Verify an agentic-os install
run_doctorAudit an agentic-os install: runs file-inspection checks natively and returns commands for host-executed tests, with a verdict of 'passed' only when all commands are run.
Instructions
Audits an agentic-os install in a target repo you name. Runs the checks that are pure file inspection natively and returns a verdict for each; the three checks that need executing Python (hook compile+import, canned-event dry-runs, HITL smoke) come back as exact commands in host_must_run for you to run yourself — this server never executes code from a target repository, and two of those three commands sets (dry_runs, hitl_smoke) write and then delete a probe file under .agentic/agents/ in the target repo when you run them (see each entry's why). verdict is "passed" only when every native check passed AND host_must_run is empty; it is "incomplete", never "passed", while host_must_run still has entries — which is every install this server finds, since host_must_run is never empty on a single server-side call. This server alone therefore never returns "passed"; that verdict is only reachable once the host has run the returned commands and folded the result back in.
Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| target_path | Yes | Filesystem path to the repository to audit — an absolute path, or one resolvable from the host process's working directory. Must already exist and be a directory; this server never creates one. |
Output Schema
| Name | Required | Description | Default |
|---|---|---|---|
| checks | Yes | Every native (pure file inspection) check this server ran, including "dependencies" — a permanent, structural placeholder that always reports passed: false because verifying installed plugin sources requires ~/.claude/plugins/installed_plugins.json, a path outside the target repo this server cannot read. It is always reported here for visibility, but it is excluded from both the verdict decision and failures below: it reflects what this server cannot check, not a defect in the target repo. | |
| verdict | Yes | "failed": at least one check other than "dependencies" reported passed: false — a real, native-verifiable problem exists in the target repo. "incomplete": every other native check passed, but host_must_run still has entries the host has not run and folded back in yet — this is the expected, correct result of a server-side-only call, not an error. "passed": every native check passed AND host_must_run is empty — this requires the host to have already run the host_must_run commands from a prior call and re-invoked run_doctor after resolving them (or otherwise established there is nothing left to run); this server alone, in a single call, never returns "passed" — a fresh install-found call always has all three host_must_run entries pending. | |
| failures | Yes | "key: detail" for every entry in checks with passed: false, EXCLUDING "dependencies" — which always reports passed: false but is not a real failure (see checks' description above). An empty failures array means every check this server can actually verify came back clean; it does not by itself mean verdict is "passed" — host_must_run may still have entries outstanding. | |
| installed | Yes | Whether .agentic/agentic-os/install.json exists at target_path and parses as JSON. false means none of the checks below ran — checks will contain the single not-installed sentinel instead. | |
| host_must_run | Yes | Command sets for the three checks that require executing Python (hook compile+import, canned-event dry-runs, HITL smoke). Empty when installed is false (nothing to check yet). Non-empty entries here are why verdict can be "incomplete" even when every entry in checks passed. |