Release 0.7

Release 0.7

This page records the notes of the 0.7 release.

Released 2026-09-28. This release cuts usage that a measured week of real sessions showed. It adds a context budget for subagents and makes subagents run in the foreground. It removes Codex support, the general workflow skills, and the CodeGraph hooks. Each of these cost more than it gave.

Note: The limits suit Pro and Plus and apply on every plan. Larger plans only reach the limits later. Usage Evidence has the measurements that docs/pro-first-usage.md held.

Added

• docs/pro-first-usage.md: where the Max 20x week of this machine went, what that means for Pro, and the Codex quota and dossier findings. It also gives the case for and against the GPT-5.6 models.
• Subagent context budget: once the context of an agent passes 150k tokens, the hook refuses its tool calls. Its next action is its report.
  • A fork starts with the context of its parent. So it gets 50k tokens of growth instead.
  • In the measured week, 83% of implementer cost and 85% of general-purpose cost came from calls above 100k tokens.
  • The hook is now enforce-agent-budget.mjs, still under the turn_limit_handoff option.
  • Subagents learn the limit at start.
• general-purpose refusal (part of subagent_guidance): refuses these agents and points to the dotclaude agent for the job. They were 17.5% of the measured week: 83 explicit calls, mostly implementation slices with no turn limit. With forks off, a call without subagent_type spawns general-purpose. So the hook refuses that call too.
• Foreground subagents: the same hook sets run_in_background: false. The settings profile sets CLAUDE_CODE_FORK_SUBAGENT=false, because fork mode forces every subagent into the background.
  • A background agent woke the main conversation when it finished, once for its report and once for its task notification. This happened in 501 of 861 main turns in the measured week and caused 20.6% of cost.
  • A foreground agent returns its report in the turn that spawned it.
  • Agents that one message spawns still run together. The main session waits while they run.
• scripts/usage-report.mjs: reports where the last N days of Claude Code usage went.
  • API-equivalent cost by agent type
  • the share from calls with context past 150k tokens
  • the share spent on full cache rewrites
  • the main turns that background agents started
• hooks/lib/_budget.mjs: holds the usage bounds. A test checks that the output style, the settings profile, and the option text agree with it. The test also checks that the output style plus the agent and skill descriptions stay under 21.5 KB.

Changed

• Mid-message skills: a /dotclaude: skill typed mid-message runs through the Skill tool. Before, the hook pasted the skill body into the context. The skill now loads the way Claude Code loads it, with forks, ! commands, and model and tool settings. A large skill is no longer cut at 9.5 KB. It costs one extra call per mid-message skill.
• Optional profile: the settings profile turns off built-in Claude Code features that cost context or turns, from profiles/optional.json. The profile applies each feature unless you skip it with --skip. The profile stamp covers both files.
  • the Artifact, Workflow, ScheduleWakeup, cron, and ReportFindings tools (about 46 KB on every request together)
  • the advisor
  • the Explore and Plan agents
  • bundled skills
  • auto memory
  • the refusal retry
  • auto-updates
• Output style: makes the main conversation the default place to work. Claude delegates only output that would flood the context, or parallel slices that the user asks for.
• AskUserQuestion: the settings profile no longer denies it. Its 4.9 KB per request is about $2 a week of cached reads at the measured volume. A structured question is worth that. apply-settings does not remove old denies. So remove "AskUserQuestion" from permissions.deny in ~/.claude/settings.json yourself.
• Compaction: the settings profile compacts at 200k tokens on every plan, where it was 400k on Max 20x. Run /dotclaude:apply-settings-profile to apply it. The session note asks for a handoff past about 200k on every plan. The output style says 200k instead of 400k.
• Agent cap: the settings profile runs 3 subagents at once, and 3 agents at once in a workflow. It was 6.
• Stale-profile check: session start detects a stale settings profile with one stamp instead of a check per release.
  • /dotclaude:apply-settings-profile writes env.DOTCLAUDE_SETTINGS_PROFILE, a hash of the shipped profile.
  • When the stamp is missing or differs, session start asks you to apply the profile again.
  • This replaces the checks for a missing Haiku mapping, a missing effort cap, and the old 400k compaction window.
  • Every existing install sees the notice once.

Removed

• Codex support: the codex-worker and codex-reviewer agents, the codex-fanout skill, and Codex setup (skills/setup-integrations/codex/, configure-codex.mjs, build-codex-catalog.mjs). Also the codex_delegation and allowed_codex_models options, the Codex Bash-guard rules, and every other Codex mention across the hooks, output style, and settings profile.
  • Reason: a second model source cost more to maintain than it gave. Delegation also spent Claude turns that forwarded briefs to GPT-6 workers and reviewed their diffs.
  • If you ran /dotclaude:setup-integrations codex, you may remove the files that it wrote under $CODEX_HOME (usually ~/.codex/):
    • dotclaude-luna.config.toml and dotclaude-review.config.toml
    • dotclaude-catalog-interactive.json, dotclaude-catalog-worker.json, and dotclaude-catalog-review.json
    • any *.dotclaude-backup-<timestamp> files next to them
  • Setup also set service_tier, model_catalog_json, [features] fast_mode, and on the Plus plan model, in config.toml. Remove those lines yourself to get the defaults of Codex back.
• General workflow skills: blind-spots, challenge, fresh-eyes, lessons-learned, polish, recap, troubleshoot, and review-code-changes.
  • dotclaude keeps the skills that its features use (settings profile, integrations, handoff, browser, CAPTCHA).
  • General workflow skills belong in xsyetopz/skills, where debug-software-failures covers troubleshoot.
  • The code-reviewer agent still runs on request.
  • Four of the removed skills were in the skill listing on every turn.
• apply-settings cleanup of old settings: the script no longer removes settings that earlier profiles wrote.
  • If your env still holds ANTHROPIC_DEFAULT_SONNET_MODEL=claude-opus-5-5 from 0.4.0, remove it yourself. Otherwise every sonnet agent runs on Opus 5.5.
  • The script no longer removes CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION from 0.2.0 either. Claude Code 2.1.283 still reads it. So the earlier claim that it had no effect was wrong.
• CodeGraph hooks and the codegraph_hint option: the session-start note and the per-prompt symbol list repeated what the CodeGraph section already says. codegraph install writes that section into ~/.claude/CLAUDE.md. The setup skill now checks for that section. It still tells you to remove the global codegraph prompt-hook entry.

Fixed

• Skill names in pasted text: a /dotclaude: skill name inside pasted text or a > quote no longer loads that skill.

Related Pages

• Previous: Release 0.6
• Next: Release 0.8
• Archive: the index of the archived pages.