Advanced Features

Output Styles

10 interaction modes in .claude/output-styles/: teaching, explanatory (recommended), concise, technical, review, emoji, minimal, structured, debug, metrics.

Specification Templates

Templates in .claude/templates/ for the Explore → Specify → Plan → TDD → Audit → Commit workflow:

Template	Used by
spec-template.md	/work:work-specify
plan-template.md	/work:work-plan
tasks-template.md	/work:work-plan

Structure: specs/[feature]/ contains spec.md, plan.md, tasks.md, clarifications.md (opt).

Conventions: P1=MVP, P2=Important, P3=Nice-to-have, [P]=parallelizable, [US1]=User Story 1.

Proxmox templates (Terraform) available in .claude/templates/proxmox/.

Automatic Memory (CLI 2.1.76+)

Claude Code automatically saves and recalls memories as work progresses (preferences, decisions, project context). Memories are stored in ~/.claude/memory/.

To memorize	To put in CLAUDE.md	To put in rules/
Personal preferences	Project conventions	Per-language/framework rules
Architecture decisions	Mandatory workflow	Code patterns
Team context	Documentation references	Verification checklist

Best practices:

• Let Claude memorize preferences and decisions (avoids repetition)
• Keep in CLAUDE.md what is shared with the team (versioned in git)
• Do not duplicate: if it is in CLAUDE.md, no need to memorize it
• Use "remember that..." to force an explicit memorization

Effort Levels (CLI 2.1.76+)

/effort command to control the reasoning level (interactive slider since v2.1.111):

Level	Command	Use case
low	/effort low	Exploration, formatting, simple tasks
medium	/effort medium	Standard development, fixes
high	/effort high	Architecture, audit, complex refactoring, debug
xhigh	/effort xhigh	Maximum reasoning — critical system architecture, advanced security audit (Opus 4.8 required)

Recommendations per foundation workflow:

Phase	Recommended effort
/work:work-explore	low
/work:work-specify, /work:work-plan	high
/dev:dev-tdd	medium
/qa:qa-audit, /qa:qa-security	high or xhigh
/work:work-commit	low

Named Sessions (CLI 2.1.76+)

--name / -n flag to name a session at startup:

claude --name "feature-auth"
claude -n "fix-login-bug"

Combine with git worktrees for isolated and identifiable sessions:

git worktree add ../myapp-auth -b feature/auth
cd ../myapp-auth && claude -n "auth-feature"

VSCode URI Handler (CLI 2.1.76+)

Open a Claude Code tab programmatically from VSCode:

vscode://anthropic.claude-code/open

Useful for: CI/CD integration, setup scripts, notification hooks.

Fable 5 (most capable tier)

claude-fable-5 is Anthropic's most capable model — the tier above Opus 4.8 for the most demanding reasoning and long-horizon autonomous work. 1M context (default and max), 128K output, same tokenizer as Opus 4.8. Pricing is ~$10/$50 per MTok — 2× Opus 4.8 — so reach for it deliberately, not as a default.

✅ Availability (July 2026): Fable 5 is generally available again. Briefly export-control-suspended on 2026-06-12, the directive was lifted on 2026-06-30 and Fable 5 returned to global availability on 2026-07-01 (Claude Platform, Claude.ai, Claude Code, Claude Cowork). Mythos 5 stays restricted to a subset of US organizations. Opus 4.8 remains the default because Fable 5 is a costlier, deliberate escalation (2× Opus), not for availability reasons. (Anthropic statement)

Behaviourally it differs from Opus 4.8: thinking is always on (the raw chain of thought is never returned) and individual turns on hard tasks can run several minutes — plan for streaming and async check-ins. Opus 4.8 remains the documented default; Fable 5 is the costlier escalation when a task genuinely exceeds it. For the API-level caveats when building with the SDK (no thinking:{type:"disabled"}, no assistant prefill, refusal classifiers, 30-day data retention), see the dev-ai-integration skill.

