JobeBOT Web
Accounts, catalogue, teams, credits, settings, and plan controls.
Start here
One platform, three connected layers
JobeBOT combines the web control plane, a local WorkSpace, and a governed model gateway.
JobeBOT WorkSpace
Project files, preview, editor, terminal, agents, diagnostics, and verification.
Gateway and bridge
Local and configured providers with policy, routing, and certification gates.
How it works
From request to verified result
A prompt is classified before action. The selected agent reads the project, plans the change, uses approved tools, and verifies the rendered result.
- Context and intent come before action.
- Native tool calls or verified recovery are required for writing.
- Completion requires disk evidence, checks, and an honest final status.
JobeBOT WorkSpace
The local surface for real project work
WorkSpace is the project-aware execution surface. It keeps the active folder, file tree, preview, browser diagnostics, terminal, agent activity, and completion contract visible together.
- Project files: inspect and open the actual working tree.
- Preview: load the local URL and verify the rendered result.
- AI Focus: inspect the prompt, model route, activity, and contract.
- Terminal: run approved commands in the project context.
Agents and models
Capability is not the same as certification
Agents coordinate work. Models supply inference. A catalogue entry or successful chat response does not certify autonomous coding.
| Role | Best for | Evidence |
|---|---|---|
| Planner | Breaking down ambiguous work | Current plan tied to the project |
| Coder | Writing files and targeted fixes | Tool calls, read-back, syntax, browser proof |
| Reviewer | Finding risks and missing tests | Findings and line-level evidence |
| Vision reviewer | Inspecting layout and responsive behavior | Rendered screenshots or browser evidence |
Writing files and proving completion
Text in chat is not a saved deliverable
A model response becomes a completed implementation only after the requested behavior exists in the target project and the result is checked.
- The requested behavior appears in the actual target project.
- Only intended files are changed.
- Read-back evidence confirms the saved content.
- Syntax, tests, and browser checks pass where applicable.
- The audit record distinguishes completed, blocked, and unverified work.
Common troubleshooting
Use evidence to find the stalled layer
| Symptom | First check |
|---|---|
| Agent writes prose only | Confirm Agent mode and run a minimal native tool-call probe. |
| Files reported absent | Compare the absolute target path and read back the content. |
| Run stays Working | Inspect provider timeout, cancellation, approval, and transcript state. |
| Preview is stale | Check HTTP status, asset errors, console output, and cache-busting. |
| Model is blocked | Discovery is not certification; read the admission state and evidence. |