Release 0.11

Release 0.11

This page records the notes of the 0.11 release.

Released 2026-09-29. Subagents moved to Sonnet 5.5, and the context bounds got tighter. Cost data from 2026-09-28 and 2026-09-29 drove the new bounds. A version in parentheses marks a later patch of this line. An entry with no version comes from 0.11.0.

Note: This line requires Claude Code 2.1.284 or later. Claude Code 2.1.284 is the first release with Sonnet 5.5. After you update, run claude update if needed. Restart Claude Code. Then run /dotclaude:apply-settings-profile again.

Added

• Version notice: session start tells the user when Claude Code is older than 2.1.284. It reads claude --version of the running binary and says to run claude update.

Changed

• Main compaction point: the settings profile compacts at 150k tokens, not 200k. The system prompt, the session notes, and the status line use the same handoff point. From 2026-09-28 to 2026-09-29, main-conversation calls past 150k were 13% of the cost.
• Subagent bound: Claude Code refuses a subagent tool call past 100k tokens of context, not 150k.
  • With the 150k bound, 34 of 69 implementer runs still passed 100k. Subagent calls from 100k to 150k were 9% of the cost.
  • code-reviewer, security-reviewer, and plan-reviewer keep 150k. A split review loses the view across files, and it writes that view to the cache again.
  • The injected context budget and the subagent status line row show the bound of each agent.
• implementer slices: the system prompt and the implementer description ask for one small slice for each implementer. Larger work goes to new agents.
• Sonnet 5.5 (claude-sonnet-5-5): replaces Sonnet 5 everywhere, at the same prices.
  • This covers the implementer, docs-writer, and mechanical-worker agents. It covers the profile's CLAUDE_CODE_SUBAGENT_MODEL and availableModels. It also covers the managed drop-in and the default allowed_models.
  • Applying the profile again removes claude-sonnet-5 from availableModels.
  • A user who set allowed_models must replace claude-sonnet-5 with claude-sonnet-5-5.
• mechanical-worker: runs at medium effort, not low. Anthropic recalibrated the effort levels of Sonnet 5.5 and gives medium as the start for well-specified agentic coding. At low, the agent sometimes reported a change as done with no check.
• test-runner: runs on Haiku 4.5, not Sonnet 5 at low effort. It runs one command and copies the failure lines. That task needs no judgment. Haiku costs half as much per token. It starts without CLAUDE.md. It starts without the CodeGraph MCP tool. Its report copies each error line exactly.
• Sonnet 5.5 reminder (0.11.0, 0.11.1): the reminder for implementer, docs-writer, and mechanical-worker has more rules.
  • A real check must come before a report that a code change is done. A syntax-only check, or a command that did not start, does not count.
  • The agent writes only in the files that the brief names. It puts scratch files in the system temp folder. It reports defects outside the brief instead of fixing them.
  • In one user test of 35 bug-fix tasks, both models fixed 34. Sonnet 5.5 wrote outside its assigned folder 4 times and Opus 5.5 0 times.
• Model docs (0.11.1): docs/models.md and the dossier page that is now Usage Evidence gave this reason. Version 0.20.0 removed docs/models.md. They also give the reported cost data. Sonnet 5.5 costs less for each task than Opus 5.5 only at low and medium effort. This is why the Sonnet 5.5 agents stay at medium.

Fixed

• CHANGELOG link (0.11.1): the [unreleased] compare link at the end of the CHANGELOG starts from the latest release, not 0.8.1. just bump now moves it to the new version.

Related Pages

• Previous: Release 0.10
• Next: Release 0.12
• Archive: the index of the archived pages.