Opus 4.8

Current frontier model (released 2026-05-28, supersedes Opus 4.7). Defaults to high effort. Anthropic reports it is roughly 4× less likely than Opus 4.7 to let a flaw in code it has written pass unremarked — a strong asset for the TDD and Audit phases of the workflow.

Adaptive Thinking: Claude automatically adjusts the depth of its reasoning based on the complexity of the task. Replaces budget_tokens (deprecated). 4 effort levels (low, medium, high, xhigh) to guide reasoning.

1M token context window by default on the Claude API, Amazon Bedrock and Vertex AI (no longer a beta opt-in). 128k output tokens, automatic Context Compaction. Reasoning is interleaved between tool calls (interleaved thinking) for agentic workflows.

xhigh unlocks Opus 4.8's maximum reasoning (introduced as a tier in v2.1.111). Auto mode available for Max subscribers (intelligent automatic permissions). Fast mode runs on Opus 4.8 (also available on 4.7/4.6).

Sonnet 5 (default tier, since 2026-06-30)

claude-sonnet-5 is Anthropic's most agentic Sonnet yet — released 2026-06-30 and now Claude Code's default model. It delivers near-Opus 4.8 quality on many agentic tasks at roughly a third of the cost, with a native 1M-token context. Pricing is $2/$10 per MTok introductory through 2026-08-31, then $3/$15 (vs Opus 4.8 at $5/$25).

The sonnet tier alias resolves to Sonnet 5 automatically, so agent model: sonnet frontmatter picks it up with no change needed. This foundation keeps Opus 4.8 as the recommended default for complex/critical work (TDD, Audit, architecture) — see best-practices.md — and uses Sonnet 5 where its price/perf wins: audits, analyses, and high-volume agentic passes.

Checkpoint / Rewind

Claude Code automatically saves the state of the code before each modification (checkpoint). To return to a previous state:

Method	Action
Esc × 2	Cancel the last modification and return to the checkpoint
/rewind	Choose a specific checkpoint in the history
/undo	Alias of /rewind (CLI 2.1.108+)

Recommended in the TDD Refactor phase: if the refactoring breaks the tests, /rewind (or /undo) is faster than a manual git revert.

Session Recap (CLI 2.1.108+)

/recap generates a structured summary of the session: decisions made, files modified, work state. Configurable in /config.

Situation	Action
Return after a break	/recap to recover the context
After /compact	/recap to verify what has been kept
Resumed session	Automatic recap on resume (if enabled in /config)

Fast Mode (Research Preview)

Same Opus 4.8 model, 2.5x faster output. Toggle with /fast. Premium cost (see Anthropic pricing).

Use case	Recommendation
Exploration, commits, simple tasks	Fast mode suitable
Architecture, audit, complex debug	Standard mode recommended

Context Compaction

Compaction automatically summarizes the context when the window approaches its limit. Manual trigger with /compact.

Command	Effect	When to use
/compact	Summarizes the context, keeps the essentials	Between long workflow phases
/clear	Erases the entire context	Total topic change
(auto)	Automatic compaction if necessary	Long sessions without action required

Associated hooks: PreCompact (before compaction, matcher manual or auto) and PostCompact (after). See docs/reference/hooks-reference.md.

Claude Code Action (GitHub)

Official Anthropic action to integrate Claude into GitHub workflows. Reviews PRs, responds to @claude mentions, implements changes.

Scenario	Trigger	Template
Automatic PR reviews	pull_request: opened, synchronize	.claude/templates/github-actions/claude-review.yml
Security review (critical files)	pull_request: paths: src/auth/**, src/api/**	.claude/templates/github-actions/claude-security-review.yml
@claude mention	issue_comment: @claude	Included in claude-review.yml

Prerequisites: an Anthropic API key (pay-per-use) or a cloud provider (Bedrock, Vertex, Foundry). The Max plan (interactive OAuth) does not work in CI/CD.

