Release 0.13

Release 0.13

Released 2026-09-29. One hook process runs for each event, the run-agent-loop skill runs large changes as slices, and the browser skills moved to a separate plugin. Two upgrades need action: brew install betterleaks and /plugin install dotclaude-browser@dotclaude. A version in parentheses marks a later patch of this line, and an entry with no version comes from 0.13.0.

Added

• run-agent-loop skill: runs a large change as slices, the workflow of the Bun, GitHub Copilot, and pnpm v12 Rust ports. The main conversation writes .dotclaude/loop/GUIDE.md and slices.jsonl. Each slice goes to an implementer in a worktree, then to diff-reviewer, then to a fixer, and then to the frozen tests. A wave runs at most 5 slices at the same time.
  • diff-reviewer agent (Sonnet 5.5, medium effort, read-only): reviews one slice from its diff and GUIDE.md only, not from the report of the implementer.
  • Protected files: while .dotclaude/loop/loop.json lists protected globs, the edit and Bash guards deny a subagent change to a matching file and its removal. The main conversation has no limit.
  • Stop hook: sends Claude back once when a loop slice has the status implemented and no review.
  • Status line: shows the loop progress, for example loop 2/7.
• Bash guard, file follows: denies tail -f, tail -F, and inotifywait -m also in the background, unless timeout bounds it. A background follow ran until the session ended, also after its file was removed. The reason names an until grep -q loop and Monitor.
• Bash guard, background stdin: denies a background command that reads stdin to its end before it starts work, when the command has no stdin of its own.
  • The commands are codex exec, cat and python, python3, or node with no file, tr, and a shell with no script.
  • A background call usually gets /dev/null as stdin. In one session it got a pipe that did not close, and two codex exec … & runs waited on it for 81 minutes.
  • A </dev/null or 0< redirect, a heredoc, a pipe, or exec </dev/null lets the command through.
• Prompt tags: the agent and skill prompts use one set of XML tags in one order. The order is <task> (skills) or a role paragraph (agents), <context>, <inputs>, <constraints>, <procedure>, topic sections, <report_format> (agents) or <output_format> (skills), then <example>. The prompting guide of Anthropic asks for consistent, descriptive tag names, and gives no fixed set.

Changed

• Hook processes: hooks/dispatch.mjs runs all actions of an event in one bun process, at the same time, and merges their output. A Bash call starts 2 hook processes instead of 7, and uses about half the CPU time (164 ms to 86 ms, measured). An error in one action does not stop the others.
• Skill descriptions: shorter, because Claude Code lists them in every main turn.
  • Each is now 150 to 230 characters, from 190 to 540.
  • drive-web-browser has no when_to_use now, because its description holds the trigger.
  • A new limit in LIMITS fails the tests above 250 characters.
• Bash guard, ignored directories: denies a walk into a gitignored directory only when that directory has 200 entries or more. Small caches such as __pycache__ or xcuserdata no longer cause a deny. In the 6 days to 2026-09-29, this rule denied 85 commands, and the usual false deny was grep -rn X Sources Tests over a small cache.
• Secret redaction: uses Betterleaks instead of gitleaks.
  • It takes the same rules and the same GITLEAKS_CONFIG, and also reads BETTERLEAKS_CONFIG. Live validation stays off.
  • Upgrade: run brew install betterleaks. Until then, tool output is not redacted, and session start says so.
  • dotclaude does not use gitleaks any more. Keep it for pre-commit hooks.
• Stop gate: no longer blocks a pass claim when nothing was edited and nothing ran. It blocked read-only research agents that reported results that others ran. After a failed check, a claim in backticks, a code block, or a > line does not count as a pass claim.
• Open-task and announced-work checks: let a turn through when it ends with ExitPlanMode. The open-task check also lets a turn through when it ends with AskUserQuestion, and it does not apply to subagents.
• Task-completion gate: no longer writes task lines to verdicts.jsonl. The input check from 0.12.0 is closed.
• Agent(general-purpose) deny (0.13.0, 0.13.1): the settings profile denies it.
  • The deny removes the agent from the list that Claude sees (checked with claude -p).
  • In one week, Claude asked for it 244 times, and the hook of the plugin refused each call. The hook stays for sessions without the profile.
  • The user runs /dotclaude:apply-settings-profile again to add the deny. The skill lists the deny in its group table, so the user sees the reason before applying.
  • The Models row no longer names general-purpose as an agent on Sonnet 5.5.
• Usage report: counts the wake turns of each entrypoint. In the week to 2026-09-29, 527 of 532 wake turns were in cli sessions. 280 were messages from other sessions, and 125 of the 128 agent wakes were before 0.7.0 kept agents in the foreground.
• Browser plugin: browser automation and CAPTCHA OCR moved to the optional dotclaude-browser plugin in the same marketplace.
  • Sessions without it do not load the two skills or their session note. The session note of the plugin also replaces the browser line in the CLAUDE.md block of the settings profile.
  • Upgrade: run /plugin install dotclaude-browser@dotclaude. Then set cloakbrowser, cloakbrowser_humanize, cloakbrowser_headless, and captcha_ocr_ddddocr again in /config under dotclaude-browser. The old values under dotclaude have no effect.
  • /dotclaude:setup-integrations shows whether the plugin, agent-browser, CloakBrowser, and ddddocr are installed.
• Context and usage notes: ask for a handoff note with the write-session-handoff skill and then /clear, not /compact. A handoff keeps the facts that Claude chooses, and /compact costs a full turn over the large context. The system prompt of the settings profile says the same.
• Output style: has a ceiling of 900 tokens (warn at 700), not 2000, and measures 515 tokens.
• integration-setup: tells the agent to trust --help over memory. Haiku 4.5 has reliable knowledge only to Feb 2025, and the tools that it sets up are newer. A new test fails when an agent on Haiku 4.5 sets effort, because Haiku 4.5 does not support it and claude plugin validate accepts it.

Removed

• Three usage notes that did not fire in the transcripts from 2026-09-25 to 2026-09-29. They are the third-correction and refusal notes (note-rewind, 0 fires), the repeated-command note (note-repeated-status, 0 fires), and the expired-cache notice (note-stale-cache, 1 fire). The status line shows the cache expiry, and the follow guard and Monitor cover polling.

Fixed

• Task-completion gate: named no task, because it read task_name and task_status, which Claude Code does not send. It now reads task_id and task_subject, and the reason names the task.
• Docs: the README, the settings profile doc, and the design doc said that 3 subagents run at the same time. The limit is 5, as _budget.mjs sets.

Previous: Release 0.12 · Next: Release 0.14