[Jules-GP]

Why

A security and performance audit of the stack, in three rounds (findings, then the fix diff twice). This is what it found in the server. Numbers are measured on a local instance with 500 run directories, about 335 events per run.

Performance

Finding	Before	After
runs.active() re-read and re-verified every event of every run at each call; /status, /admin-panel.json (twice), updates.json and the update agent call it, the pages every 2 s	panel 2.5 s per call; live with one admin tab: panel median ~100 s, 4 of 5 /status timed out, which the update agent's wait_idle reads as "server down" and restarts it with runs in flight	panel 9.8 ms, /status 10 ms, no timeout
Staged inputs deleted (rmtree/remove) on the event loop	worst /health 200 ms during a 2 GB discard	4 ms
Admin and status pages fired the next poll whether or not the last one answered	11 requests stacked in 20 s on a slow answer	at most 1 or 2
disk() walked TEMP_DIR and DATA_DIR at every panel poll	~25k stat per poll	1 walk per 30 s
runs.finish() parsed the whole events file on the loop	0.5 s for a chatty tool's 100k lines	off the loop
One nvidia-smi per /status and per panel poll	3 execs for 3 calls	1 per trace interval
Upload part copied once more than needed	2.10x the part in memory	1.10x

The listing is now incremental: one reader per run, kept between listings and keyed on inode, size and mtime, reads only what a run appended, still verifies the signature of every new line, and stops at the terminal event as a whole read does (a 12-seed random-append test checks it gives the same answer as a whole read). Also: the event stream sends a : ping every 15 s of silence (the client's read timeout is 30 s, so a quiet stretch was a reconnect and a full replay); runs a previous server process left pending or running are failed at startup with "The server restarted during this run." (paused runs untouched); container logs are bounded (json-file, 5 x 20 MB).

Security

• Uploads to a model argument are refused (422). deployment.toml opts 14 model arguments out of the name-only convention with "none", which also made them uploadable files; AREG IOS then unpickled such an upload (fixed in the tool, SADT-VISOR#99). The server now refuses an upload to model / *_model in both staging paths, and hides those arguments from the published schema so a panel offers no upload row. *_reference arguments stay uploadable: none feeds a pickle-based loader.
• A cap on each workstation's runs in flight, MAX_RUNS_PER_CLIENT (default 10, 0 = none), 429 before anything is staged. Pending and waiting-for-the-card runs count, paused ones do not; a run sent without X-Run-Id holds a slot while it runs. The count, the registration and the slot claim are one locked step in a worker thread, so two requests cannot both pass, and a cancelled request keeps no slot. A benchmark battery an operator launched is exempt through a per-battery random key that never reaches a tool's environment. The cap reads the client address uvicorn derives, so behind a proxy it depends on FORWARDED_ALLOW_IPS and the proxy overwriting X-Forwarded-For.
• ToolExecutionError was raised in main.py without being imported: a detached run whose result could not be stored failed with a NameError.

Tests

New tests count (calls, stats, fetches, bytes), not time; each failed before its fix. A test that a cancelled /run keeps no slot fails when _off_loop's shield is removed. Server suite: 1554 passed; the known test_dev_mounts failure is unchanged, and the update-agent suite passes 57/57 when rerun. bandit and zizmor clean.

Deferred: AREG CBCT copies the oriented scans out of supervisor scratch instead of moving them (unmeasured).

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.