Versions and Upgrading

0.2.12 (2026-11-01)

• Docs: new habits hand over deltas (a diff, the changed facts) instead of snapshots, and keep a state capsule (HANDOFF.md). Extended: quiet modes for test and build runners, start effort low and raise it on a failed check, a stop rule that says when to ask the user, a subagent brief that names its output file and proof. Nothing to change when upgrading.

0.2.11 (2026-11-01)

• Fix: the hook meter counts an image Read at its token cost (about 1.6k) instead of the length of its base64, so one screenshot no longer adds ~170k to the meter. Nothing to change when upgrading.

0.2.10 (2026-11-01)

Docs only. New habits: mention a line range (@file#5-10) instead of a whole file, queue follow-ups while Claude works. levers.md gains a formatter and failures-only lint hook section, what an @file mention loads, LSP plugins, and search over RAG for code. Output price ratio stated as exactly 5x.

0.2.9 (2026-10-31)

Docs only. references/levers.md gains a Subagents section: delegating keeps the main context small but raises total tokens (cold start, no shared cache, 5-minute TTL); it pays on big reads with small answers and loses on small lookups and coupled work. Brief with a length cap; return a file path for big results.

0.2.8 (2026-10-31)

• Fix: tw.py score counts a repeated tool record once; a transcript that wrote one tool result twice no longer lists its sink twice. Nothing to change when upgrading.

0.2.7 (2026-10-31)

• tw.py meter and a hook meter note every 25k tokens of tool output. The post hook now also runs after Read and Glob.
• tw.py score --compare N: this session against the best and median of the newest N past sessions in the same project.
• Shorter skill descriptions; the eval log fixture is generated instead of stored. Nothing to change when upgrading.

0.2.6 (2026-10-31)

Two new habits (put must-follow rules at the edges of a long prompt, place a fact by how often it is needed) and a result-caching lever in levers.md, from AI Engineering. Nothing to change when upgrading.

0.2.5 (2026-10-31)

• tw.py score gains the debug-loop kind: three failed calls in a row to one tool.

0.2.4 (2026-10-30)

• tw.py score gains the poll kind: the third and later run of one exact shell command.

0.2.3 (2026-10-30)

Three new habits (cut the always-loaded load instead of leaning on compaction, let a check finish before the next edit, brief a subagent with the purpose and a size), a leverage ranking for fixes in levers.md and a shouted-rules check in tokenwright-skills, from Thinking in Systems. Nothing to change when upgrading.

0.2.2 (2026-10-30)

Six new habits in the catalogue (set the stop rule first, bisect instead of scanning, read the error you have, change approach after two misses, sample before a sweep, wait with a deadline), from The Pragmatic Programmer and Algorithms to Live By. tidy.py conflicts prints shared words in a stable order. Nothing to change when upgrading.

0.2.1 (2026-10-30)

Fixes the two known issues of 0.1.0: score counts a second whole read once, as a repeat read, and changed reports the right line count for a file that ends in a newline.

Upgrading from 0.2.0: update the plugin. Totals from score drop for sessions that read a big file whole more than once.

0.2.0 (2026-10-30)

skill-tidy is folded in as a fourth skill, tokenwright-catalog, and tw.py catalog forwards to its tidy.py. apply now keeps its backup beside the file when the state folder is read-only. The known issues of 0.1.0 are still open (fixed in 0.2.1).

Upgrading from 0.1.0: update the plugin. Nothing else changes.

Coming from skill-tidy: uninstall it, install tokenwright, then run this once so the parked-skills index points at the new script:

python tw.py catalog parked index

Parked skills and backups in ~/.skill-tidy keep working. With both plugins installed side by side, the descriptions loaded in every session came to about 2,400 characters; tokenwright 0.2.0 alone loads about 1,920.

0.1.0 (2026-10-30)

The first release: the three skills, tw.py with six commands, and the two hooks, off by default. Nothing to upgrade from.

Known issues in 0.1.0, fixed in 0.2.1:

• score counts a second whole read twice (FAQ).
• changed reports one line more than a file with a final newline has.

The full list of changes is in the CHANGELOG.