Release 0.8

Release 0.8

This page records the notes of the 0.8 release, with its patches 0.8.1 and 0.8.2.

Released 2026-09-28, in three patches: 0.8.0, 0.8.1, and 0.8.2. This release moves the engineering and git rules from the output style into a replacement system prompt. A version in parentheses marks a patch. Each patch is dated 2026-09-28.

Version	Change
0.8.0	Replacement prompt, launcher function, instruction-file checks
0.8.1	Writes to Claude Code settings ask the user in auto mode
0.8.2	Agents stop searching gitignored directories

Note: After you update to 0.8.0, run /dotclaude:apply-settings-profile again to install or refresh the launcher.

Added

• Replacement system prompt (0.8.0): replaces the built-in prompt of Claude Code. It holds the engineering and git rules. It names the installed Claude Code version. The output style keeps only how Claude talks and reports. The output style shrinks from about 15 KB to about 2.3 KB.
  • Only a CLI flag can replace that prompt. /dotclaude:apply-settings-profile installs a claude shell function that passes --system-prompt-file. It supports zsh, bash, fish, and PowerShell on macOS, Linux, and Windows.
  • The function adds nothing when you pass your own system prompt flag or set DOTCLAUDE_SYSTEM_PROMPT=0.
  • Session start keeps its prompt copy current after plugin and Claude Code updates.
• Missing-prompt notice (0.8.0): session start tells you when a session runs without the dotclaude system prompt. The launcher may be missing. An IDE or another program may have started Claude Code without the shell function. Such a session has only the rules of the output style.
  • With ANTHROPIC_BASE_URL set, the notice explains how to run a proxy such as Headroom with the function. Use headroom proxy and an exported ANTHROPIC_BASE_URL, not headroom wrap claude.
  • Set DOTCLAUDE_SYSTEM_PROMPT=0 to run without the prompt and without the notice.
  • 0.8.1: when a session starts without the claude function but the startup file already has it. The notice names source <file> or a new terminal as the fix. This happens when the terminal opened before the launcher was installed.
• Instruction-file checks (0.8.0): session start checks every CLAUDE.md, CLAUDE.local.md, AGENTS.md, and .claude/rules/ file. It also checks every file that they @import, without block-level HTML comments.
  • One file warns above 150 lines and fails above 200.
  • The text that loads at session start warns above about 3,000 tokens and fails above 5,000.
  • It also reports a file above 4 MiB, an import more than four hops deep, and an AGENTS.md that a CLAUDE.md hides without @AGENTS.md.
  • Names that link to one file (CLAUDE.md, AGENTS.md, GEMINI.md) count once, under the path of the original. A symlink to a missing file gets its own warning.
• git_attribution option (0.8.0, on by default): the includeGitInstructions: false setting of the profile makes Claude Code drop its Co-Authored-By commit trailer and pull request footer. The session notes now give them back. The attribution setting of Claude Code still changes or removes them.
• Ask on settings writes (0.8.1): writes to the own configuration of Claude Code ask the user. In auto mode, the classifier denied them as self-modification even when the user asked. So the 0.8.0 migration could not finish. A hook ask skips the classifier and shows a permission prompt (checked in Claude Code 2.1.283).
  • The Bash guard asks for the apply-settings, apply-claude-md, apply-launcher, and install-managed scripts of this skill with --apply. It also asks for commands that write a .claude/settings*.json or managed-settings file. It follows the bash_guard option.
  • The edit guard already asked for settings edits. It now also covers managed-settings.d/*.json.
• install-managed from Claude Code (0.8.1): scripts/askpass.sh asks for the admin password in a desktop dialog through sudo -A, after the prompt of the Bash guard. It uses osascript on macOS, and zenity, kdialog, or ssh-askpass on Linux. Without a dialog, the user runs the command in a terminal as before.
• Ignored-directory guard (0.8.2): the Bash guard denies recursive searches and listings that would walk gitignored directories, such as build output, dependencies, and caches. An agent's grep -r in a repository with an 8.9 GB .build/ directory hung. It would have flooded the context with generated files.
  • It covers grep -r, find, tree, ls -R, ack, rg and ag with ignore-bypass flags (--no-ignore, -u), fd -I and -u, and git grep --no-index.
  • It asks git which ignored directories are under the search path. A walk passes when there are none. A walk also passes when it starts inside an ignored directory, or when it is at most two levels deep. A walk passes when it excludes each ignored directory (--exclude-dir, -prune, -g '!…', -E, tree -I, including {a,b} lists).
  • The deny message names the directories. It points to rg, fd, or git grep, which skip ignored files.
  • It follows the bash_guard option.
• CLAUDE_CODE_GLOB_NO_IGNORE (0.8.2): the settings profile sets it to "false", so the Glob tool skips gitignored files as the Grep tool already does. Claude Code 2.1.283 runs Glob as rg --files --no-ignore unless this variable is false. In a test repository, **/*.swift returned files under .build/ and node_modules/ without it. With it, the same search returned only the tracked file. The variable comes from the Claude Code source and has no documentation. So a later release may change it. Run apply-settings again to add it.
• AGENTS.md (0.8.2): gives coding agents the commands and rules of the repository. CLAUDE.md and GEMINI.md are symlinks to it.
• Prompt surface doc (0.8.0): docs/claude-code-prompt-surface.md, now part of Prompt, records the text of the lean prompt that CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT selects. It also records the conditions for each part.
• Two letters that list gaps for xsyetopz/skills (0.8.0), both local files of the maintainer that were never in the repository:
  • docs/letter-to-skills-developer.txt lists dossier gaps that belong in skills.
  • docs/letter-to-skills-developer-skill-lint.txt gives that repository a skill lint gate and the fixes that it needs first.

