no longer OOM-killed after hours of use with multiple opencode instances.Screenshots / recordings
N/A — not a UI change.
Checklist