Quick setup: /install-github-app in Claude Code, or add ANTHROPIC_API_KEY in the GitHub secrets then copy the template into .github/workflows/.

Source: anthropics/claude-code-action

Agent Teams (Experimental)

Parallel coordination of agent teams. Activation: CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in .claude/settings.json.

Modes: auto (default), in-process, tmux. Command: /work:work-team "description".

See .claude/skills/agent-teams/SKILL.md for the full documentation.

Subagent reliability (CLI 2.1.113+)

A subagent stuck for more than 10 minutes without progress fails with an explicit error message instead of remaining in a silent hang. Isolated worktrees grant Read/Edit on the files of their own worktree. Permission dialog crashes during tool requests by a teammate are fixed (CLI 2.1.114+).

Cross-session messaging hardening (CLI 2.1.166+)

Messages relayed via SendMessage from other Claude sessions no longer carry user authority: a teammate or remote session cannot approve permissions or trigger privileged actions on the user's behalf. Treat inter-agent messages as data, not as user instructions.

Dynamic Workflows (Opus 4.8)

Introduced with Opus 4.8: a native Workflow capability that orchestrates work across tens to hundreds of agents in the background for large, complex tasks. Unlike the two mechanisms above, control flow is deterministic and scripted (loops, conditionals, fan-out, fan-in) rather than model-driven — you describe the structure (pipeline, parallel fan-out, adversarial verification) and the harness drives the agents.

Ask Claude to "create a workflow that…" and it generates a script orchestrating the fleet. Typical shapes:

Shape	When to use
Pipeline	Each item flows through N stages independently (migrate → verify per file) — no barrier between stages
Parallel fan-out	Independent tasks that must all complete before the next step (multi-dimension review)
Adversarial verify	Spawn N skeptics per finding; keep only what survives a majority refute — kills plausible-but-wrong results
Loop-until-dry	Unknown-size discovery (bugs, edge cases): keep spawning finders until K rounds find nothing new

When to use which mechanism

Mechanism	Coordination	Best for
parallel-agents skill	Task-based fan-out, manual	A handful of independent sub-tasks within one session
Agent Teams (experimental)	Inter-agent messaging, live	Long-running collaborative work with a lead + teammates
Dynamic Workflows	Deterministic script, background	Scale (dozens–hundreds of agents): audits, migrations, exhaustive review

Explicit opt-in: dynamic workflows can spawn many agents and consume a large token budget — they run only when you request that scale, not by inference.

MCP Configuration

.mcp.json ships empty ({"mcpServers": {}}) — no server is active by default. There is no per-server enabled flag: a server is active iff its block is present in .mcp.json. The curated catalogue ships alongside as .mcp.json.example. Servers you can copy from it:

Server	Usage
filesystem	Advanced file access
memory	Persistent memory
github	GitHub integration
postgres	PostgreSQL connection
puppeteer	Browser automation
slack	Team communication
sentry	Error monitoring
linear	Project management

To enable a server, copy its block from .mcp.json.example into .mcp.json (there is no enabled flag — presence = active). Provide the referenced environment variables in .env.

MCP Channels

MCP servers can push messages into a session via --channels. Available through channel plugins (Telegram, Discord, iMessage) that install as MCP servers.

Channel	Plugin	Usage
Telegram	telegram-channel	Messages and commands from Telegram
Discord	discord-channel	Messages from a Discord server
iMessage	imessage-channel	Messages from iMessage (macOS)
Slack	slack (native MCP)	Slack notifications and messages

Activation: claude --channels at startup. Channels have access to the filesystem, MCP and git of the local session.

Permission relay: channels declaring the permission capability can relay approval requests to your phone.

MCP Elicitation (CLI 2.1.76+)

MCP servers can request structured input from the user during a task via interactive dialogs. Associated hooks: Elicitation (request) and ElicitationResult (response).

