Retry taxonomy

Vendor dashboards usually answer “how many tokens?” RetrySight classifies how those tokens were spent, including retries that never look like a successful completion.

The visual taxonomy on Product is the canonical illustration. The classes below are what collectors try to emit.

Edits and compaction

Edit loops and context compaction burn tokens without visible output. Compaction re-ingests prior conversation when the window resets.

Test and diff outcomes

Test failures, diff rejected, and diff accepted — the retry patterns behind wasted spend.

Subagents and sessions

Subagent loops, session boundaries, and task aliasing across collector restarts.

Token usage

Input, output, cache read/write, and reasoning tokens by model, user, and source tool. Fleet cost views are on Cloud (/cost) and Enterprise Analytics+.

Collector health

Agent heartbeats and stale-collector alerts so reports do not silently go quiet. GraphQL subscriptions include agentHeartbeat.

Evaluations

Eval definitions, runs, annotation queue, and playground — Analytics+. A/B eval experiments — Integration+.

Why it matters

Seat fees are the obvious line item. Retry loops, rejected diffs, and compaction often dominate the rest. A short FinOps framing is in the blog post The hidden cost of coding agent retries.