v0.2.14

levers.md: when a tool must retrieve, return the top few reranked results; search small, read around hits (Iusztin and Labonne, LLM Engineer's Handbook, sec. 97 and 219).

v0.2.13

Budget-fit reads (the biggest part that fits) and fewer tools in view, from Berryman and Ziegler, Prompt Engineering for LLMs (2024). See skills/tokenwright/CHANGELOG.md C-20261102-1.

v0.2.12

Habits: hand over deltas, keep a state capsule, quiet test and build runners, start effort low, a stop rule with an escalation point, a subagent brief that names its output and proof (R-20261101-2).

v0.2.11

The hook meter now counts an image Read at its token cost (about 1.6k) instead of the length of its base64; one screenshot used to put ~170k tokens on the meter.

v0.2.10

Docs: mention a line range instead of a whole file (@file#5-10), queue follow-ups while Claude works, a formatter and failures-only lint hook, @file load costs, LSP plugins, search over RAG for code. Output price ratio corrected to exactly 5x.

v0.2.9

When subagents pay: a Subagents section in references/levers.md (total tokens rise, cold start and cache, when delegation wins and loses, briefing and return size), a sharper Delegate big reads habit. Docs only.

v0.2.8

Fix: tw.py score counts a repeated tool record once, so a transcript that repeats a tool_result no longer lists one sink twice.

0.2.7

• tw.py meter and a hook meter note every 25k tokens of tool output (hooks on).
• tw.py score --compare N: this session against the best and median of the newest N past sessions in the same project.
• Back to a green size budget: the server.log eval fixture is generated, and the skill descriptions are shorter.

v0.2.6

AI Engineering habits (Edges, not middle; Memory by frequency) and a Result caching note in levers.md.

v0.2.5

tw.py score gains the debug-loop kind: three failed calls in a row to one tool, calibrated at about p99 of 300 sessions.