Problem
estimateLowWorthRecoverableTokens returns the entire session token total when editTurns === 0 (src/optimize.ts:1971). That per-session number feeds tokensSaved, the headline "Potential savings: X tokens ($Y, ~Z% of spend)" (src/optimize.ts:2557-2563), and the JSON potentialSavingsCostUSD / per-finding estimatedSavingsUSD. The finding text hedges (review candidate, not proof of waste), but the number asserts the whole session cost is savable. An exploration-heavy user sees an upper bound presented as a point estimate, and it inflates the setup-health picture.
The retry path already does this right: it applies a bounded recovery fraction. Only the no-edit branch claims 100%.
Evidence
- src/optimize.ts:1967-1978, specifically :1971
if (editTurns === 0) return tokens
- Headline roll-up: src/optimize.ts:2557-2563; JSON: :2619-2650
- Verified on main at 1a75484.
Proposed fix
Either apply a conservative recovery fraction to the no-edit case (mirroring the retry branch), or exclude low-worth candidates from the summed dollar headline entirely and present them as a review list without a claimed dollar figure. Realized-savings measurement (act report) makes honesty here cheap to keep.
Problem
estimateLowWorthRecoverableTokensreturns the entire session token total wheneditTurns === 0(src/optimize.ts:1971). That per-session number feedstokensSaved, the headline "Potential savings:X tokens ($Y, ~Z% of spend)" (src/optimize.ts:2557-2563), and the JSONpotentialSavingsCostUSD/ per-findingestimatedSavingsUSD. The finding text hedges (review candidate, not proof of waste), but the number asserts the whole session cost is savable. An exploration-heavy user sees an upper bound presented as a point estimate, and it inflates the setup-health picture.The retry path already does this right: it applies a bounded recovery fraction. Only the no-edit branch claims 100%.
Evidence
if (editTurns === 0) return tokensProposed fix
Either apply a conservative recovery fraction to the no-edit case (mirroring the retry branch), or exclude low-worth candidates from the summed dollar headline entirely and present them as a review list without a claimed dollar figure. Realized-savings measurement (act report) makes honesty here cheap to keep.