Daily Edition Sources +3

Builder Work: Pi Counts the Work Its Agent Hides

Pi's latest accounting patch treats compaction, branch summaries, and tool-result hooks as model work with receipts—not invisible overhead beneath an agent's final answer.

A rough marker ledger diagram shows branch summaries, compaction, and tool results flowing into one Pi session usage total, with a caveat that tested totals are not invoices.
Diagram Punkcount the model work that happens between visible turns.
repo earendil-works/pi evidence
3 source signals 1 repo commit 2fd3868
Evidence: commit 2fd3868 / July 21, 2026 / Daily Edition
Open Edition Evidence below

“What did this agent cost?” is an easy question to answer badly. A visible assistant reply may be preceded by a branch summary, a compaction pass, or a tool hook that called a model of its own. Pi's public coding-agent README says its maintainer regularly publishes Pi monorepo work sessions; its new usage-accounting patch shows the corresponding engineering discipline: carry the receipt with every operation that spent model work.

The repo evidence

The change reaches across the agent harness and coding-agent layer rather than adding a single counter. A branch-summary result now retains response usage. Compaction returns both summary text and usage; in the split-turn path, Pi combines the history-summary and turn-prefix receipts. Tool-result hooks can preserve usage supplied by a tool. Session entries retain it, and the coding-agent session code rolls it into totals. The accompanying tests check the combined record, while the patch's coding-agent tests follow the numbers into session statistics.

The transferable skill

Make cost and usage part of the event contract, not a dashboard calculation. When a system creates a summary, compacts history, calls a tool, or lets an extension patch a result, attach typed usage right there and aggregate it at the session boundary. That preserves a traceable answer when a user asks why a short response used more work than it appears to contain.

The limit

The public session statement is evidence that coding agents are part of Pi's work, not proof that an agent authored this exact commit. The tests show intended accounting behavior, not a cloud-provider invoice or universal extension support. The next useful signal is whether provider and extension adapters consistently return the metadata this ledger expects when a real session crosses multiple models.

Evidence Trail

Receipts below the story

The article above is the public narrative. This section keeps the source trail and limits on the same page.

Edition
DateJuly 21, 2026
LaneDaily Edition
Confidence78%
Sources3
Reposearendil-works/pi

Primary Evidence

Evidence Limits

  • The public session statement does not attribute this exact patch to a coding agent. The repository tests prove intended accounting behavior, not provider billing, deployment, or universal extension support.
Letters & Corrections

Send a note to the desk

Corrections, missing context, or a follow-up lead.