Give a support agent a reviewed skill

Keep the agent runtime and its instructions separate. AgentTrunk stores the support skill; your runtime handles tickets, tool use, and message delivery.

1. Preview a skill package

From a built public client checkout, set the intended IDs:

1
export AGENTTRUNK_WORKSPACE_ID="WORKSPACE_ID"
2
export AGENTTRUNK_SCOPE_ID="SCOPE_ID"
3
export AGENTTRUNK_CONTEXT_KEY="support-response"
4
node examples/publish-skill/index.mjs

This previews two files: SKILL.md and references/escalation.md. The example does not upload until you explicitly enable writes. Review the preview and use a new context key if you do not intend to replace an existing package.

2. Upload and verify staging

With authentication supplied securely:

1
AGENTTRUNK_WRITE=1 node examples/publish-skill/index.mjs

The example uploads the complete package, inspects its returned revision, and verifies both files. Its receipt includes the revision ID and confirms production was not changed. Evaluate that revision in your support harness.

3. Review before production

Follow review and release. The release includes the entire scope snapshot. Check unrelated staged changes too.

4. Load approved context

1
node examples/pinned-context/index.mjs

The read example discovers production context within your selected workspace, pins a revision, and verifies a file. It selects a matching result for demonstration; your application should choose by its task and expected context key.

The script emits a receipt, not private context bodies. Integrate the returned bytes in your runtime and retain the pin for debugging.

Read the example source.