Skip to content

WORKSPACE_AGENT.md

Latest commit

 

History

History
131 lines (102 loc) · 5.06 KB

WORKSPACE_AGENT.md

File metadata and controls

131 lines (102 loc) · 5.06 KB
Raw

LabCanvas Workspace Agent

LabCanvas now exposes one persistent, tool-capable agent through both the web studio and CLI. The web chat is a direct transport to the agent, not a keyword router. Focused APIs such as scene rendering, figure grids, target dispatch, and WeChat controls remain available as fast routines the agent can call.

Capabilities

The packaged capability contract covers:

  • parametric CAD and Shapr3D-compatible STEP handoff;
  • KiCad PCB generation, ERC/DRC, Gerbers, STEP, and renders;
  • Blender scenes, engineering renders, and .blend artifacts;
  • TeX reports, papers, diagrams, and compiled PDFs;
  • WeChat status, messages, files, queues, and worker routines;
  • LabVIEW/MCP, camera, and isolated virtual-desktop workflows;
  • AgInTi image generation, BioRender, Unity, Unreal, and target bridges.

The agent receives a compact packaged knowledge file at src/agenticapp/knowledge/workspace_agent.md and links to measured repository evidence such as the Shapr3D batch history, OpenHI thread/fit tables, Shapr STEP repair lessons, and cage-print feedback. It still inspects the actual source design before changing geometry.

Web Chat

Start the studio:

PYTHONPATH=src python -m agenticapp webapp start --port 19483

The chat composer provides:

  • Model: Auto for dynamic model selection;
  • GPT-5.6 SOL for an explicit gpt-5.6-sol turn;
  • Effort: Auto, Low, or Medium;
  • Execute or Plan only mode;
  • cancellation and live durable-task polling.

Each browser profile receives a stable conversation id. Codex resumes that conversation's prior thread, so follow-up requests retain context without mixing users or workspaces.

Rooms

The /rooms surface provides durable room-scoped conversations using the same workspace-agent runtime and artifact store:

http://127.0.0.1:8787/rooms

Owner access is local-only unless LABCANVAS_ROOMS_TOKEN is configured. Owner-created invite links are room-scoped, expire, and force participant turns into plan mode. Each room has an independent conversation ID, SQLite message ledger, durable task links, and artifact URLs constrained to the webapp storage directory.

The allowlisted WeCom LabAgent transport uses the shared worker routines for research, TeX/PDF, editable figure generation, CAD, PCB/KiCad, OpenSCAD, and Blender work. It must return both review previews and editable/source artifacts to the originating group. It does not authorize public video publication, payment, manufacturing submission, or other irreversible external actions.

CLI

labcanvas agent capabilities
labcanvas agent chat "Design and render a C-mount sensor holder"
labcanvas agent chat "Rebuild this exact Shapr3D part" --model gpt-5.6-sol --effort medium
labcanvas agent chat "Inspect the KiCad board" --mode plan
labcanvas agent chat "Run a long Blender task" --detach --json
labcanvas agent tasks
labcanvas agent status TASK_ID
labcanvas agent cancel TASK_ID

Use --conversation NAME to preserve a separate CLI thread. --dry-run prints the selected model policy and full operating contract without launching an agent.

Model Policy

Auto routing classifies only the current request:

Request Default effort Default model
Short answer or status low auto-code-review
Analysis, documentation, or ordinary tool work medium auto-code-review
Complex implementation and broad engineering validation high gpt-5.6-sol
Demanding autonomous research, design, or presentation synthesis xhigh gpt-5.6-sol

Explicit low, medium, high, and xhigh values are supported. max and Ultra normalize to xhigh. Environment variables LABCANVAS_AGENT_FAST_MODEL and LABCANVAS_AGENT_STANDARD_MODEL can replace the low/medium auto tiers. High and xhigh use their explicit shared-policy entries. Backend quota/unavailable failures may fall back to AgInTi when the saved setting permits it. New AgInTi versions use the noninteractive aginti run --stdin contract. Older installed versions receive a short command that points to the task's private durable prompt file. The default workspace . keeps either form inside LabCanvas instead of depending on an AgInTiFlow source checkout.

Tasks and Artifacts

Durable state is written under the ignored directory:

output/webapp/agent/tasks/<task-id>/
  task.json
  prompt.md
  agent-result.json
  response.md
  artifacts/

The worker runs outside the HTTP request, so browser refreshes do not cancel a task. It registers declared PNG/PDF/STEP/STL/3MF/Blend/KiCad/TeX and other safe artifact types in the normal LabCanvas canvas. Source designs remain in their proper repository folders; the canvas receives review copies.

Action Boundaries

Normal local design, editing, rendering, compilation, and validation can run in Execute mode. The agent must stop for explicit current-user authorization before payment, final manufacturing submission, public publication, credential changes, destructive deletion, or another irreversible external action. Completion must be supported by real command, file, render, or external-status evidence.