Agent tools: no upper caps, instrument flags pass through, bench seed, web tools

[AdvayMonga]

The user's rule: every tool works as intended, the agent controls it however it wants, and no search space is cut. Cost is the budget's job, not an argument cap.

Changes

• trace / kernel / hostprof:
  • Integers keep only a lower bound (requests, prompt length, output tokens, launch count/skip, seconds, rate).
  • kernel: any regex, or none (profiles every kernel), and any ncu section set.
  • New nsys_args (replaces the default nsys flags), ncu_args and pyspy_args: the instruments' own flags, passed as argv items.
  • Safety: the command line is never a shell and it runs inside the jail, where the agent's own code already runs. So the old character-set rule on the regex protected nothing.
• bench: takes seed (seen split only), recorded in its ledger config. submit and the held-out guard are unchanged.
• Agent built-ins: WebSearch and WebFetch are enabled for Claude.

Not here

• Open internet for the agent's shell. srt refuses an allow-all domain list (* is rejected), so on macOS the agent's shell still reaches only the API host. Open internet comes with the workbench container (lab: container jail on Linux; runtime design (RUNTIME.md) #19 and the stack after it).
• Under srt, WebFetch is limited to the same allowlist.
• Corpus workloads can't be larger than the corpus class (sampling is without replacement); synthetic has no limit.

Tests: lab suite passes; the engine-side test_lab_profile.py passes (11); ruff is clean.

🤖 Generated with Claude Code