diagnose

One self-diagnosis scan, on demand: read the-loop's own event log for harness-level failures — error-level events, terminal give-ups — debug each new one in an isolated agent one-shot, and file the findings as an issue on the-loop's own repository, labeled the-loop: self-diagnosed (issue-242).

the-loop diagnose [--dry-run]

The automatic form of the same capability is the background watcher the poller and gh-webhook daemons host when selfDiagnosis.enabled is true; this command is for deployments running neither daemon, and for previewing.

Flag	Default	Meaning
--dry-run	off	Build and print the redacted report(s) without posting. Works while the feature is disabled — this is how you see exactly what would leave the machine before opting in.

Without --dry-run, the command refuses unless selfDiagnosis.enabled: true is set in the CLI config — self-diagnosis posts publicly with your own gh credentials, so nothing short of that explicit opt-in posts anything.

What a run does

For each new failure fingerprint in the log: build the redacted dossier (field allow-list, free text scrubbed), run the configured harness (claude | cursor) as a one-shot subprocess in a private temporary directory, compose the issue (summary, root cause hypothesis, suggested fix, trigger evidence, environment), and create it via your gh CLI with the the-loop: self-diagnosed label requested. Every outcome is printed and recorded — in the event log (diagnosis.* events) and in the self-diagnosis ledger, which maps each fingerprint to the issue it produced.

A self-filed issue is never armed: no auto-execute label, no control-keyword comment, control keywords in the body visibly defanged, and the self-authored marker so the-loop's own ingress drops it.

Failures are bounded: an agent that keeps failing abandons the fingerprint after maxRetries attempts, and at most maxIssuesPerDay issues post per rolling day — the excess is deferred to a later scan, never dropped.