Every production system breaks; the question is what happens next. AgenticSelfHealingCode closes the heal step of the run → remember → measure → heal loop: it ingests production signals over a signed, HMAC-verified boundary — native Sentry webhooks included — grounds incident diagnosis in real evidence like deploy timelines and repo history, and either recommends the fix with a confirmed verdict or escalates with the named missing evidence. It never guesses. Telemetry content is treated as data, never as instructions, so a prompt-injection attempt inside a stack trace is flagged instead of obeyed; duplicate incidents don't double-page the on-call; and every automatic action passes an earned-autonomy gate with a kill switch and an auto-action ledger. Durable incident memory lives in Postgres + pgvector and survives restarts.
Everything an on-call team wants from a healing agent — and every guardrail a security team demands.
Deploy-linked root-cause analysis grounded in the actual repo and deploy timeline — it recommends a rollback with evidence, or escalates naming exactly what evidence is missing. Never a guess.
Every inbound signal crosses an HMAC-signed boundary; spoofed webhooks are rejected at ingestion, and Sentry is supported natively — no shim.
Telemetry is data, never instructions: a prompt-injection attempt inside an exception message is flagged as a finding, not executed as a command.
Automatic actions live behind an autonomy ladder with human confirmation first, an append-only auto-action ledger, and a signed kill switch that stops everything instantly.
Duplicate incidents are deduplicated instead of double-paging, delivery lands in Telegram with tap-to-ack, and the whole service is operable over HTTP — status, incidents, kill and release.
Incident memory, notify state and the action ledger live in Postgres + pgvector and survive restarts — the system remembers what broke and what fixed it.
To help you with any questions that are not listed here, we offer free, no-obligation consultations
Only as far as autonomy has been earned. The design is adversarial-first: repair proposals start human-confirmed, every automatic action is gated, logged in a ledger, and behind a kill switch. Autonomy is a ladder, not a default.
Just Bun: bun run demo starts the real signal-intake service and plays five scenarios at it over real HTTP with real HMAC signatures — no keys, no database, no config. For real use, add Postgres + pgvector, your Sentry and an Anthropic key for grounded diagnosis.
Apache 2.0, on GitHub (Moai-Team-LLC/AgenticSelfHealingCode). Free to use, self-host and build on, including commercially.
It closes the heal step of run → remember → measure → heal: AgenticOps runs the fleet, AgenticMind remembers, AgenticPerformance measures, and AgenticSelfHealingCode repairs what breaks. Sibling integrations are optional adapters, never hard dependencies.
The full design was red-teamed on paper first, then deliberately reduced to the subset that is safe to operate today — grounded diagnosis, signed intake, gated actions — with the more autonomous loops deferred until they can be earned.
More from our agentic practice — the standard we build to, the open-source layers that implement it, and the team that ships it.
Tired of 3am pages an agent could have triaged? Tell us about your production stack and we'll map what's safe to automate first.