queued prompts (no files/images/skills/text attachments) are steerable.
- **Background recovery** mirrors Codex's survivable app-server pattern:
host-backed PI runs store the host PID on the run, are treated
Check[]`
- No hardcoded mappings — LLM-driven via `src/ai/detect.ts`
- `collectFingerprint()` async with LLM enrichment (`src/fingerprint/index.ts`)
- MCPserver workspace: `packages/mcp-server/`
## Before Committing
**IMPORTANT:** Before every git commit, you MUST ensure Caliber syncs
silently if
the network is unavailable — it never blocks a commit.
## MCP integration invariants
`mcp_server_ids` (in `AgentDraft`) is the **sole source of truth** for which
MCP integrations
install
npm start
# Interact with characters through the web interface
```
3. **For MCP lessons, use the inspector tool:**
```bash
cd lessons/07-mcp/solution
npm run build
npm run inspect
# Test individual tools
land. After three failures on one
operation, report the blocker.
**Maintain effectiveness without MCPservers.** Where one is unavailable, fall back to web search
and fetch for library documentation
test file per source file**, mirroring the path.
- Run tests via the dart-mcpserver, **only for the files you actually touched**. Do not run the whole suite locally
computing agent. The long-term product direction is a research workbench: local, WSL, SSH servers, GPU hosts, schedulers, literature tools, runs, data assets, artifacts, papers, and decisions should be represented
same goes for code comments.
## Project layout
- `backend/` — Go backend (Fiber HTTP, GORM, MCPserver)
- `frontend/` — Vue3 frontend, and the single source of UI for **both** the web and desktop
brain`, gated by `factoryBrainPolicy` so unattended runs can't write shared context).
- **Connections** are MCPservers: `linear` (`https://mcp.linear.app/mcp`, app-scoped auth shared through `linearAuth` in `agent/lib/constants.ts`). Its connection
following in MATLAB:
```matlab
addpath(' ')
setupAgenticToolkit("install")
```
This handles platform detection, MCPserver download, toolkit installation, agent configuration, and skill registration. Do not attempt to run setup commands
credential values; redact
them via the helpers in `internal/credential/`.
- External processes (Agent Engines, MCPservers, scripts) run in the selected
runtime (`none` or `opensandbox`). When adding new shell execution paths