Release 0.14

Release 0.14

Released 2026-09-30. A contribution guard checks the AI policy of a project before Claude writes to it, a context note starts at 100k tokens, and a model lock covers effort levels. A version in parentheses marks a later patch of this line, and an entry with no version comes from 0.14.0.

Added

• Contribution guard (in the Bash guard, see Contributions): covers git commit, git push, and gh pr, gh issue, and gh discussion writes. It also covers gh api writes and GraphQL mutations that post content.
  • It denies a contribution to a project that forbids AI work. The catalog comes from open-source-ai-contribution-policies by melissawm (183 projects, 101 forbid).
  • It asks before a push or write to a GitHub repository whose owner is not the gh login of the user. The prompt shows the catalog entry of the project, or says that its policy is possibly unwritten.
  • A local git commit in the clone of another owner passes.
• scripts/update-ai-policies.mjs: updates the catalog into the plugin data directory.
  • At session start, a detached process checks the upstream README hash at most once a day.
  • The guard reads only the stored hash, and tells the user to update when it differs from the catalog.
  • DOTCLAUDE_OFFLINE=1 turns the check off.
• contribute-upstream skill: reads the AI policy of the project, stops when the project forbids AI work, and treats a missing policy as unknown. It verifies the claim before a draft and writes the draft in plain English for the user to send. Claude does not reply in the thread after that unless the user asks.
• Settings profile: adds permissions.ask rules for gh pr create, gh pr comment, gh pr review, gh issue create, gh issue comment, gh discussion create, and gh discussion comment.
• Context note (usage_notes option): when the main context is 100k tokens or more, each prompt that the user types tells Claude the context size and tells it to start a handoff with write-session-handoff.
  • No hook input gives Claude its context size.
  • With the settings profile, Claude Code compacts at about 117k tokens (median 121k in 145 measured compactions), before the 150k handoff point. The note at 100k gives Claude time to write a handoff first.
  • In a long run with no typed prompt, the first tool call past 100k gives the note once. It comes again after the context goes under 100k and back over it.
  • The size comes from the end of the transcript.
  • Release 0.15 changed when the note asks for a handoff.

Changed

• Model lock (0.14.1): denies a subagent or a claude command at an effort level that dotclaude does not support for its model.
  • Opus 5.5 supports low to xhigh.
  • Sonnet 5.5 supports only low and medium, because work that needs high needs judgment, and Opus 5.5 gives more for the same cost there.
  • The deny reason names where the effort came from and what to do.
  • EFFORT_LEVELS in hooks/lib/_models.mjs holds the table.
• System prompt, approvals: approval also covers actions that spend money or speak for the user. A request says what the action does, why, what it changes, and how to undo it. Open questions go in one AskUserQuestion call. Claude does not ask in text for an approval that a permission prompt gives, because many small approvals teach the user to approve without reading.
• System prompt, gaps: Claude does not call an unsupported case "intended" or a correct skip only because the code does not handle it. It checks public implementations and docs, and reports a gap unless the project or the user excludes the case. In one session an agent called four controllers correct skips although public drivers for them exist.
• apply-settings-profile: runs all previews and asks one question for the profile groups, the switches, and the extras. The permission prompt of each --apply is the approval of the write, with no second question in text.
• Usage notes: give the reset time of the session and weekly limits. At 90%, Claude writes a handoff when the remaining work does not fit before the limit, and tells the user the reset time.
• Hook speed on Bun:
  • The Edit guard reads the transcript only when an edit removes assertions (a long session: 35 to 17 ms for each edit).
  • The Bash and Edit guards find the latest prompt of the user from the end of the transcript and stop there. In a long session, an edit that removes assertions took 26 ms, not 40 ms.
  • The contribution catalog and the settings stamp hash with Bun.CryptoHasher, not node:crypto (a Bash check: about 3 ms less).
  • The status line measures width with Bun.stripANSI, not node:util (about 2 ms less).
  • The secret scan starts Betterleaks with Bun.spawn, not node:child_process, which uses about 15% less CPU after a Bash call.
  • The contribution guard reads the gh login from hosts.yml, and starts gh only when it cannot read the file. A push to the repository of another owner took 37 ms, not 91 ms.
• Dossier: fixes the cost comparison of Sonnet 5.5 and Opus 5.5 from the Artificial Analysis suite. diff-reviewer stays on Sonnet 5.5, so a risk: high slice still gets two different models as reviewers.
• docs/models.md (0.14.1): listed the supported effort levels for each model and the limits of the check. 0.20.0 removed it, and Plans and models now has the effort levels. Model Fit adds the ProjectArchitect Bench A/B as reported evidence, with its limits.

Fixed

• Search guard: it denied rg -uu, fd -I, and git grep --no-index into large gitignored notes or docs. It now denies an explicit bypass only when it walks a build or dependency directory, such as node_modules or dist.
• rm -rf "$TMPDIR/name": asked for approval. A named child of $TMPDIR now passes, and $TMPDIR itself, a glob, and .. still ask.
• Inline code: python3 -c "s.replace('os.unlink(', …)" gave a destructive-code warning for text inside a string literal. The check now ignores string literals.
• Stop gate: it counted an edit of .gitignore, .git/, .claude/, or a gitignored file as a code edit, and then asked for a check run. It now counts only files that git does not ignore, in both the Edit and Bash paths.

Previous: Release 0.13 · Next: Release 0.15