What agents see
When an agent runs, it receives a compiled context made up of three parts:
The manual context is the part you write and maintain.
Edit context
Web UI
Open your workspace and click Context in the left sidebar. The editor lets you write in Markdown. Changes are saved as a draft until you publish them.CLI
Theguild workspace context edit command opens your $EDITOR with the current published context. When you save and exit, a draft is created automatically.
Versions
Every change creates a new version. Versions are eitherDRAFT or PUBLISHED — only the latest published version is used when agents run.
Programmatic access
Agents can read the workspace context history and append new versions through the Guild service. IncludeguildTools in your agent so task.guild is available.
Workspace contexts are append-only versioned snapshots. There is no update or delete API — every call to
create_workspace_context produces a brand-new row. To “edit” the published context, fetch the latest version, merge in your changes, and create a new PUBLISHED version.Read the context history
get_workspace_contexts returns { items, pagination }, with items ordered newest-first. Each Context has id, status ("DRAFT" or "PUBLISHED"), manual_context, generated_context, summary, created_at, updated_at, and author fields.
Publish a new version
create_workspace_context:
Only the latest
PUBLISHED row is used when agents run. Drafts are persisted but inert until you publish them by creating a new row with status: "PUBLISHED".
Writing good context
Context is injected verbatim into the agent’s prompt, so write it as if you’re briefing a capable new team member. Include:- What the project does and its tech stack
- Repository structure and important directories
- Coding conventions and style preferences
- Tools and services the team uses
- Links to relevant internal resources