Separate probability heads remain visible
The result retains operation_probabilities and target_probabilities; a single confidence number is not the entire decision trace.
Inspect the supporting source ↗REPOSITORY ANALYSIS
jev-ego
A browser Agent for ego lite that numbers actionable elements for Jev to choose the next step.
E2 · source reviewedPublic material or source evidence was reviewed. A fixed-version source review requires checking a pinned revision; this is not a reproduction, performance endorsement, or safety guarantee.
- Stars
- 15
- Forks
- 1
- License
- —
- Last commit
- 2026-09-17
What Jev does here
Jev selects an operation and target in one request; a separate helper handles free-form text when needed.
Offers observation, suggestions and execution, while uploads and dialogs still require other browser tools.
Reviewed the fixed-version README and call path.
What task does this project fit?
Use Jev’s indexed browser-action policy in an ego lite TaskSpace.
Material analysed：
From input to action
- Observation and model calls stay in a single ego-browser Node process.
- model.ts validates operation-specific Choice distributions.
- agent.ts records history, executes guarded actions and stops on limits or no progress.
Implementation details that matter
A suggestion is distinct from execution
The command surface separates observe/act from suggest/step. The agent also bounds both history and decisions, so continued predictions cannot run forever.
Inspect the supporting source ↗When to choose it
Choose it when ego lite is already your browser runtime. It avoids the Browser Harness hop but introduces an ego-specific environment and profile model.
How to inspect it yourself
Read validateChoice in model.ts, then the step/run paths in agent.ts. Check that an invalid target fails validation and that history and decision limits can block another step.
Limits of this review
- Requires the ego lite environment described in the pinned README.
- No live browser run or claimed speed improvement was reproduced.