Managed Agents — private MCP & sandbox (Enterprise)

As of May 2026, Claude Managed Agents can run in a sandbox you control and connect to your private MCP servers — both the execution environment and the services it reaches stay within your enterprise boundaries. Paired with Compliance API integrations (security/compliance tooling), this lets IT and security teams govern Claude across the platform. Enterprise-only; out of scope for the local foundation, listed here as a pointer. See the Anthropic news feed.

MCP OAuth RFC 9728 (CLI 2.1.85+)

Automatic discovery of Protected Resource Metadata for OAuth MCP servers. Simplifies OAuth 2.1 authentication by exposing the authorization server URL via a standard endpoint. Servers can provide a headersHelper and use the CLAUDE_CODE_MCP_SERVER_NAME and CLAUDE_CODE_MCP_SERVER_URL environment variables.

MCP Step-up Authorization (CLI 2.1.84+)

RFC support for step-up authorization: MCP servers can return a 403 insufficient_scope to trigger a refresh token with an extended scope. Useful for sensitive operations that require reauthentication without breaking the session.

MCP Result Size Override (CLI 2.1.84+)

MCP tools can declare _meta["anthropic/maxResultSizeChars"] (up to 500K) to override the result persistence limit. Useful for tools that return large payloads (exports, reports, diffs).

Async Hooks (CLI 2.1.70+)

"async": true property to run a hook in the background without blocking the session. Recommended for logging and notification hooks. Security hooks (gitleaks, pre-commit tests) must remain synchronous.

Hook	Mode	Reason
SessionStart, PreToolUse, PostToolUse, Setup	sync	Critical actions (security, formatting)
SessionEnd, PreCompact, PostCompact, SubagentStop, Notification	async	Logging, no impact on the workflow
TeammateIdle, TaskCompleted, InstructionsLoaded	async	Observability, non-blocking
Elicitation, ElicitationResult	async	MCP logging

HTTP Hooks (CLI 2.1.70+)

"http" type to send a JSON POST to an external URL (webhook). Generic webhook configuration example:

{
  "type": "http",
  "url": "https://your-webhook-url.example.com/hook",
  "headers": { "Authorization": "Bearer ${WEBHOOK_TOKEN}" },
  "timeout": 5000,
  "async": true
}

Recommendations: always async: true and onFailure: "ignore" to avoid blocking the session if the remote service is unavailable.

Claude Code Security (Enterprise/Team)

Vulnerability scanning tool that reasons about code beyond traditional static analysis — data flows, interactions between components, and architectural patterns.

Claude Security entered public beta on 2026-05-22 for Claude Enterprise customers: it scans code repositories for vulnerabilities and generates proposed fixes. It is the productized form of Project Glasswing, whose first quantified results (also 2026-05-22) reported 10,000+ high/critical-severity vulnerabilities found across widely used internet software via ~50 partner organizations.

Prerequisites: Enterprise or Team plan. Complement to /qa:qa-security (local, OWASP-based) for an in-depth audit. See Anthropic announcement.

CLAUDE.md @imports

@path/to/file syntax to import files. Relative and absolute paths supported, recursive imports (max 5 levels). View loaded imports with /memory.

Plugins

Ecosystem of community extensions for Claude Code. A plugin can contain skills, agents, hooks and MCP servers.

Action	Command
Load a local plugin (directory or .zip)	claude --plugin-dir ./my-plugin
Load a remote plugin	claude --plugin-url https://example.com/my-plugin.zip
Namespaced skills	/my-plugin:skill-name
Plugin executables	Files in bin/ invocable as Bash commands

Plugins can be distributed via an Anthropic-managed directory. Setting disableSkillShellExecution to disable shell execution in unverified plugins.

