opus-4-7-features

name	opus-4-7-features
description	Optional guidance for sessions using Claude Opus 4.7: adaptive thinking, xhigh effort, prompt caching, /btw side questions.

Opus 4.7 Features & Optimization

Optional guidance for sessions using Claude Opus 4.7 (v2.1.111+). Load this rule when you want to leverage Opus 4.7-specific capabilities.

Adaptive Thinking (Breaking Change)

Opus 4.7 changes the API contract for extended thinking. The legacy form is rejected:

• thinking: {type: "enabled", budget_tokens: N} → returns 400 Bad Request on Opus 4.7.
• Only thinking: {type: "adaptive"} is accepted.

With adaptive thinking, the model dynamically decides when and how much to think based on query complexity. There is no fixed token budget — Claude self-regulates depth.

How to control depth:

• Use the effort parameter (xhigh / high / medium / low) instead of budget_tokens.
• Thinking is also promptable in-message:
  • "Think carefully and step-by-step" → encourages deeper reasoning.
  • "Prioritize responding quickly" → encourages a shorter thinking pass.

Older models: enabled/budget_tokens is deprecated on Opus 4.6 and Sonnet 4.6 (still functional today, but on the way out). Migrate to adaptive proactively so your code keeps working when the deprecation lands.

Effort Level: xhigh

Opus 4.7 introduces a new effort level xhigh between high and max. Other models fall back to high when xhigh is requested.

Effort	When to use	Phase fit
low	Quick lookups, file searches, trivial edits	—
medium	Standard implementation, build tasks	Build
high	Architecture, planning, detailed review	Research, Plan, Review
xhigh	Complex multi-file reasoning, deep architectural analysis, hard debugging	Plan (large features), Review (cross-cutting PRs)

Availability: xhigh became selectable in Claude Code v2.1.111+ when running Opus 4.7.

Default change: From Claude Code v2.1.117+, xhigh is the default effort level for Opus 4.7. On v2.1.111–v2.1.116 the default is still high and you must opt in with /effort xhigh. Other models continue to default to high regardless of Claude Code version.

Set per-session with /effort <level> (opens an interactive slider if called without arguments). Persist across sessions with the effortLevel setting in settings.json.

Use xhigh when the task requires reasoning across many files or making subtle architectural judgments. Drop to high or medium for routine work — xhigh trades speed for depth.

Auto Mode

Auto mode is now GA — no longer requires --enable-auto-mode. A permission classifier handles tool approval automatically: safe read-only actions run without interruption, risky or destructive actions get blocked.

When to use: trusted repos where you are the sole developer and want minimal friction.

When NOT to use: unfamiliar repos, shared machines, or when onboarding to a new codebase where you want to see what the agent is doing.

Configuration: customize the classifier with the autoMode setting:

{
  "autoMode": {
    "environment": ["Trusted personal repo, no production access"],
    "allow": ["Run tests", "Read any file", "Search codebase"],
    "soft_deny": ["Delete files", "Modify CI configuration"]
  }
}

The environment array gives the classifier context about the workspace. allow and soft_deny are prose rules — the classifier interprets them, not pattern-matches them.

Cloud Commands

/ultrareview — Parallel multi-agent code review

Runs a comprehensive code review in the cloud using parallel analysis agents. Results are synthesized into a single review.

• /ultrareview — review current branch vs base
• /ultrareview 42 — review GitHub PR #42

Use for: complex PRs, cross-cutting changes spanning many files, or when you want a second opinion beyond the local /review agent.

/ultraplan — Cloud-based planning

Drafts a plan in the cloud, opens a web editor for review and commenting, then runs it remotely or pulls it back local. Auto-creates a cloud environment on first run.

Use for: large feature planning that benefits from cloud compute, or collaborative review of the plan before execution.

Permission Friction Reduction

/less-permission-prompts

Scans your recent transcripts for common read-only Bash and MCP tool calls, then proposes a prioritized allowlist for .claude/settings.json. Run this after a few sessions to reduce approval friction on safe operations you use repeatedly.

Bash improvements (v2.1.111+)

These no longer trigger permission prompts:

• Read-only bash commands with glob patterns (e.g., ls *.ts)
• Commands starting with cd <project-dir> &&

Prompt Caching

ENABLE_PROMPT_CACHING_1H

Set this environment variable to enable 1-hour prompt cache TTL (default is 5 minutes). Useful for long multi-phase sessions where the system prompt and early context stay warm across many turns.

export ENABLE_PROMPT_CACHING_1H=1

Or set in settings.json:

{
  "env": {
    "ENABLE_PROMPT_CACHING_1H": "1"
  }
}

The 5-minute TTL is fine for short tasks. Use 1-hour when a session will span 30+ minutes with the same context.

Caching Awareness

Per Anthropic's April 30 2026 guidance, prompt caching is the single most important optimization for Claude Code workflows — but adaptive thinking interacts with it in ways worth knowing:

• Switching between adaptive and enabled/disabled thinking modes breaks cache breakpoints for messages. System prompts and tool definitions remain cached, but per-message cache hits are lost.
• Pick one thinking mode per session and stay with it. Toggling mid-session quietly halves cache effectiveness.
• For long sessions on Opus 4.7, the recommended profile is: adaptive thinking + ENABLE_PROMPT_CACHING_1H=1 + a single sustained effort level.

Session Continuity

Session recap (/recap)

When you return to a session after being away, Claude shows a one-line recap of what was happening. Enable with /config or the awaySummaryEnabled setting. Invoke manually with /recap.

Useful for multi-session workflows where you context-switch between projects.

Monitor tool

The Monitor tool streams background events (stdout lines from a process) into the conversation. Claude can tail logs, watch builds, and react to events in real time.

Use for: tailing a dev server during UI work, watching a long build or test suite, or monitoring a background process you just started.