Release 0.5

Release 0.5

Released 2026-09-26 (0.5.0) and 2026-09-27 (0.5.1). Makes dotclaude aware of the Claude plan and cuts the cost of long agent runs. 0.5.0 puts Sonnet 5 back on the cheaper agents, and adds plan notes, turn-limit handoff, and usage notes. 0.5.1 stops a /goal loop that wasted turns. This line requires Claude Code v2.1.283 or later.

Note: Plans and models has the research behind this release, which docs/subscription-tiers.md held. It covers how the Fable limit works, plan names in Claude Code and Codex, prices, dossier findings, and a scan of real session usage.

To update:

1. Restart Claude Code.
2. Run /dotclaude:apply-settings-profile again. It adds Sonnet 5 to availableModels, replaces the Sonnet deny with Fable denies, and removes the 0.4.0 mapping of sonnet to Opus 5.5.
3. If you set allowed_models yourself, add claude-sonnet-5 to it.
4. If you use the Codex agents, run /dotclaude:setup-integrations codex again.

Added

• Claude plan awareness: the claude_plan option (default auto) reads the plan from the account that Claude Code caches in ~/.claude.json. It uses organizationType, the rate-limit tier, billingType, and hasExtraUsageEnabled, and it never touches the token.
  • Plans: pro, max_5x, max_20x, team_standard, team_premium, enterprise, api.
  • A session-start note says what the plan means for model choice:
    • Max and premium seats: Fable draws up to 50% of the same weekly limit as every other model.
    • Pro: Fable runs on paid credits.
    • API: per-token price ratios.
    • Plans with a small 5-hour window: hand off or compact earlier.
• Codex session note: after Codex setup, a note points bounded, fully specified tasks at codex-worker. It uses the quota of the ChatGPT plan instead of the quota of Claude.
• Agent turn limits: each dotclaude agent learns its maxTurns at start and writes its report before it reaches the limit. The transcripts that the release scanned showed 39 implementer runs and 8 code-reviewer runs that stopped at their limit.
• Sonnet 5 scope reminder: agents on Sonnet 5 get it, because Sonnet 5 follows instructions literally (Anthropic's prompting guide).
• Turn-limit handoff (turn_limit_handoff, on by default): Claude does not resume an agent that stopped at its turn limit.
  • The hook rewrites the first SendMessage to that agent into a request for a handoff report. It denies later messages and points at a fresh agent.
  • Reason: a resumed agent keeps its whole context, and every turn re-reads it. In the scanned transcripts, implementer tasks ran a median of 68 turns (90th percentile 115, maximum 320), with context up to 856k.
• Usage notes (usage_notes, on by default):
  • They read the /usage numbers that Claude Code caches in ~/.claude.json, at most an hour old, including the Fable cap.
  • They tell Claude once when the session or weekly limit passes 75%, and again at 90%.
  • They read no token and get nothing.
• Message prefix: each message that dotclaude shows to Claude or the user starts with [dotclaude].
• Goal loop guard (0.5.1, goal_loop_guard, on by default): a /goal with an unmet condition no longer burns turns after Claude has been told to stop.
  • Reason: Claude Code's goal check blocks every stop. In one session, Claude replied "I'm staying stopped" to 9 blocks in a row until Claude Code gave up, and each round re-read about 710k tokens of context.
  • A new Stop hook ends the turn after two goal blocks in a row with no tool call between them. Claude Code then pauses the goal. The note names /goal <new condition> to change the goal, /goal clear to end it, and a message to resume.
  • The output style tells Claude to propose a replacement condition with ProposeGoal when a goal no longer matches the user's request. Otherwise it names those commands once, instead of restating that it is stopping.

Changed

• Sonnet 5: allowed again in the default allowed_models, in the profile's availableModels, and in the managed drop-in. The sonnet alias no longer maps to Opus 5.5.
• implementer: runs on Sonnet 5 at medium effort. Claude passes model: "opus" for a slice that needs design judgment or failed on Sonnet.
• CLAUDE_CODE_SUBAGENT_MODEL: the profile now sets Sonnet 5. The built-in general-purpose, Explore, and Plan agents run on it.
• mechanical-worker, test-runner, docs-writer: mechanical-worker and test-runner run on Sonnet 5 at low effort, and docs-writer at medium effort. They were Opus 5.5 at low effort. Sonnet 5 costs half as much as Opus 5.5 for input, output, and cache writes, and cache reads cost the same on both.
• Fable: denied as a subagent model on every plan. A fresh Fable context costs about $0.47 API-equivalent before any work, measured with claude -p.
• Pro and standard Team: with extra usage off, the model lock and the profile's availableModels leave Fable out.
• Fable note: adds Anthropic's instruction to edit files surgically and not to rewrite them.
• Output style: asks for these.
  • dotclaude agents over general-purpose
  • briefs sized to finish well inside the turn limit of an agent
  • a fresh agent, briefed from the handoff report, when an agent stops at its turn limit, and no resume
  • a handoff or compaction once the main context passes about 400k tokens
• Codex plans: groups by quota.
  • Pro 5x and the $100 self-serve Business plan review with Astra.
  • Team, Standard Business, Enterprise, and Edu get the Plus rules: Sol reviewer, no Astra.
  • Free and Go do not delegate to Codex.
  • The GPT-5.6 models stay out of the defaults, because they cost about twice the credits of their GPT-6 counterparts.

Previous: Release 0.4 · Next: Release 0.6