Release 0.3

Release 0.3

Released 2026-09-26. Limits usage and cuts prompt size. Caps effort and workflow size, replaces the heavy CodeGraph prompt hook with a light one, and replaces the built-in Codex instructions with dotclaude templates. This line requires Claude Code v2.1.283 or later.

To update:

1. Restart Claude Code.
2. Run /dotclaude:apply-settings-profile again.
3. If you use the Codex agents, run /dotclaude:setup-integrations codex again.

Added

• Effort cap: the settings profile sets maxEffortLevel: "xhigh". The session, agents, skills, and workflow stages cannot use max. The claude.ai effort picker warns that max uses about 5.5x the usage on Opus 5.5 and 3.5x on Fable 5.1.
• Workflow bounds: the settings profile sets two values.
  • CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS=6, a hard cap per run.
  • workflowSizeGuideline: "medium", advice to aim for fewer than 10 agents.
• $schema: the settings profile has it.
• install-managed.mjs: an optional lock that you run with sudo in your own terminal.
  • It writes managed-settings.d/50-dotclaude.json with the effort cap, fast mode off, and the model list.
  • It never touches managed-settings.json.
  • It checks the file before it moves the file into place, and asks before it replaces a different file.
  • It keeps backups outside the drop-in directory.
• Skills anywhere in a message: /dotclaude:<skill> works anywhere in a message. When the command is not at the start, a UserPromptSubmit hook injects the skill with the rest of the message as $ARGUMENTS. For a skill that must run as a command, the hook tells Claude how to start it.
• PostModelSwitch hook: gives the Fable 5.1 note to a session that switches to Fable, and retracts it on a switch away. The session-start note alone missed those sessions, because the model field is not always present there.
• Codex prompt replacement: /dotclaude:setup-integrations codex builds three model catalogs. They are dotclaude-catalog-interactive.json, dotclaude-catalog-worker.json, and dotclaude-catalog-review.json.
  • The model list comes from codex debug models, run under a temporary Codex home that links to your login. A catalog stops Codex from refreshing its own list, so this method lets a new run of setup find new models. When the request fails, models_cache.json is the fallback.
  • For GPT-6 Astra, Sol, and Luna, dotclaude's templates replace the built-in instructions of 18 to 21 KB. The persistent-mode and multi-agent role text is removed.
  • config.toml and each dotclaude profile point at their own catalog through model_catalog_json.
  • Both profiles set include_collaboration_mode_instructions = false and include_apps_instructions = false.

Changed

• CodeGraph prompt context: a light dotclaude hook, on with codegraph_hint, replaces codegraph prompt-hook. That command injects up to 9 KB of explored source on any prompt with a word such as "how" or "call", including task notifications and subagent hand-backs.
  • For prompts you type, the hook looks up the code-shaped names in them (camelCase, snake_case, name(, a.b) with codegraph query. It lists only the names that the index has, as name, kind, and file:line. Claude calls codegraph_explore for the source when it needs it.
  • The setup skill tells you to remove the global codegraph prompt-hook entry that codegraph install adds.
• web-researcher: runs on Opus 5.5 at low effort, with maxTurns 60, instead of Haiku 4.5.
  • It backs each "exists" or "does not exist" claim with the source line that it read.
  • It gets raw pages instead of WebFetch summaries.
  • It reports as soon as the question has an answer.
• Skill loading: Claude can load recap, challenge, blind-spots, and lessons-learned when you ask for what they do. polish, troubleshoot, fresh-eyes, review-code-changes, and apply-settings-profile still run only when you name them.
• Output style:
  • Claude answers a correction with the right fact or the changed action. The quoted banned phrase is gone, because quoting it primed it.
  • The line before the first tool call names what Claude does to your task, not how tools or skills load.
  • Claude ends a turn with a question only when the answer changes the next step.
  • New guidance covers a subagent that stopped at its turn limit, and cheap agentType and effort choices in workflow scripts.
  • The style is no longer than before.
• Staleness notice: the session-start notice also checks for the effort cap. It looks at user, project, and local settings, and at managed settings and their drop-ins.
• Codex worker profile: drops compact_prompt, which has no effect with remote compaction on GPT-6. Its developer_instructions now match the worker template: Claude lists assumptions instead of stopping to ask.
• verify-before-stop gate: counts these commands as check runs.
  • bun run validate (and the validate scripts of other package managers)
  • claude plugin validate
  • markdownlint and markdownlint-cli2
• Codex load check: runs codex -p <profile> debug prompt-input for both profiles.
• Test command: bun test ./tests/ replaces bun test tests/ in package.json and CI. The bare form is a path filter that also found tests under ignored directories.
• Markdown lint: the Markdown passes markdownlint, and the shared rules moved to .markdownlint.jsonc.
  • Prompt files in agents/, skills/, output-styles/, and evals/ are exempt from the 80-column limit and keep their step numbers.
  • Two skills had a code fence directly after an XML tag, which Markdown read as part of the tag. A blank line now separates them.

Removed

• codegraph_prompt_context option: it ran codegraph prompt-hook on typed prompts. The codegraph_hint hook replaces it.
• CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION: removed from the settings profile. Claude Code removed it in v2.1.224, so it has no effect. A new run of the profile removes it from settings that an earlier version wrote it to.

Fixed

• Stop gate: it counted any quoted string in inline interpreter code as an edited file. A regex literal, or a file that the script only read, counted as an edit and caused "Code changed … no test ran". Now the gate records only the path argument of an actual write call, and paths after a cd resolve against that directory.
• Guard tests: they failed inside a session with the dotclaude profile applied, because ANTHROPIC_DEFAULT_SONNET_MODEL leaked into alias resolution. The tests now clear the model alias variables.

Previous: Release 0.1 and 0.2 · Next: Release 0.4