Evaluating a plugin before adoption (CLI 2.1.128+)

Both --plugin-dir <path> (local directory or .zip) and --plugin-url <url> (remote .zip) are session-scoped: the plugin is loaded for the current claude invocation only and disappears at session end. They are repeatable, so multiple plugins can be combined for a single trial. This is the foundation's recommended way to validate a plugin against your workflow before requesting it for permanent inclusion in a preset's marketplacePlugins list — consistent with the validation-first policy described in docs/recipes/recommended-vendor-skills.md.

Recipe — try a plugin without committing to it:

1. Get the plugin. Either clone the repo (git clone <repo>) or download the release .zip to a temp dir.
2. Validate the manifest. Run claude plugin validate <unzipped-path>. The validator reads <path>/.claude-plugin/plugin.json; pass an unzipped directory, not a .zip (the validator reads the path argument as a JSON file). Confirm at minimum name, version, description are present; author is recommended.
3. Load it transiently. Either claude --plugin-dir ./plugin/ or claude --plugin-dir ./plugin.zip or claude --plugin-url <url>. The plugin is active for this session only.
4. Use the plugin in your real workflow. Invoke its skills (/<plugin-name>:<skill>), trigger its hooks, exercise the surface you care about. Take notes.
5. Cleanup is automatic. Exit the session — no installed state remains. Repeat with --plugin-dir/--plugin-url if you want to compare with another plugin or against the un-augmented baseline.

If after this trial the plugin is worth adopting, raise an issue or pull request against the relevant preset under .claude/presets/ with the validation evidence (the marketplace-audit methodology under specs/marketplace-audit/ describes the bar).

Scheduled Tasks (Cloud)

Recurring jobs executed on Anthropic's cloud infrastructure. Useful for ongoing operational tasks without an active local session.

Use case	Description
PR reviews	Automatic review of pull requests
CI monitoring	Continuous monitoring of the CI pipeline
Dependency audits	Periodic audit of dependencies
Doc syncing	Documentation synchronization

Configuration via /tasks, /schedule or the API. Requires a Pro/Max/Team/Enterprise plan.

See also Routines (section above) for more complex automated workflows combining prompts, repos and connectors.

Computer Use

Direct integration in Claude Code (Pro/Max). Allows opening files, launching dev tools, clicking and navigating in the interface without additional setup.

Useful for: visual tests, UI interactions, workflows requiring a browser or emulator.

Routines (CLI 2.1.108+)

Routines are automated workflows that run on Anthropic's cloud infrastructure. A routine combines a prompt, one or more repos, and connectors into a single configuration executable on schedule, via API, or on a GitHub event.

Property	Description
Prompt	The instructions to execute
Repos	One or more target repositories
Connectors	MCP servers, GitHub events, API triggers
Execution	Anthropic cloud — runs even with laptop turned off

Use cases with the foundation:

Routine	Description	Foundation equivalent
Automatic PR reviews	Review every new PR	/qa:qa-review in cloud version
Periodic audit	Weekly security/quality audit	/qa:qa-audit in scheduled version
Automatic standup	Daily activity summary	/ops:ops-standup in cloud version
Dependency check	Audit deps every Monday	/ops:ops-deps in scheduled version

Configuration via the Anthropic console or /schedule. Requires a Pro/Max/Team/Enterprise plan.

Ultraplan and Ultrareview (CLI 2.1.101+)

Cloud commands that delegate work to parallel agents on Anthropic's infrastructure.

Command	Description	When to use
/ultraplan	Plan in cloud: draft, review in a web editor, remote or local execution	Complex architecture, multi-file plans
/ultrareview	Parallel multi-agent review in cloud	Large PRs, in-depth reviews

/ultraplan automatically creates a cloud environment on first launch. The plan can be revised via a web editor before execution.

/ultrareview launches several agents in parallel for a more exhaustive review than local /qa:qa-review. Ideal for PRs of more than 500 lines.

