APIbase v1.6.0 — Autopilot

Five months and 761 commits since v1.5.0. The headline is Autopilot: APIbase now watches its own
providers, opens incidents on its own, and escalates anything involving money to a human.

Catalogue — measured live at release time

	v1.5.0	v1.6.0
tools	393	1,322
providers	114	389 (376 with at least one live tool)

Two provider counts appear above because the project uses two definitions and publishes both rather than picking the flattering one: 389 is every provider in the catalogue, 376 is those with at least one tool currently live. The site's own surfaces use 376.

Payment paths both green at release: pre-funded balance and x402 (USDC on Base, facilitator
reachable at 231 ms).

New — Autopilot

A deterministic health-and-incident system. It costs nothing to run while everything is healthy: the
incident engine makes zero model calls, and the only place a language model is involved is
classifying provider emails, capped at three calls a day.

• Incident lifecycle — OPEN → REMEDIATION_QUEUED → VERIFYING → RESOLVED, with
  WAITING_HUMAN as a terminal-until-a-person-acts state.
• Deduplication is a database constraint, not a convention: a unique index over open incidents
  means the same outage cannot be worked twice in parallel.
• Money is never automated. A PAYMENT_REQUIRED incident goes straight to WAITING_HUMAN and
  writes an operator file. There is no automatic branch that spends anything.
• Provider email is treated as data, not instruction — classified into a fixed set of outcomes
  with a schema check, never executed.
• Public endpoint /api/v1/incidents and a rebuilt /dashboard.

Honest notes

The autopilot only started running on 2026-09-04. The code had been merged over the preceding
days, but its scheduled jobs were never installed — the engine's own header said the schedule was
not installed by the script, and that line waited for a human who missed it. For that window the
dashboard displayed autopilot: OK, which did not mean "no problems" but "never measured".
The schedule was installed on 2026-09-04, and the first real run immediately opened 13 incidents,
including a SEV1 provider outage. 17 incidents are tracked as of publication.

Directly because of that, the dashboard no longer renders absent data as healthy:

111 of 389 providers currently report health unknown — not yet measured. They are shown as
unknown, not as green. 235 are green and 43 are degraded.

Also in this release

• /api/v1/incidents mounted (it was declared and returning 404 while a parity check reported green)
• /dashboard rebuilt and shipped English-only
• All repository access moved to SSH keys after anonymous HTTPS fetches were found to be silently
  IP-rate-limited (public repositories never issue the challenge that triggers credential lookup)
• Guides section and archive on the site