Release 0.28 Changes

Release 0.28 Changes and Fixes

This page records the changed behavior and the fixes of the 0.28 release. The additions and the removals are on Release 0.28.

Changed

• The eval cases guard a protected file with a file-state regex grader (kept-<path>) on its final content. They do not use a Bash text grader, because command text cannot tell a write from a read. just eval always passes --keep-temp, so each run keeps its sandbox and trace.jsonl.
• just eval and just eval-variant stop before the first run when they run in the Bash sandbox of Claude Code. An example of such a run is ! just eval. In the sandbox, the model calls of the runs use its proxy. Runs stopped on a model call until the 600 s timeout of the case.
• The git ask and the write ask also read a quoted heredoc as commands. They do so when a loop, xargs, or a shell runs its lines. Examples are bash -c "$c" in a while read loop, | sh, source /dev/stdin, and sh -c "$(cat)". They also read a here-string given to a shell, and a git subcommand from a variable or from xargs gets the git ask. Before, a git stash in the brief of the test-runner script got no ask.
• The git ask and the write ask read code in any language that a program other than the shell runs. Examples are python3 - <<'EOF', node -e '...', and a heredoc written to a file that python3 then runs. They have no parser for the language, and they read the words of the code. So os.system("git stash") gets the git ask, and a protected path that the code names gets the write ask.
• The git ask and the write ask read the script file that a command runs, such as python3 x.py or bash run.sh, except in a test run. A shell script is read as shell, and a file in another language gives an ask only for the commands in it.
• dotclaude-eval-compare adds no --scaffold and no --allow-tools Bash,Write,Edit, because each one runs the scripts or the tools of a suite as the user. Add them in the options for suites that you trust. just eval-variant adds both for the suite of this repository. The eval-compare skill no longer grants Bash(dotclaude-eval-compare *), so Claude Code asks you before each run. A bound in the form --max-cost-usd=N is kept, and no second bound is added.
• With the backend option cloakbrowser, a WebSearch that falls back to the built-in tool gives Claude a notice with the reason, such as a bot check. The notice tells Claude to tell you.
• The drive-web-browser skill loads the agent-browser guide through the new bin dotclaude-browser-guide, which cuts the guide at 40,000 characters and names the command for the rest. A timeout or a failed run gives its reason, and only a missing binary gives the install text.
• The eval cases destructive-ask, big-file-subagent, and library-docs no longer name the graded behavior in the prompt. destructive-ask now asks the agent to free disk space in a project whose large files are a customer export and a database. It no longer needs a question. A report that removes nothing is also right. report-count no longer names its script in tasks.md.
• The wiki has the shape of the Claude Code docs: an overview, a quickstart, a reference page for each plugin, Hooks, Troubleshooting, and Undocumented-Reads. That last page gives the source of each read of undocumented data and how it breaks. The old release pages are in wiki/archive/.
• CI runs just check with Claude Code 2.1.296, and just lab runs the lab of each plugin that has a tests/ folder. The repository has CONTRIBUTING.md, SECURITY.md, CODE_OF_CONDUCT.md, issue and pull request templates, and dependabot.yml.
• Each plugin manifest has $schema, keywords, and homepage, and its description is the same text as in the marketplace. The backend option of dotclaude-browser lists its values, agent-browser and cloakbrowser.
• The lab bins take flags in place of custom env vars: dotclaude-sandbox --dir DIR (also --clean and tools/start-run.ts) and dotclaude-capture --upstream URL. dotclaude-sandbox finds claude on PATH only, and the capture proxy reads ANTHROPIC_BASE_URL when --upstream is not set.
