Cursor integration: local truth check loop
May 6, 2026 · View on GitHub
Canonical first-run: first-truth-check.md — same agentskeptic check command and stdout/stderr contract as everywhere else; this page adds Cursor-specific wiring.
This page is for teams using Cursor or agent-assisted coding in their own repositories. It helps agents verify database or state outcomes before they claim a workflow change succeeded.
1) Copy the rule into your repo
Start from examples/cursor/agentskeptic-check.mdc and copy it into your project's Cursor rules folder, typically:
.cursor/rules/agentskeptic-check.mdc
Cursor rule locations can vary by setup; if your workspace uses a different rule path, keep the same file content and contract.
2) Configure a project command
Add a script your agent can run consistently:
{
"scripts": {
"agentskeptic:check": "agentskeptic check --workflow-id wf_complete --events examples/events.ndjson --registry examples/tools.json --db examples/demo.db"
}
}
Replace workflow-id, events path, registry path, and database/state-store path with your own project values.
3) Expected result and claim policy
- stdout contains the Outcome Certificate.
- stderr begins with
truth_check_verdict: trusted|not_trusted|unknown, then the human report. - Agents should only claim the checked workflow is verified when verdict is
trusted. - If verdict is
not_trustedorunknown, treat the workflow as not verified until the blocker is resolved.
4) Local Cursor loop vs CI
- Cursor rule: a thin wrapper that prompts coding agents to run the same local
agentskeptic checkcommand before claiming a workflow is verified. - GitHub Action: a thin CI wrapper around the same
agentskeptic checkcommand (defaultmode: check); seeexamples/github-actions/agentskeptic-check.ymland.github/actions/agentskeptic-check. - Both paths use
agentskeptic checkand the same stdout/stderr semantics. Neither requiresAGENTSKEPTIC_API_KEY.
For CI wiring, see docs/ambient-ci-distribution.md.
5) Optional next step: enforce
agentskeptic enforce remains a later opt-in commercial/stateful step for baseline, drift, and acceptance workflows.
Keep local onboarding focused on agentskeptic check first.
6) Troubleshooting
- Missing events file: verify
--eventspath and generated NDJSON location. - Missing registry file: verify
--registrypath and trackedtools.json. - DB path wrong/unavailable: verify
--db(or equivalent state-store config) points to readable data. - Verdict
unknown: gather missing evidence or reduce the checked scope so trust can be determined. - Verdict
not_trusted: fix mismatched expected vs observed state before claiming success.
For full integration details and command reference, see docs/integrate.md.