[hardy30894]

What

secondwind eval replays every recorded Claude Code session through the optimizer and reports, per tool and per command (Bash:cat, Bash:grep |, Bash:git log, Read, ...):

• results kept inline, offloaded, or left alone
• how often the model later needed what an offload hid
• raw, sent and handed-back tokens, and net savings
• unfaithful: blocks that did not come back intact

--blocks <path> writes one line per result with its byte size, the context the model held, and the turns after it, so policies can be compared on held-out sessions.

Fix in the trace reader

A tool result's content array was serialized as raw JSON: text results carried a wrapper the model never sees, and images and PDFs carried their base64 as text. On the author's 1,491 sessions that made tool output look like 739M tokens when the text the model read is 101M. The reader now joins text blocks and replaces an image or document with a short placeholder. This also corrects cache-savings and the audit, which read traces the same way.

How a reopen is counted

A block counts as reopened when a token only it introduced, and its preview hid, appears in a later assistant message or tool call before any other result or user message supplies that token again. A grep followed by a read of the same file charges the read, not the grep. Net assumes a reopened block is fetched back whole, the costliest case.

Faithfulness

Checked per block, independently of the optimizer's own gate: an inline wire must decode to the certified content; an offload must resolve from the store to the original, byte-identical text or the same JSON value. On the author's corpus: 0 of 138,209 blocks unfaithful.

Tests

Unit tests for reopen attribution (including the re-supply rule), the command classifier on heredocs, loops, quoted programs, git -C and timeout prefixes, faithfulness across JSON, log and code shapes, and the reader's text/image handling; an end-to-end test runs secondwind eval on a synthetic recorded session. cargo test --workspace: 354 passed. Clippy clean.