Skip to content

operations.md

Latest commit

 

History

History
63 lines (48 loc) · 2.01 KB

operations.md

File metadata and controls

63 lines (48 loc) · 2.01 KB
Raw

Operations

Local development

Install Bun, then run:

bun install
bun run start
bun test

Run a Worker locally with the corresponding bun run dev:<component> script. Worker configuration, bindings, cron triggers, and required secrets are in each Worker's wrangler.jsonc.

Setup and deployment

The normal deployment path is:

cp .env.example .env
# fill in provider and deployment values
bun run setup

The setup script is incremental: it creates or reuses KV namespaces, deploys Workers, pushes missing secrets, and registers the Telegram webhook. Deploy all Workers explicitly with bun run deploy, or deploy one component with its bun run deploy:<component> script.

OAuth helper scripts are available for Google, GitHub, MantisHub, and Todoist. Never commit .env, Worker secrets, refresh tokens, or local CLI secrets.

Teardown and state

bun run undeploy removes deployed Workers, KV namespaces, and the Telegram webhook while preserving .env for a later setup. The pull and push scripts are available for deployment state synchronization; inspect their behavior before using them against a shared environment.

Configuration checklist

  • Agent: AGENT_TOKEN, AGENT_KV, service bindings, and Telegram dispatch values.
  • LLM Gateway: OPENROUTER_API_KEY.
  • Telegram: bot token, webhook secret, and optional allowed chat IDs.
  • Google: OAuth client ID, client secret, and refresh token.
  • GitHub: GITHUB_TOKEN.
  • MantisHub: connector KV and instance configuration.
  • Todoist: TODOIST_TOKEN.

Operational characteristics

  • Scheduled work has one-minute heartbeat resolution; a task can be about a minute late.
  • Prompt schedules consume model tokens; action schedules do not call the model.
  • Agent memory work is best effort and asynchronous.
  • Telegram chat history is process-local and bounded; Agent KV is the durable memory layer.
  • Agent and Telegram requests can time out on long-running multi-tool work; keep scheduled prompts and connector operations appropriately scoped.