Bug type
Behavior bug (incorrect resource scheduling without a process crash)
Beta release blocker
No
Summary
The automatic local full-suite profile multiplies its host-aware worker budget across nested Vitest shards, causing avoidable peak RAM pressure and OOM risk.
Steps to reproduce
- Resolve local Vitest scheduling for a representative roomy host with 14 CPUs, 48 GB total memory, and no load; the scheduler returns a six-worker host budget.
- Resolve the current local full-suite profile for the same host; it launches ten shard processes with two Vitest workers each.
- Run the isolated full-suite inventory on a fresh Blacksmith Testbox. Individual two-worker leaf runs reach over 2 GB RSS, so multiplying them across 20 workers creates substantially more aggregate memory pressure than the inferred six-worker budget permits.
Expected behavior
The automatic full-suite profile should spend the inferred host worker budget once across the full process tree and keep each shard to one Vitest worker.
Actual behavior
The six-worker host budget selects a ten-shard by two-worker profile, allowing up to 20 Vitest workers plus their separate process and module-graph overhead.
OpenClaw version
Current main at d7e2096dc78.
Operating system
Linux on a fresh Blacksmith Testbox; scheduling behavior is platform-independent.
Install method
Source checkout with pnpm install.
Model
Not applicable; test-runner infrastructure.
Provider / routing chain
Not applicable; test-runner infrastructure.
Additional provider/model setup details
Not applicable.
Logs
The full isolated Testbox inventory and paired aggregate-RSS run will be linked from the fixing PR.
Screenshots, recordings, and evidence
Source-level scheduler reproduction plus fresh Testbox memory sampling; no visual surface is involved.
Impact and severity
- Affected: developers and maintainers running the full local suite on roomy machines.
- Severity: high resource pressure; can OOM the test run or make the workstation unresponsive.
- Frequency: every default full-suite run that selects the large local profile.
- Consequence: avoidable retries, lost test progress, and interference with other local work.
Additional information
The existing Vitest worker cap is sufficient when treated as a global budget. The fix should preserve explicit operator overrides and CI serialization while removing nested automatic concurrency.
Bug type
Behavior bug (incorrect resource scheduling without a process crash)
Beta release blocker
No
Summary
The automatic local full-suite profile multiplies its host-aware worker budget across nested Vitest shards, causing avoidable peak RAM pressure and OOM risk.
Steps to reproduce
Expected behavior
The automatic full-suite profile should spend the inferred host worker budget once across the full process tree and keep each shard to one Vitest worker.
Actual behavior
The six-worker host budget selects a ten-shard by two-worker profile, allowing up to 20 Vitest workers plus their separate process and module-graph overhead.
OpenClaw version
Current
mainatd7e2096dc78.Operating system
Linux on a fresh Blacksmith Testbox; scheduling behavior is platform-independent.
Install method
Source checkout with
pnpm install.Model
Not applicable; test-runner infrastructure.
Provider / routing chain
Not applicable; test-runner infrastructure.
Additional provider/model setup details
Not applicable.
Logs
The full isolated Testbox inventory and paired aggregate-RSS run will be linked from the fixing PR.
Screenshots, recordings, and evidence
Source-level scheduler reproduction plus fresh Testbox memory sampling; no visual surface is involved.
Impact and severity
Additional information
The existing Vitest worker cap is sufficient when treated as a global budget. The fix should preserve explicit operator overrides and CI serialization while removing nested automatic concurrency.