-
-
Actual app: the baseline fails, the isolated patch passes, and check evidence survives closing and reopening the work session.
Model update — October 30, 2026
v0.1.2 defaults to nvidia/Nemotron-3.5-Super-VL ahead of the previous Serverless model’s November 9 retirement. Live inference checks, including tool calls and tool-result follow-ups, and a disposable local coding fixture passed on the replacement. Migration validation records the scope. The video and sandbox/browser results below used the original Nemotron 3 Super; v0.1.1 and the existing video remain available.
Inspiration
A coding agent saying it finished leaves important questions unanswered. Did the patch pass the tests? Did those tests already fail before the change? Did the agent install something that will be missing when the patch reaches your machine?
Prax Bot puts those answers beside the diff. Developers can inspect the evidence, compare alternatives, and choose which changes reach their repository.
What it does
Give Prax a task in a Git repository. NVIDIA Nemotron 3 Super reads files, edits code, and runs commands inside a Nebius Token Factory Sandbox. Commands and edits advance an immutable image, preserving checkpoints of the work.
Prax independently runs configured checks against a clean baseline and against the proposed patch. The Before / After view shows both results, including commands, exit codes, timing, and logs. This distinguishes pre-existing failures from regressions and catches changes that depend on the agent's incidental environment.
Compare starts two attempts from the same checkpoint. Each retains its patch, trajectory, checks, and usage. Choosing A or B loads that candidate for review. Applying selected files is a separate action that writes to the developer's repository.
A persistent Main session can delegate to named work sessions and inspect their results. Sessions retain pending changes and evidence across close, reopen, and server restart. Stop cancels queued instructions and settles the current call before stopping; completed edits remain available.
How we built it
Prax is a Python application with a localhost web interface and terminal front ends. Nemotron 3 Super makes coding decisions through Token Factory's inference API using structured tool calls. The ConTree SDK provides sandbox execution and immutable images for checkpoints, independent attempts, and forked verification.
The verification fork begins with the baseline environment plus the patch that Apply would write. A second fork checks the unchanged baseline. The runtime records that evidence independently of the model's final summary. Verification statuses distinguish passing checks, failures, unavailable evidence, and stale results.
Challenges we ran into
Saved results must keep their meaning after a workspace changes, a session reopens, or a process stops. Prax preserves check provenance, marks stale evidence, prevents double counting of comparison usage, and separates choosing a candidate from applying it.
Our first browser rehearsal exposed a completion guard that prompted the coordinating Main session to edit its own workspace. We corrected that guard. A second attempt exhausted its step budget through non-waiting polling and reported unsupported success. The successful prompt explicitly waits for the child to finish and prohibits duplicate follow-ups or stopping it. Both unsuccessful attempts remain in the published evidence.
Accomplishments that we're proud of
The successful real-browser workflow used 19 Nemotron calls and 73,145 tokens. Main delegated without editing its own workspace; the child passed independent checks against a failing baseline; both comparison candidates passed; and choosing A left the host unchanged. Explicit Apply changed only the intended file, after which the host test passed. Rename, close, and reopen preserved the child's patch and check evidence.
All three browser attempts used 49 calls and 231,692 tokens. The application also passed 383 automated tests plus lint, formatting, type, and JavaScript syntax checks. These small functional fixtures do not establish general coding quality or a SWE-bench Verified score.
The 2:42 public YouTube demo shows actual app footage. Waiting-time cuts, detail crops, held frames, captured host-output cards, synthetic narration, and rehearsal context are disclosed. English captions are included.
What we learned
Immutable sandbox images are useful beyond isolating commands: they make it practical to test proposed changes independently and compare alternatives from the same starting point. Clear evidence and recovery behavior are product features.
Evaluation needs the same care. Our benchmark reports preserve failures, account for incomplete attempts, and distinguish development samples from broader evidence. Unsuccessful prompt experiments remain documented.
What's next for Prax Bot
Broader task evaluation and more language/toolchain coverage are planned. Improvements to session navigation and check-result inspection will follow user feedback.
Source, test build, and evidence
- Public source and MIT license
- v0.1.2 test build
- Setup and judge testing instructions
- Browser validation, including unsuccessful attempts
- Technology feedback
Agent execution requires Nebius Token Factory inference and project Sandbox access. The public release contains the complete source, installable wheel, checksums, and reproducible fixture.
Built With
- javascript
- nebius-token-factory
- nvidia-nemotron
- python
