Detectors

Each detector is an independent, individually toggleable plugin that reads a normalized session and emits zero or more findings — a named waste pattern with severity, wasted tokens, wasted dollars, an evidence trail, and a prescribed fix. Third-party detectors register the same way the built-in ones do, via a wattage.detectors entry point (see CONTRIBUTING.md for how to write one).

Detector	Detects	Basis	Quality risk
prefix_churn	Re-sent context not being cached	Measured	None
cache_gap	Caching attempted but under-redeemed by reads	Measured	None
nonconvergence	Agent loops thrashing/oscillating/stalling without progress	Measured	None
retry_storm	The same request re-sent back-to-back — a retry loop	Estimated	None
tool_result_bloat	Oversized tool results re-fed into later context	Estimated	Low
verbosity	Output over-generation with no cap	Estimated	Low
redundant_tool_calls	The same tool call repeated pointlessly	Estimated	None
retrieval_thrash	Repeated retrieval yielding no new evidence	Estimated	Review
model_mismatch	A premium model used for a step a cheaper one would handle	Estimated	Review
reasoning_overspend	Excess reasoning tokens on a simple step	Estimated	Review

Basis: measured vs estimated

Every finding declares how its dollar figure was derived, and the two tiers are treated differently on purpose:

• Measured — computed from tokens the provider actually billed, at the real rate card: a resent prefix, an unredeemed cache write, a non-convergent loop's real spend. These drive the Token Efficiency grade, the headline recoverable figure, and the CI gate.
• Estimated — a heuristic projection: chars÷4 token conversions, policy ceilings, hypothetical downgrade savings. These are reported in full (marked with ~) and totaled separately as "estimated findings" — but they never move the grade and can never fail a CI build. A guess, however useful, is not something to gate a merge on.

Quality risk, explained

Every finding is also tagged with a quality risk tier:

• None — the fix is purely a billing/caching change; it can't affect what the model actually sees or produces.
• Low — the fix nudges output format/length; a small, usually-safe behavior change.
• Review — the fix could plausibly change output quality (a cheaper model, less reasoning, less retrieval). These findings always show up in the report so you know about them, but they never count toward the Token Efficiency score unless you supply a --quality quality.json map showing the fix is actually safe. That's a hard rule, not a default that can drift: a cheap-but-wrong agent can't score well just because nothing was measured.

A shared limitation, stated once

Several detectors approximate signals that would be exact if message content were captured (prompt diffing, precise per-call token attribution) — most production traces don't capture full message content for privacy/size reasons, so Wattage falls back to token-count-based heuristics with explicit guards against the most obvious false positives. Each detector's page names its specific approximation and the guard that keeps it honest, rather than glossing over it.