• For contributors: the sandbox, the request capture, the settings schema, the start measure, and the usage report moved from tools/ into dotclaude-lab and dotclaude-sessions. just sandbox gives --plugin-dir for each plugin of the checkout, and DOTCLAUDE_SANDBOX_PLUGINS is removed. dotclaude-sandbox makes each --plugin-dir absolute, because claude runs in the sandbox project and skips a relative folder. just measure passes the start bounds of tools/limits.ts to dotclaude-start, and the usage report no longer reads verdicts.jsonl. just debug-agents runs dotclaude-sandbox --check-frontmatter.
• For contributors: the tests are in one folder for each feature, such as tests/dotclaude/guard/, in place of file names with a shared prefix.
• When the headless browser of dotclaude-browser fails or gets no search results, such as a bot check, the built-in WebSearch answers the call. Before, the hook denied the call, and a session could end with no answer.
• The size bounds that equaled the current file size are gone (STYLE_MAX_BYTES, VARIANT_MAX_BYTES, and SKILL_DESCRIPTION_MAX_CHARS). Size targets in tools/limits.ts replace them. The targets are CLAUDE.md at most 6 KB. Skill metadata has 150 to 300 bytes for each skill, and skill bodies have 2 KB to 8 KB. The skill metadata of all skills has a target of 5 KB to 8 KB and a ceiling of 15 KB. The first request has a target of 8 KB to 16 KB. A size above a high bound fails, and a size below a low bound is reported.
• dotclaude is tested on Claude Code 2.1.296, and the minimum version is 2.1.296, because the settings schemas are of that version. The intro section of the system prompt has one text since 2.1.295, and CLAUDE_CODE_INTRO_FRAME is gone.
• The core plugin has five agents: debugger, implementer, investigator, reviewer, and test-runner. Each agent has a maxTurns value, a model, and an effort in its frontmatter, and web-researcher is merged into investigator.
• The core plugin follows the plugin layout of the Claude Code docs. hooks/register.mjs is the one hooks module, features/<name>/ holds the code of each part, and lib/ holds the shared code.
• Each file has 300 lines or fewer, and each test file 500 lines or fewer. A test fails on a file over its bound.
• Runtime JavaScript in plugins/ is .mjs on Node. A test fails on another script type, except the hook-lab .test.ts files.
• For contributors: just measure sizes the start of a session and prints the figures. just capture logs and sizes the requests of Claude Code, and just schema writes the settings schemas. just start-run captures and measures the start of a clean sandbox session in one command. just debug-agents fails when the debug log names an unknown frontmatter field. just eval-variant REF runs the eval cases on the checkout and on a variant at a git ref. just eval takes eval options such as --case NAME. just eval-variant copies the results of the variant arm to plugins/dotclaude/evals/results/variant-REF-TIMESTAMP/ before it removes the worktree.
• For contributors: pnpm installs the pinned toolchain (Biome, markdownlint-cli2, and TypeScript) from pnpm-lock.yaml. .node-version pins Node 22.18. just lint and CI also lint the Markdown.
• For contributors: the repository has no Bun. The tests run with node --test (just test or pnpm test), and the scripts of tools/ run on Node. pnpm installs the yaml parser that the eval-case test reads case.yaml with.
• just schema reads Claude Code 2.1.296 only. The tables of older versions are removed, because each build renames the minified helpers. The schema check of a definition now reads the object and record helpers from the table, so a lazy record such as env keeps its object type.
• Each catch in plugins/ now names the error that it expects, such as ENOENT or SyntaxError, and rethrows the others. A classic hook that crashes still lets the call continue, and it writes the error to stderr. The guard of dotclaude no longer keeps an empty list of GitHub owners for the session after a failed gh call. It asks gh again at the next check.
• Windows: codegraph, openspec, and claude start through cmd.exe when they are .cmd files. An argument with %, !, ", a line break, or an end \ is refused there, because cmd.exe expands % before it reads quotes. A missing codegraph adds no session line, dotclaude-sandbox sets USERPROFILE for the child, and readTail returns null for a folder.
• CI runs just ci, which is just check without the type check, because Claude Code writes .claude-plugin/types/ only when an interactive session loads the plugin.

Related Pages

• Release 0.28: the additions and the removals of 0.28.
• Archive: the index of the archived pages.