Release 0.27

Release 0.27

Released 2026-10-07.

0.27.0 is a breaking release. The core plugin is built again from an empty tree on the documented extension points of Claude Code 2.1.292: a settings profile, a forced output style, one hooks module, agents, and skills. The Design page gives the source of each decision.

Migration from 0.26

• Install Node.js 22.18 or later, because the scripts and hooks no longer run on Bun.
• Update, restart Claude Code, and run /dotclaude:setup. Setup removes includeGitInstructions: false, autoCompactWindow, advisorModel, the 0.26 subagentStatusLine and its stub, and the 0.26 block in <config dir>/CLAUDE.md, because they work against the new profile. It makes a backup of each file before it writes.
• The plugin options are gone, so the /config entries of dotclaude have no effect.
• Handoff notes from 0.26 in .claude/handoffs/ stay readable, and no hook writes new ones.

Added

• Compaction is off: autoCompactEnabled: false and DISABLE_COMPACT=1, which also turns off /compact.
• The context window is 300K tokens by CLAUDE_CODE_MAX_CONTEXT_TOKENS=300000, which Claude Code honors only together with DISABLE_COMPACT. At the limit the session stops, and the user runs /clear. CLAUDE_CODE_AUTO_COMPACT_WINDOW=300000 makes /context and the context warnings use the same window. Without it, Claude Code 2.1.292 uses 200K for Opus 5.5 there.
• Auto memory is off: autoMemoryEnabled: false.
• The forced dotclaude output style, with keep-coding-instructions: false, together with CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=1. It replaces the built-in coding instructions and the 0.26 spec injection, and the prompt cache holds it.
• The sandbox is on, with network access only to GitHub and the package registries.
• templates/managed-settings.json for users who want the compaction, memory, and sandbox keys enforced. Setup prints the copy command and does not run it, because it needs sudo. With it, the built-in guard sec-default loads and skips the context_management text of the module.
• OpenSpec is a first-class integration. Setup checks the openspec CLI, offers the install and openspec init --tools claude, and the status line shows the active change with its task count.
• The profile turns off the tools that it does not use, so that their schemas leave each request: enableArtifact: false, enableWorkflows: false, and deny rules for ScheduleWakeup and ReportFindings. /code-review then gives its findings as text. In the sandbox, /context at the start of a session fell from 32k to 8.1k tokens, with the trimmed modder plugin.
• The hook lab: just lab runs claude plugin test plugins/dotclaude, which loads the hooks module and fires stubbed tool.check events.

Changed

• Permissions ask only before dangerous commands, with native rules and no Bash parser: force pushes, history rewrites, piped shells, sudo, publishes, public gh writes, and SQL drops. Read-only git and the usual build and test commands are allowed.
• One hooks module replaces the 0.26 command hooks. It asks before a recursive rm outside the project, and before the first call to a GitHub repository of another owner with an AI policy. It keeps a deny of the engine and fails open. With DISABLE_COMPACT set, its prompt.section handler replaces the context_management section of the system prompt, which says that the system summarizes prior messages, with text that says compaction is off. The built-in guard sec-default skips this handler on a machine with managed settings or a Team or Enterprise login.
• In auto mode, the classifier can allow an ask of the module, so one classic PreToolUse hook, hooks/auto-mode-guard.mjs, asks again for the same calls. It runs only in auto mode, only for commands that its if filters match, and needs Node.js.
• The scripts and hooks of the core and dotclaude-jev plugins run on Node.js 22.18 or later, not on Bun. Setup changes a bun status line command to node. The um command of dotclaude-modder stays on Bun, because it uses the YAML parser of Bun. The allowed-tools of the second-opinion skill allow only node with jev.mjs, not each bun command. The drive-web-browser skill and the Browser page give npm install -g agent-browser.
• A prompt audit against Opus 5.5 removed two numeric reply caps from the output style and the reviewer agent, and the Reason: line of the old spec format from the Jev note. The modder skill says that a /clear also removes what is not in MODLOG.md.
• /dotclaude:handoff is user-invoked only. It writes a note that primes the next session: the goal, the constraints, the decisions, the rejected approaches, and the proof.
• maxEffortLevel is xhigh, and the profile sets no effort level, so each model keeps its default.
• The status line is one file, and it shows the context against the 300K window, yellow at 75% and red at 90%.
• The add-on plugins work without the core plugin. dotclaude-browser is skill-only, and dotclaude-jev gives its own session note.
• dotclaude-modder has one skill, mod-any-game, and no session note. The 8 other skills are step files in references/ that Claude reads only at their step, so the plugin adds one skill listing to each session, not nine. The rules of the session note are in the skill.
• includeGitInstructions stays at its default, and the profile sets attribution.commit and attribution.pr.

Removed

• The operating spec, its SessionStart and SubagentStart injection, and lib/terms.mjs.
• The compaction fork, the cold-cache note, and each hook that tells the user to run /clear.
• The Bash, edit, secret-redaction, spawn, and attribution guards, and the dependency on betterleaks.
• The CodeGraph augment, the sembr rewrap, the long-run note, the setup notice, and the Stop prompt hook.
• The Ponytail integration and the minimal-code rule. The output style has its own rule to write the least code that does the task.
• All plugin options, the Concise output style, and the contribute skill.
• The agents translator, docs-writer, digest-writer, eval-designer, fuzz-engineer, infra-engineer, and reverse-engineer.
• The evals in plugins/dotclaude/evals/ and tools/compaction-report.mjs.
• advisorModel and CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS. The advisor tool adds its instructions to each request, and each advisor call reads the whole context again.
• The reverse-engineer agent of dotclaude-modder. The reverse-engineer-binary skill in xsyetopz/skills replaces it.