Self-diagnosis options

Options under selfDiagnosis — the-loop noticing that it broke, and filing the bug itself (issue-242). When the daemons hit a harness-level failure — an error-level event, or a terminal give-up like #240's poll.comment_failed — the evidence is already in the event log. Self-diagnosis scans that log, debugs each new failure in an isolated agent one-shot, and posts the findings as an issue on the-loop's own repository, labeled the-loop: self-diagnosed.

Strictly opt-in, off by default. Enabling it means this machine may post redacted failure reports publicly, with your own gh credentials. Before opting in, run the-loop diagnose --dry-run — it prints exactly what would leave the machine, and works while the feature is disabled.

Three properties hold regardless of configuration:

• Everything posted is redacted. Reports are built from a field allow-list (event types, levels, enums, counters); free text is scrubbed of paths, usernames, hostnames, e-mails, tokens and sensitive environment values. Work-item refs and repository names never appear.
• A self-filed issue is never armed. No auto-execute label, no control-keyword comment; keywords inside the body are visibly defanged, and the body carries the self-authored marker so the-loop's own ingress drops it.
• Storms are bounded. One issue per failure fingerprint ever; a rolling daily cap defers (never drops) the excess; repeated agent failures abandon the fingerprint.

The watcher runs inside the daemons that already exist (the poller and the gh-webhook receiver) — there is no fourth service. Deployments running neither daemon use the-loop diagnose.

selfDiagnosis:
  enabled: false
  repo: MadaraUchiha-314/the-loop
  label: "the-loop: self-diagnosed"
  harness: claude
  model: ""
  timeoutSeconds: 900
  intervalSeconds: 3600
  maxIssuesPerDay: 3
  maxRetries: 3

Options

enabled

• Type: boolean
• Default: false

The opt-in. Only a literal true enables anything; an absent section, false, or a malformed section all mean no scan, no watcher thread, no agent run, no post — fail closed.

repo

• Type: string
• Default: MadaraUchiha-314/the-loop

owner/repo the issues are filed on. Defaults to the-loop's own repository — the point of the feature is fixing the-loop, not your project. Point it at a fork if you triage there first.

label

• Type: string
• Default: the-loop: self-diagnosed

Label requested on each created issue. GitHub silently drops the request for callers without triage rights on the target repository; the body names the intended label either way, so the marker survives the permission gap. Must never equal routing.autoExecuteLabel or any polling.sources[].label — the label must mark, never arm.

harness

• Type: string — claude | cursor
• Default: claude

Which agent harness runs the isolated diagnosis one-shot (the same adapter set as routing.defaultHarness). The run is a subprocess with an argv list — never a shell — under timeoutSeconds, in a private temporary directory.

model

• Type: string
• Default: "" (the harness's own default)

Model passed to the harness's one-shot invocation.

timeoutSeconds

• Type: number
• Default: 900

Wall-clock budget for one diagnosis agent run.

intervalSeconds

• Type: number
• Default: 3600

How often the background watcher scans the event log for new failures. A manual the-loop diagnose scans regardless.

maxIssuesPerDay

• Type: integer
• Default: 3

Rolling 24-hour cap on created issues (storm control). Candidates over the cap are deferred to a later scan, never dropped — a real bug is late, not lost.

maxRetries

• Type: integer
• Default: 3

Failed diagnosis attempts (agent run or post) per failure fingerprint before it is abandoned and never retried. The ledger of reported, retrying and abandoned fingerprints is the self-diagnosis state file.