Skip to content

Archive 0.28

Jump to bottom
xsyetopz edited this page Oct 13, 2026 · 2 revisions

Archive: 0.27 and 0.28 Text

This page keeps the text that described behavior of the 0.27 and 0.28 releases. Current pages do not describe these parts. The Release History page has the notes of each older release.

The 0.27 and 0.28 Rebuilds

0.27.0 rebuilt the core plugin from an empty tree and removed most 0.26 parts. 0.28.0 removed the setup profile with its diff and backups, the router table, the usage state file, the handoff notes, and the dotclaude-modder plugin. The settings snippet replaced the setup profile.

Update Notes for 0.28.0

  • Files that 0.27.0 wrote stay on your machine, such as the stub statusline.mjs and the <file>.dotclaude-backup-<time> files.
  • The statusLine command of 0.27.0 runs the stub, which shows nothing. Replace the command with the path of the installed plugin folder, and remove the stub and the backups by hand.

Usage Notes

  • The setup profile of 0.20 to 0.27 denied Agent(general-purpose).
  • The 0.27.0 profile set cleanupPeriodDays to 14, so Claude Code removed old transcripts.
  • 0.28.0 restored the search context in the add-on dotclaude-codegraph, with a PreToolUse hook on Grep|Glob. Claude Code 2.1.296 has no Grep or Glob tool, so that hook is gone.

Removed Browser Note

The recognize-captcha skill and the offline ddddocr-rs OCR no longer exist in the browser plugin. The CHANGELOG removed them in 0.20.0, because CloakBrowser avoids most CAPTCHAs.

The CLI Against the LSP Tool

Task 17.1 of the change dotclaude-0-28-0 compared the CLI with the built-in LSP tool. The reason was the rule that the plugin does not add runtime code for a feature that Claude Code has. The user decided that the plugin ships whatever the comparison shows.

Setup:

  • The model was claude-sonnet-5-5, in a headless sandbox run (plugins/dotclaude-lab/bin/dotclaude-sandbox) with no plugin of dotclaude.
  • The project was 9 Python files (shop/ and tests/) of about 80 lines.
  • The question was "If I change the signature of parse_amount, which functions are affected, directly or through other calls, and in which files?"
  • The true answer is 8 functions: Cart.add, load_prices, apply_discount, build_cart, summary, main, test_parse, and test_discount.
  • The LSP tool needs a language server. The run used pyright-langserver (installed on the machine) through a scratch plugin with a .lsp.json file, and ENABLE_LSP_TOOL=1. The run did not test the tool without the variable. The tool is deferred, so Claude finds it with ToolSearch first. A machine with no server for the language has no LSP tool result.
  • Input tokens are input_tokens, cache_read_input_tokens, and cache_creation_input_tokens, summed over the requests of the transcript. One run for each arm.
Arm Correct Tool calls Requests Input tokens Output tokens
No graph tool (Grep, Glob, Read) 8 of 8 9 5 55,752 1,003
CLI on PATH, index built, no instruction 8 of 8 4, none of codegraph 4 43,555 601
LSP tool, no instruction 8 of 8 11, none of LSP 4 43,750 1,152
CLI, told to use codegraph impact and callers 6 of 8 1 Bash 2 17,493 204
LSP, told to use the LSP tool 8 of 8 12 LSP, 1 ToolSearch, 2 Read 8 77,004 2,351

What the result showed:

  • With no instruction, Claude used neither tool. It read the files of the small project instead.
  • When told to use it, the CLI answered in one call and used 4.4 times fewer input tokens than the LSP tool. The LSP tool took one call for each hop of the call chain.
  • codegraph impact has a default depth of 2. It missed summary and main, which are three calls away. The SessionStart line gave -d 5 for this reason.
  • One run for each arm is not a rate.

Related Pages

Home | README | CHANGELOG | Issues

Clone this wiki locally