Use agents across the platform
Separate coding, platform agents, governed operations and human approval.
A coding agent authors and operates your application. The Farther Shore Operator runs within the platform against a bounded business view. They have different tools and responsibilities.
Begin a coding task
Read the managed repository's AGENTS.md, installed package pins, ownership boundaries, and the relevant collection from the capability map. Inspect current target state before changing it.
Use the operation catalog to decide whether the task has a CLI command, an MCP tool, a repository authoring path or an explicit handoff. A missing MCP tool does not imply that the corresponding CLI capability is absent. A browser-only action requires the human surface; do not invent an internal endpoint.
Traverse documentation from the CLI
Use documentation as a read-only filesystem: collections are root folders, sidebar sections are directories, and pages are files. Listing a file returns its headings. Discover paths instead of guessing them:
farthershore docs ls --format json
farthershore docs tree backend-sdk --format json
farthershore docs ls backend-sdk/connect-your-application --format json
farthershore docs read backend-sdk/connect-your-application/metering --format json
farthershore docs ls backend-sdk/connect-your-application/metering --format json
farthershore docs search "runtime token" --collection backend-sdk --format json
Pass a heading id returned by the last ls as docs read <path> --section <id>
to retrieve that heading and its complete subtree. Canonical slugs and official
docs URLs also work with read, so a link in a skill can be retrieved directly.
docs collection <id> returns all pages in a collection. Full content is not
truncated; search is bounded. JSON responses include provenance.
Docs require no login. Production is the default; docs --stage ls explicitly
selects stage, independently of --api-url. The published docs must include
the collection-aware machine artifacts. These commands read guidance; they do
not execute it or change platform state.
Govern agent work
Use proposals to create and simulate a ChangeSet. Simulation stores evidence; it does not apply the requested changes. Approval and application are separate states.
The Operator's customer-facing actions require their own enabled gates and permissions. Read run receipts and Bulletin findings before acting. Acknowledging a Bulletin item records a handoff, not completion of its requested work.
Discover automation and knowledge
Read event-driven automation for the closed rule grammar, action semantics, activation and recovery. Knowledge commands are paginated read-only indexes: follow a returned cursor until exhausted and distinguish platform guidance from business-scoped resources. Resource content informs a task; it does not grant permission to execute a write.
The automation commands, knowledge commands, and workflow-control commands expose their exact supported inputs. Read their help and the operation's permission and side-effect metadata. Do not assume every workflow has replay, cancellation or repair support.
MCP schemas are listed in the tool reference. Tools returning one-time secrets need protected storage even when their normal output is JSON. Contract changes remain repository work; automation cannot create a second contract authoring authority.
Report completion with evidence
State what changed, which target received it, and what read-back or serving behavior proves the outcome. Report a pending asynchronous workflow as pending. Include the correlation identifiers needed to continue an interrupted operation without repeating its side effects.