Skip to content

Conversation

@gwarmstrong
Copy link
Contributor

Summary

  • Add disable_session_restore flag to skip O(n) session history replay after sandbox worker restarts (enabled by default in config)
  • Add verbose_tool_logging flag to optionally collect per-session timing metrics and log per-call execution times (disabled by default)
  • Decrease code execution timeout from 30s to 10s for faster failure on runaway code
  • Pass --disable-session-restore CLI flag through to python_tool subprocess
  • Update nemo-skills requirement to georgea/super-rl-02062026 (nemo-skills side changes for session restore and dependencies)

- Add disable_session_restore flag (enabled by default in config) to skip
  session history replay after sandbox worker restarts, avoiding O(n)
  replay overhead that caused 76% of >31s tool call wall time
- Add verbose_tool_logging flag (disabled by default) for per-call and
  per-session timing metrics
- Decrease code execution timeout from 30s to 10s for faster failure
- Track tool execution timing per session and report metrics in verify
  response (total time, avg time, timeout counts)
- Detect both internal sandbox timeouts and HTTP/request-level timeouts
- Update nemo-skills requirement to georgea/super-rl-02062026

Signed-off-by: George Armstrong <georgea@nvidia.com>
Only track per-session timing data and aggregate metrics when
verbose_tool_logging is enabled. Avoids unnecessary dict accumulation
and perf_counter overhead when timing is not needed.

Signed-off-by: George Armstrong <georgea@nvidia.com>
@copy-pr-bot
Copy link

copy-pr-bot bot commented Feb 6, 2026

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

bxyu-nvidia
bxyu-nvidia previously approved these changes Feb 6, 2026
verify() now takes a request param for session-based timing. Pass
None in tests since verbose_tool_logging defaults to false.

Signed-off-by: George Armstrong <georgea@nvidia.com>
@bxyu-nvidia bxyu-nvidia merged commit f0c830d into main Feb 7, 2026
5 checks passed
@bxyu-nvidia bxyu-nvidia deleted the georgea/ns-tools-stability branch February 7, 2026 00:11
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants