Honey Eco

Surface the numbers the carbon badge is based on. Never estimate by hand — run the committed report so the figure matches hooks/eco.js (the verified EcoLogits v0.8.2 port) exactly.

Do

Run, from the repo root — zero-dependency, reuses the same hooks/eco.js the badge uses:

node scripts/eco-session.js              # newest session
node scripts/eco-session.js --transcript PATH --mode full

Then report terse — output tokens, CO₂eq (g), and the saved line — in a line or two. Keep the exact numbers the script prints; don't round away precision.

Two numbers, two epistemic statuses — never merge them

• CO₂eq is measured for this session: real output tokens through the EcoLogits port.
• saved is not. It is a modelled counterfactual — how many tokens a no-Honey run

would have emitted — taken from a committed bench stamp. Always pass through the basis: line the script prints (stamp, model, n, p). Never state a savings figure without it, and never re-derive one by hand.

When the script prints saved : —, no committed stamp covers this session's model. Say that. Do not extrapolate from another model's ratio — that is exactly how a tool ends up reporting savings while the invoice goes up.

For the authoritative breakdown (usage + embodied + primary energy) the user can run the Python path, which needs the real package:

pip install ecologits && python scripts/eco_report.py

Say once, don't bury

The params are speculative (Anthropic discloses none). The badge shows served impact — EcoLogits' single-stream (batch-1) ceiling divided by serving_concurrency (continuous-batching amortization); eco-session.js/eco_report.py also print the raw ceiling. State that once, briefly, then give the number. The badge renders only in Claude Code (its transcript is all Claude models).

Boundaries

Report only — don't change config or mode (that's the /honey command). The number is a range, not a meter reading; present it as one.