fix(always-on): the turn-1 prefix includes what it READ from cache

[pa-arth]

computeAlwaysOn sampled input + cacheWrite5m + cacheWrite1h from the first main-chain turn (src/alwaysOn.ts:212-214). That is the prefix only when the session opened cold. A session resuming in a project it has run in before finds the whole prefix already cached — it reads ~everything and writes ~nothing — so observed standing context came out smallest exactly where the standing context is largest.

The measurement

522 local sessions, main-chain turn 1, /subagents/ files excluded:

turn 1 reads from cache in	97.9% of sessions
median prefix, input + cacheWrite	40,272
median prefix, + cacheRead	59,645 (1.48x)

59,645 lands within 1.8% of the same quantity measured independently by differencing claude -p "ok" --output-format json runs with one context component removed at a time (60,738). Two methods sharing no code agreeing to 1.8% is what makes this a correction rather than a preference.

Why no test caught it

Every existing turn-1 fixture in audit.test.ts carries cache_read_input_tokens: 0 — the spawn-economics block asserts 11,000 and 17,500 on exactly that shape. The suite covered the cold case thoroughly and the warm case not at all, which is the 97.9% case. New test locks a warm prefix at 54,200 and asserts explicitly it is not 4,200, so a future edit dropping the term again lands there.

Related

The same defect shipped in the product: promptster-backend/packages/engine/src/jobs/costEfficiency.ts:591 had the identical formula, fixed in pa-arth/promptster-backend#713 (which also needs a rail term — its rows can be OpenAI, where cached tokens sit inside input; cc-audit is Claude-Code-only so it does not).

Still open, not addressed here

report.ts printed 31,365 for this number, where a faithful replication of the pre-fix formula over the same transcripts gives 40,272. So ~9k of the originally-observed gap is something else in the span/merge handling — mergeUsage folds streamed rows by per-field max, and the main-chain span selection is stricter than my replication. Worth chasing separately; this PR does not explain it.

Checks

src/__tests__/audit.test.ts 50/50 · full suite 386 passed, 1 failed — pricingDrift.test.ts, which fails identically on clean main (our pricing table vs litellm) and is unrelated · tsc --noEmit clean.

🤖 Generated with Claude Code

[greptile-apps]

Greptile Summary

This PR corrects the observed turn-1 standing-context measurement by including tokens read from cache, with a regression test for warm-cache sessions.

• Adds cacheRead to the first main-chain turn’s sampled prefix.
• Adds a warm-cache fixture asserting the complete 54,200-token prefix.
• Adds a pnpm lockfile without changing declared dependency constraints.

Confidence Score: 5/5

The PR appears safe to merge, with the corrected calculation matching the documented empirical turn-1-prefix contract.

The implementation adds the previously omitted cache-read component, and the new regression test covers the common warm-cache shape without revealing a reachable behavioral or security failure.

Important Files Changed

Filename	Overview
src/alwaysOn.ts	Includes cache-read tokens in the empirical first-turn prefix, aligning warm and cold cache representations without changing downstream pricing structure.
src/tests/audit.test.ts	Adds focused regression coverage demonstrating that a warm-cache prefix includes input, cache-write, and cache-read tokens.
pnpm-lock.yaml	Adds deterministic pnpm dependency resolution; no changed-code-reachable security issue was established.

_{Reviews (1): Last reviewed commit: "fix(always-on): the turn-1 prefix includ..." | Re-trigger Greptile}