Links sessions to todo notes (note_claim.sh, todo-images prompts,
fuzzy first prompt >= 0.5, handoff chains), dates completions from
the status HLC in the first Firebase dump, and adds usd_per_note,
note_linked_usd_share, notes_done(_linked) to the snapshot, a WEEKLY
section and notes.json. token_dashboard stores the build-class A/B
(agent-ab-build.jsonl) as meta ab_build.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WzxdWWMeapxnDcgk7HSw67
- Subagents that outlived /clear land in a new session's subagents/ dir with
no meta: borrow the spawning session's agent-<id>.meta.json, else the first
prompt line. Clears the $35.93 '?' row; no launch lacked a description.
- bash_tokens_per_call rose 41% with Bash tokens/day flat (cheap calls fell
28%). Compare bash_tokens_per_day and bash_big_result_share (results >2k)
instead; the Large-Bash lever keys on the share. Tool counters grouped in
ToolPayloads.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TjyHmRrjZssaxXDWcGuCeQ
Counting transcript records reports 1.00 tool calls per message no matter
what the model did: the harness writes one JSONL record per tool_use block,
all sharing message.id. A hand-rolled per-record count read the real 1.9%
batching rate as 0.0%, which nearly shipped as a headline finding.
Group by message.id instead, and report it, so the number comes from a
tested code path with the window handled by find_transcripts rather than an
ad-hoc timestamp. Three snapshot fields give next week's run a baseline.
Axes was at the 12-attribute cap, so the counters live in their own Batching
dataclass with the two ratios as properties.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NiSSDu4gKbp9GGxTTtXvYW
Answers "what is billed on every turn regardless of use" by measurement
instead of recall, which is the part earlier audits kept getting wrong.
A surface is reported dead only when unused in BOTH windows. One earlier
audit's seven-day dead list included six servers that had been used within
thirty days, so a single window is not evidence. Shares are weighted with
cache-read weight 0.1, so a cached surface is not billed as though it were
re-read in full each turn.
Verified by running it, not just by tests: the report renders over 44,311
turns, putting rules/ at 4,907 tokens (1.77%), MEMORY.md at 4,312 (1.55%),
memories/ at 3,758 (1.35%), skill descriptions at 3,279 (1.18%) and CLAUDE.md
at 2,057 (0.74%), with per-MCP 7d/30d call counts behind each verdict.
The files were found untracked in the tree from an earlier session; every
module is already under the 250-line cap, so nothing needed splitting, and
the package was already at 100% branch coverage, so nothing needed new tests.
This lands the instrument only -- no surface is unregistered on its say-so.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
android_ui, app_icons, token_audit, focus_policy and random_jpg are now
entirely under 250 lines. Only wsg_grabber's six test modules remain in
python_pkg.
android_ui/driver.py 411 -> 205 took two cuts and a class split: the
element and error types plus the XML parser to _elements.py, and the
text-entry methods to a TextEntryMixin, using the ABC-with-abstract-
methods shape because the tests call them as bound attributes.
app_icons/glyphs.py 291 -> 93 by moving the twelve path fragments out.
Three intermediate failures, all module-level globals that travelled
with their code (_logger, re, _SAME_FIELD_X_TOLERANCE_PX) - the same
class of miss as _make_verifier, which ruff cannot flag.
Counts unchanged: 81, 137, 14, 44, 57.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011w2Pj4Ek3ZNMQVYcFHgXDY
