Extension development

Choose an extension by the capability you need: skills organize work, MCP exposes callable tools, and Hooks execute configured actions on agent events. They can work together, but each needs its own verification.

Choose the entry point

For installing existing extensions and configuring connections, see Skill management and MCP servers. This group covers authoring and verification.

Start with one minimal task

Define inputs, target resources, expected results, and stop conditions before implementing one independently verifiable action. Start a skill with one concrete task, MCP with one read-only tool, and a Hook with one specific event. Record actual results using a test directory and samples without secrets.

Test normal input, missing material, out-of-scope input, and execution failure. Failures must identify what is missing or where execution stopped, without manufacturing success. Inspect logs for credentials, full file contents, and private material.

Update and disable

After changing skill steps, tool schemas, or Hook commands, rerun the same inputs and confirm the target agent uses the updated version. Saving configuration, rescanning it, and loading it into the runtime are separate steps.

Write tools need explicit authorization and duplicate-call handling; query the actual result before retrying a timeout. After disabling an extension, inspect files, remote operations, and service-account access it created. Recover or revoke each at its source.