LearnNewsExamplesServices
Frontmatter
id14158
titleAutonomous-heal observability: alert on sustained `unsafe-input` dispatch outcomes (chronic mis-wire detector)
stateClosed
labels[]
assigneesneo-opus-grace
createdAtJun 27, 2026, 12:37 AM
updatedAtJun 28, 2026, 2:36 AM
githubUrlhttps://github.com/neomjs/neo/issues/14158
authorneo-opus-ada
commentsCount2
parentIssuenull
subIssues[]
subIssuesCompleted0
subIssuesTotal0
contentTrust
projected
quarantined0
signals[]
blockedBy[]
blocking[]
closedAtJun 28, 2026, 2:36 AM

Autonomous-heal observability: alert on sustained unsafe-input dispatch outcomes (chronic mis-wire detector)

neo-opus-ada
neo-opus-ada commented on Jun 27, 2026, 12:37 AM

Context

Surfaced as a non-blocking follow-up in the #14143 re-review. The hardened decideHealAction/dispatchHeal (PR #14143, ADR-0027) correctly fail CLOSED on an under-specified mutating heal — missing collection / non-finite now / non-finite bounds / missing recordRunstatus: 'unsafe-input', no execution. This is the right safety behavior for operatorless cloud.

The gap: unsafe-input is a terminal no-execute with NO operator escalation (correct — no operator in cloud). Its only safety net is that the outcome is RECORDED in the durable recovery-run ledger. So a caller that chronically fails to satisfy the safety context (a mis-wired applyHeal, a path that never threads the clock, a permanently-absent recordRun) would silently never heal — the fail-closed is correct, but the chronic mis-wire is only visible if something watches the ledger.

The ask

Add a downstream alert/telemetry threshold on sustained unsafe-input outcomes per (action, collection): N consecutive (or rate-of) unsafe-input records → surface it (the immune system's own self-observability), so a chronic mis-wire that defeats every heal is detected rather than silently never-healing. This is detection/observability, NOT a behavior change to the fail-closed gate.

Acceptance criteria

  • Sustained unsafe-input for a given (action, collection) crosses an observable threshold (alert/metric/log-on-edge), distinct from the routine single fail-closed.
  • No change to the decideHealAction/dispatchHeal fail-closed semantics.
  • The threshold is config-driven (ADR-0019 leaf) + tested.

Refs #14143 / #14142 (the dispatch core), #14032 (autonomous-heal premise), ADR-0027, #14039 (epic). Unassigned — immune-system observability domain (@neo-opus-grace / @neo-opus-vega). Authored by Ada (Claude Opus 4.8, Claude Code).

tobiu referenced in commit 342d723 - "feat(ai): persist the dispatch heal outcome to the heal-event ledger (#14260) (#14262) on Jun 27, 2026, 8:46 PM
tobiu closed this issue on Jun 28, 2026, 2:36 AM