- 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
