Release 0.6

Release 0.6

Released 2026-09-27, in three patches: 0.6.0, 0.6.1, and 0.6.2. Cuts background usage and disk use, and makes agents finish with a report. 0.6.0 adds a justfile, CI hardening, and settings that lower background requests. 0.6.1 and 0.6.2 fix the goal loop guard and the turn-limit failures of codex-worker. A version in parentheses marks a patch.

Note: After you update to 0.6.0, run /dotclaude:apply-settings-profile again.

Added

• justfile (0.6.0): has the recipes test, lint, validate, check, and bump. just bump <major|minor|patch|X.Y.Z> sets the version in both manifests and dates the [Unreleased] CHANGELOG entries.
• StopFailure hook (0.6.2): shows a terminal notification when a turn stops on a usage limit.
  • The notification names the limit, its reset time from the usage that Claude Code caches, and claude --resume <session>.
  • The hook is part of the usage_notes option.
  • Claude Code shows it only in an interactive session, and only terminals that support OSC 9, 99, or 777 notifications show it.
• README notes (0.6.2):
  • The prompt cache TTL: CLAUDE_CODE_PROMPT_CACHE_TTL=1h for API keys, cloud providers, and usage credits (subscriptions already get an hour on the main conversation).
  • Claude Code relays your latest chat message to every workflow agent (#95369).
• Turn-budget hook (0.6.2): refuses every tool except SubagentHandback once about 5% of a dotclaude agent's maxTurns remains, with a minimum of 3 turns. The next action of the agent is then its report. The hook counts the API calls of the agent since its latest prompt, resume, or wake-up, as Claude Code does.
• codex-worker Bash hook (0.6.2): limits codex-worker to three actions: write its brief file, run run-codex.mjs, and read the report. Given the task text, Haiku edited the files itself with a shell heredoc and never started Codex. The prompt alone did not stop it.
• Optional integrations (0.6.0): /dotclaude:setup-integrations offers two.
  • tgrep: indexed text search. It installs the tool, indexes the project, and adds .tgrep/ to the global excludes file of git.
  • fast-compact: /fc, which trims old tool output with Jev. Setup uses the right Jev provider and model, and states dotclaude's measured trade-off first.
• Disk use controls (0.6.0):
  • dotclaude prunes its per-session hook state at session start when it is 30 days old. This matches the default transcript retention of Claude Code.
  • The scratchpad_prune_days option (off by default) removes Claude Code session scratchpads and loose entries under /tmp/claude-<uid> that nobody touched for that many days. It runs at session start and in the background, and it never prunes the current session.
  • For a one-time cleanup, run bun hooks/session-start/prune-scratchpads.mjs --days <n> from the plugin folder. It is a dry run unless you give --apply.

Changed

• CI (0.6.0): runs with a read-only token, actions pinned by commit, per-job timeouts, and cancelled superseded pull-request runs. It checks format and import order as well as lint, covers scripts/, and reports one all-green check to require.
• Background usage (0.6.0): the settings profile cuts it. Per the costs and prompt-caching docs of Claude Code, each of these settings sends a request that re-reads the conversation.
  • autoCompactWindow: 400000, or 200000 on Pro, Max 5x, and Team seats. The default on current models is about 967k, and every turn re-reads context up to that size.
  • promptSuggestionEnabled: false. Suggestions cost an extra request after every response.
  • awaySummaryEnabled: false. There is no automatic session recap when you step away, and /recap still works.
  • crossSessionInbound: "hold". Messages from your other sessions wait and do not start idle turns.
• Lean prompt (0.6.0): the profile sets CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=1, the lean built-in prompt of Claude Code. Claude Code already uses this prompt in some interactive sessions. It is about 6.6k tokens shorter per request, and the tools are unchanged. A one-word claude -p request measured 26,671 input tokens without it and 20,038 with it. Most of the cut is the long auto-memory instructions.
• /goal conditions (0.6.0): the output style asks for three parts. They are an end state that the output of Claude shows, how to check it, and a turn bound. Reason: the goal evaluator reads only the transcript.
• Temp cleanup (0.6.0): the output style and the subagent conventions ask Claude to remove used build output, clones, and large dumps from the temp folder. Nothing else removes them.
• Fable deny rule (0.6.0): the settings profile drops Agent(model:claude-fable*). Permission rules match the alias that the Agent tool sends (fable), never a full model ID, so that rule could not match.
• Frontmatter check (0.6.0): the edit guard denies a Write or Edit that would leave the YAML frontmatter of a Markdown file unparseable. An example is an unquoted description: Does a thing: then another. The denial names the parser error, so Claude quotes the value.
• Global CLAUDE.md block (0.6.0): names tgrep, scc, tokei, dasel, and mlr when they are installed. It gives an example only for tools that are present.
• Turn-budget hook (0.6.2): the hook and the notice that describes it follow the turn_limit_handoff option.
• code-reviewer (0.6.2): may use 60 turns instead of 40. Its median review used 36 turns, and half of the reviews since 0.5.0 hit the cap. A review that stops at its cap delivers no findings.
• codex-fanout (0.6.2): gives each worker a brief file instead of the task text. It checks the report file and git status of a stopped worker before it dispatches the item again.
• Bash guard (0.6.2): applies its Codex model and plan rules to the --model option of run-codex.mjs.

Fixed

• /goal loop guard (0.6.1): it never fired unless the goal's condition started with Goal:. Claude Code puts the condition itself in the feedback text ([@GOAL.md]: …). The guard now counts the goal_status records of the transcript.
• codex-worker out of turns (0.6.1, 0.6.2):
  • 0.6.1: Codex's workspace-write sandbox keeps .git, .agents, and .codex read-only, so workers used their 12 turns without an edit. codex-worker now stops before it starts Codex when a file in scope is under one of those folders, and codex-fanout keeps those files out of the queue.
  • 0.6.1: codex-worker no longer reads the files in scope or polls with sleep while Codex runs. It waits for the background completion notice.
  • 0.6.2: on 0.6.1, all 7 runs in one session stopped at the 12-turn limit with no report, and two finished only after the orchestrator continued them. Haiku polled with sleep and until loops, ran Codex in the foreground, and did the acceptance steps of the brief itself.
  • 0.6.2: the worker now runs one script, skills/codex-fanout/scripts/run-codex.mjs. The script checks the setup, runs Codex, and prints a report with the git state. It flags writes that the sandbox of Codex denied under .git, .agents, or .codex. The worker runs as a background agent, so a run longer than the Bash timeout moves to the background, and the worker waits for the completion notice.
• Agents out of turns (0.6.2): agents still ran into their turn limit without a report. Of 13 capped runs after 0.5.0 told agents their limit, none reported before it. The turn-budget hook above is the fix.
• Git settings that run a program (0.6.2): the Bash guard asked only about core.hooksPath. It now also asks when -c, --config-env, git config, or git environment variables (GIT_SSH_COMMAND, GIT_CONFIG_KEY_n, and others) set one of these keys:
  • core.fsmonitor, core.sshCommand, core.pager, or diff.external
  • a filter driver
  • a ! alias to a program
  • Values such as cat, less, true, and empty still pass.
• Agent prompts (0.6.2): they and the subagent conventions said that the final message reaches the caller. In auto mode only the SubagentHandback report reaches it, so they now say "report".

Previous: Release 0.5 · Next: Release 0.7