Savings accounting
The estimated transformation had three separate defects, each found by pulling on the previous one.
- Premium delegation no longer erodes the headline. The subagent pool netted cheap and premium delegation together, so choosing a more capable model read as a routing regression.
- The sidechain pool is normalized to a signed per-session rate. It was a raw window sum, so a heavy delegation day moved a monthly figure by hundreds of dollars.
- Every pool is now priced at its real window value, and short one-shot sessions get their own pool instead of being dropped. The main "now" arm no longer picks up rolled-up subagent tokens.
Each arm is traceable to one source with one inclusion rule.
Subscription runway
New card on the savings and trends surfaces for rate-limited plans, where a dollar figure is the wrong unit. Reports how much more work fits inside the 5-hour and weekly limits, with the current state beside the counterfactual. Window state is measured; the comparison is estimated and labelled as such. Renders nothing when the meter is missing or stale.
Cross-platform
session_log.platformis recorded in all three ingest paths and backfilled. It was NULL for every row, so per-platform reporting was impossible.- The estimated tier is blocked on non-Claude runtimes rather than applying Claude-shaped pricing to different billing. Copilot meters premium requests, not tokens.
Version drift
TOKEN_OPTIMIZER_VERSION was hardcoded and had fallen four releases behind, so fresh installs rendered a stale version in the dashboard header. It now derives from the plugin manifest. Reported by @danikdanik in #96, who also suggested the generalized fix.
Docs
Corrected counts, thresholds and env vars across the docs site, removed an unsupported benchmark claim, and added a CI gate that parses ground truth from source so these numbers cannot drift again.
Other
- The installer announces
TOKEN_OPTIMIZER_SKIP_VERIFYon stderr at both skip points. - Dashboard population metrics separate main-work from delegated sessions.
- The lean-output nudge fires from 20% context fill and is tunable.
283 tests.