Statusline and savings
This is the statusline of a session inside a PA3 project, taken during phase 3.10.6 of PA3's own build.
Every x/y pair reads the same way: white is what was used, blue is what the helpers saved, in the same unit.
1%/2% of 5h means this session used 1 % of your five-hour allowance, and the work its helpers kept out of your
context would have cost another 2 %. The blue figures after them (66% fewer, lasts 2.9x longer, 2.5w) are
savings too. Green, yellow and red judge a reading against your window (green is comfortable), and magenta is the
context your session is carrying.
ProjectArchitect · 3.11 runs in another session: the project and where its phase stands. In the router's own
session this line reads like ProjectArchitect · Phase 3.10.6 · 4/9 (tasks done out of the plan's total), with the
running task and the next ones beneath it; here another session owned phase 3.11.
Opus 5.5 · medium · 114k/1M · 5h 5% ↻ 1h41m · 7d 21% ↻ 5d07h · fable 32%: the session's model and effort, the
context it carries (114k tokens of a 1M window, with a bar), and your plan's usage straight from Claude Code: 5 % of
the five-hour window, which resets in 1 h 41 min; 21 % of the seven-day window, which resets in 5 days 7 hours; and
32 % of the separate weekly Fable limit.
Pace: 0.9x · 13%/day used · 15%/day left · 47% fewer tokens than vanilla · lasts 1.9x longer: how you are
spending the week. 0.9x means you use the seven-day window at 0.9 times the even pace that would last exactly to
the reset. 13%/day used is your average so far; 15%/day left is what you can spend each day from now and still
last to the reset. 47% fewer tokens than vanilla is your account's saving this week: saved over paid plus saved.
lasts 1.9x longer is the same saving as a multiplier: the week's allowance goes 1.9 times as far.
Session: 1%/2% of 5h · 2%/4% of 7d · 66% fewer · lasts 2.9x longer: this session. It used 1 % of the
five-hour window and its helpers saved 2 %; of the seven-day window, 2 % used and 4 % saved. Its work used 66 % fewer
tokens than it would have without the helpers.
Project: 5%/2% of 5h · 21%/19% of 7d · month ~1.6w/2.5w · lifetime ~1.6w/2.5w · 62% fewer · lasts 2.6x longer · 2 accounts:
this project, as your account sees it: 5 % of the five-hour window used and 2 % saved; 21 % of the week used and 19
% saved. month ~1.6w/2.5w counts weeks of allowance: this billing month the project used about 1.6 weeks' worth and
its helpers saved 2.5 weeks' worth; lifetime is the same since the project began. 62% fewer and
lasts 2.6x longer cover the project's whole life, every account included, and 2 accounts have worked on it.
Account: 5%/~2% of 5h · 21%/~18% of 7d · month ~1.2w/1.4w · lifetime ~1.2w/1.4w · 55% fewer · lasts 2.2x longer:
the same figures for your whole account, every project together.
⚑ waiting on you: REVIEW.md: an alert. The router is waiting on you, here for a review. Model fallbacks and a
low disk show here too. The last line, with auto mode, is Claude Code's own.
A ~ marks a figure converted through a fitted rate: a saved share on the Account line, or weeks of allowance. It
turns into ≈ when the rate is less certain: the fit is only a band, or more than a tenth of the dollars behind a
weeks figure have no rate yet. month — means PA3 does not know yet which day your plan renews; it asks once (the
Claude app shows the date under Settings → Billing), and the month figures count from that day. Every figure is
defined in
statusline-definitions.md.
PA3 shows two kinds of numbers: what was counted, straight from the API's usage figures, and what was modeled from those counts under a stated assumption.
What is counted. A helper (a retriever, a coder, an expert) reads material its parent never has to see. For
each helper run the ledger counts the tokens kept out of the parent's context: the API-measured growth of the
helper's context, times the share of that growth that was tool results, minus the answer it returned. Scripts that
read or edit a file on an agent's behalf add a credit of their own: the carry of the file they read or wrote, plus
the text an Edit would have had to send, net of what the script printed.
The saving you see: the carry model. A token that entered the parent's context would have been re-read with
every later request of the session. The carry model prices that: each kept-out token is re-read at the cache-read
price on every later request, written once at the one-hour cache-write price and again after a cold gap, inside
one 1M-token window that starts over when it would overflow; the helper's own starting context is subtracted. The
definition's version is vanilla-net-v4, and pa_ledger.py report names it beside the kept-out counts. This is
the figure behind every blue value on the statusline. The one assumption is the 1M window, so the report also walks
the same figure at 450k and 200k.
The replay: a second, separate model. pa_ledger.py report also replays every request of your account as if
it had all been one monolithic session (pa/replay.py), under stated assumptions about the cache lifetime, the
starting context and the handoffs. It answers a different question: what the same work would have cost without the
harness's shape.
Why the two are never added. The replay is a counterfactual whose assumptions can be tuned, and folding it into the headline would let the harness flatter itself. Its monolithic context also already contains the kept-out tokens, so adding the two would count them twice (spec decision D39). They are shown apart, and they check each other: over the author's billing period up to 2026-09-25, the per-run figure came to $1,254 and the replay to $1,093.
Dollars and allowance. Dollars are API list prices from pa/prices.py, applied to every request the ledger
records. On a subscription they are a yardstick, not a bill, so the statusline converts them into a share of your
five-hour and seven-day allowance through a rate fitted to your own account's usage, each week at its own rate.
What no figure counts. PA3's install turns off Claude Code's Task tools and its session recap. Both save tokens, but the ledger has nothing to count them from, so no figure includes them: savings the ledger can't see.