Cache read is charged per API round-trip on the whole conversation so far, so a
long session's cost is driven by TURN COUNT, not message length. One measured
session spent 267M cache-read against 856k output — a ratio of 290:1. The report
priced tokens but never counted the turns spending them.
Across the 263 stored transcripts this now shows 33,508 tool turns at an 8%
batching rate, of which 4,437 (13%) decided nothing: polling a counter, running
a linter, checking whether a push landed. At 606k average context re-read per
turn that is ~2.7B cache-read tokens on turns a script could have taken.
- turns.py classifies a command as poll / lint_test / vcs_check / substantive,
deliberately conservative: state-changing commands (git commit, push) count as
substantive even though they are scriptable, so the reported waste is a floor.
A turn is only mechanical when EVERY command in it is.
- Batching is grouped by requestId, not by transcript line: one API response can
span several lines, and counting per line reports every turn as single-tool
(0% batching where the real rate is 12%).
- records.py gains TurnEfficiency (schema v3, additive — old records read as
zeros) with derived rates written into to_dict so readers need not recompute.
- report.py gains a "Turn efficiency" section pricing each mechanical kind.
- _TurnTracker owns both the totals and the open turn, after pylint flagged
_Accumulator over its attribute budget; _Accumulator.turns is now a property,
so every existing caller is unchanged.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The transcript analyzer behind /compile-candidate lived only in the working
tree. Untracked meant unlinted — pre-commit only inspects files handed to it —
so nothing had ever checked it, and it had been abandoned mid-refactor with
every parse raising AttributeError: '_Accumulator' object has no attribute
'first_timestamp'.
The accumulator had been moved onto the grouped records dataclasses, but three
increment sites and parse_session still read flat attributes that no longer
existed. parse_session now hands over acc.meta and acc.counts directly; only
subagent_count comes from elsewhere, since no transcript line carries it.
Two modules broke the 500-line limit, so they were split along seams that were
already there. signatures.py takes the text-signature helpers — traces.py was
importing them from parse.py, making parse an accidental hub, and now neither
imports the other. stats.py takes the pure number and time helpers. parse.py
521 -> 458, detectors.py 559 -> 467.
The structural lint findings were refactored, not configured away.
_counter_candidates took eight arguments; the five naming which counter field to
mine and at what thresholds became a frozen _CounterSpec, so both call sites are
one line. _extract_skill carried in_span and a line budget as loop variables
across thirteen branches; a _SpanBuffer owns that state now, leaving the walk at
four. record() took twelve flat kwargs and now takes one override per record
block, with session_meta() so a partial meta override keeps the standard
timestamp — a bare SessionMeta(slug=...) would have nulled started_at and
quietly changed the per-week arithmetic. All 36 call sites rewritten by an AST
pass. tests/factories.py moved into conftest.py, where every other package here
keeps shared builders and where name-tests-test does not object.
965 tests pass at 100% statement and branch coverage, every hook is green, and
no suppression was added anywhere. `candidates` and `traces skill-finish` both
run against the real corpus.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