Changed

• ASD-STE100 (0.8.0): the output style, the agents, the skills, the option descriptions, and the hook messages that Claude reads follow ASD-STE100. They also follow the current prompting guides of Anthropic. They use no semicolons, short sentences, and active voice with a named actor. They use no phrasal verbs or metaphor, one word per action, and a reason with each rule. No rule changed meaning.

• Size tests (0.8.0): dotclaude's tests check its own files against token and line limits. They replace the 21.5 KB byte cap.

  File	Warning	Failure
  Output style	1,000 tokens	2,000 tokens
  System prompt	5,000 tokens	15,000 tokens
  Each agent body	2,000 tokens	5,000 tokens
  Each SKILL.md	450 lines, 4,500 body tokens	500 lines, 5,000 body tokens

  Skill names and descriptions stay inside the Agent Skills limits of 64 and 1,024 characters.

• New system-prompt rules from the dossier reports (0.8.0):

  • Apply a correction to every similar case.
  • Make sure that the edited file is the file that runs.
  • Count a bug test only after it fails without the fix.
  • A mock of the part under test cannot catch its defect.
  • Review the diff of the whole task.
  • Do not remove a feature to pass a test.
  • Repeat a search or fix only when you expect new evidence.
  • Open a search hit before you rely on it.
  • Revisit an earlier decision that proved wrong.
  • Do not turn prose guidance into tests or rule files.
  • Wait for background jobs with Monitor.
  • A priority order settles conflicts between rules.

• Eval prompts (0.8.0): mention files as bare @file, which Claude Code expands. The earlier @`file` form did not expand.

• apply-settings cleanup (0.8.1): removes exact entries that older dotclaude profiles wrote and 0.8 dropped. These are the AskUserQuestion deny, the two Codex allow rules, and ANTHROPIC_DEFAULT_SONNET_MODEL=claude-opus-5-5. The preview lists each one. The same keys with other values stay.

• CLAUDE_CODE_FORK_SUBAGENT (0.8.1): the settings profile sets "0" instead of "false". Claude Code reads both as off. The settings JSON schema accepts only "0" and "1".

• Dossier (0.8.2): docs/dossier.md replaces the five research docs. It indexes six parts under docs/dossier/ that hold the design decisions and the measurements behind them. The dossier is now the evidence pages of this wiki, from Design to Open Items. The README is about half as long.

• Old releases (0.8.2): releases before 0.8.0 moved from the changelog to docs/changelog/. They are now the Release History pages. Each Markdown file now has 300 lines or less.

Fixed

• Fork test (0.8.0): the subagent test for forks no longer fails when the shell that runs it sets CLAUDE_CODE_FORK_SUBAGENT=false.
• Skill names in code fences (0.8.1): a /dotclaude: skill name inside a fenced code block no longer counts as an invocation. A pasted session-start notice in a code fence made Claude ask the user to send /dotclaude:apply-settings-profile again.

Related Pages

• Previous: Release 0.7
• Next: Release 0.9
• Archive: the index of the archived pages.