What Snapback covers: 32 failure families, and what agents ask for

An agent can fail in a thousand ways, but they cluster. Snapback organises failures into a taxonomy — behavioural failures on one axis, infrastructure failures on the other — so a cryptic error becomes a named family with a verified fix. Here's the map.

Two kinds of failure

Behavioural failures are about how the agent reasons and acts: it loops, hallucinates a tool schema, selects the wrong tool, contradicts itself, or acts on a truncated output. Infrastructure failures are the cryptic errors from the systems the agent touches — a chain error, an API quota, an expired certificate, a rebalanced queue. Both get diagnosed; the second is where the depth is.

The 46 infrastructure families

Each family is a cluster of real, documented error signatures with a verified fix — sourced from official docs, specs, and SDKs. Grouped by domain:

Try the free one right now

The infrastructure diagnosis tool — diagnose_infra_error — is free and needs no token. A known error returns from the curated library with no LLM call, instantly; a novel one falls back to an LLM-assisted diagnosis (slower, never auto-applied). Hand it a cryptic error string and it returns the family, root cause, and fix:

diagnose_infra_error({error: "34651 SEGMENT SELL FAILURE could not sell segment"})
→ { family: "travel-gds",
    root_cause: "availability changed between search and book",
    fix: "re-price the offer, don't retry the stale one" }

What agents actually ask for — and how feedback grows the library

The library isn't static. It grows from what real agents hit and what they tell us. Two feedback channels drive that:

• submit_feedback — after a verdict, an agent (or developer) rates whether the diagnosis was correct. This tunes confidence and surfaces patterns that need refining.
• request_pattern — when Snapback couldn't solve something, an agent asks us to add it. Requested patterns move through a review queue; check status with get_request_status.

The pattern in the requests is telling: agents don't ask for more behavioural coverage — they ask for the gnarly infrastructure errors that have a single correct fix the base model keeps missing. That's why the infrastructure families go deepest, and why recent expansions have been in financial rails, messaging, auth, and travel — the domains where a wrong retry is expensive.

An agent can also see its own contribution: my_impact shows how your feedback and requests shaped the shared library. It makes the library a commons — every agent that reports a miss makes the next agent's diagnosis better.