Local /code-review --fix (CLI 2.1.152+)

The local /code-review flow gained --fix: review findings are applied automatically to the working tree instead of only being reported. Combine with the foundation's qa-loop (audit + iterative fix to score 90) for a tight local loop, or --comment to post findings as inline PR comments. The same release added native skill management (list/enable skills from within Claude Code) — complementary to the foundation's writing-skills / base-maintenance conventions.

Rate limits: Anthropic doubled Claude Code rate limits (May 2026) for developers, startups and enterprises — fewer throttling interruptions on agentic/parallel workloads.

TUI Fullscreen (Research Preview, CLI 2.1.89+)

Alternative rendering mode that takes control of the terminal surface like vim or htop. "Fullscreen" refers to taking over the drawing surface, not to maximizing the window.

Activation: /tui fullscreen (CLI 2.1.110+) or CLAUDE_CODE_NO_FLICKER=1 before launch. Deactivation: /tui default.

Three key benefits

Benefit	Impact
Flicker-free	No more flickering in VS Code terminal, tmux, iTerm2 on long sessions
Constant memory	Only visible messages in the render tree → flat RAM even on conversations of several hours
Mouse support	Click-to-expand tool results, click URLs/file paths, click-and-drag selection with auto-copy

Visual signal: in fullscreen, the prompt input stays fixed at the bottom instead of scrolling up with the output.

Associated commands

Mode	Command	Description
Fullscreen	/tui fullscreen	Activates the mode (persists via the tui setting)
Default	/tui default	Deactivates the mode
Status	/tui	Displays the active renderer
Focus	/focus	Condensed view: prompt + 1 line per tool + final response (separable from /tui)
Transcript	Ctrl+O	Toggle transcript mode with less-style navigation

Navigation in fullscreen

Shortcut	Action
PgUp / PgDn	Half-screen scroll (or Fn+↑/Fn+↓ on Mac)
Ctrl+Home / Ctrl+End	Start / end of conversation
Ctrl+O then /	Search in the transcript
Ctrl+O then [	Dump the conversation into the terminal's native scrollback
Ctrl+O then v	Open the transcript in $EDITOR

Environment variables

Variable	Usage
CLAUDE_CODE_NO_FLICKER=1	Activates fullscreen at startup (equivalent to the tui setting)
CLAUDE_CODE_DISABLE_MOUSE=1	Keeps flicker-free + flat memory, but disables mouse capture (useful in SSH/tmux)
CLAUDE_CODE_SCROLL_SPEED	Scroll wheel speed multiplier (1-20, terminal-dependent default)

tmux compatibility

• Requires set -g mouse on in ~/.tmux.conf for the scroll wheel
• Incompatible with tmux -CC (iTerm2 integration mode)

Push Notifications (CLI 2.1.110+)

Claude can send push notifications to mobile when Remote Control is enabled. Useful for long background tasks.

Activation: enable Remote Control + "Push when Claude decides" in /config. Claude notifies at task end or when a human decision is necessary.

/loop Command

Run a prompt or command at regular intervals:

/loop 5m "run tests and report failures"   # every 5 minutes
/loop "check CI status"                     # auto-paced by Claude (CLI 2.1.101+)

Alias: /proactive (CLI 2.1.105+). Without an interval, Claude auto-determines the optimal frequency.

Wakeup control: Esc cancels pending wakeups (CLI 2.1.113+), a "Claude resuming /loop wakeup" message confirms restart at each tick.

Monitor Tool (CLI 2.1.98+)

Native tool that spawns a watcher in the background and streams its events into the conversation: each event arrives as a new transcript message that Claude reacts to immediately. Replaces Bash sleep loops that block an entire turn.

Use case	Example prompt
Application log tail	Tail server.log and notify me as soon as a 5xx appears
Babysit CI on a PR	Watch the CI of this PR and auto-fix the lints
Watch a dev server	Watch npm run dev and restart on crash
Track a training run	Monitor the training log and alert on loss spike

Recommended pairing with /loop (auto-pace): Claude chooses Monitor over polling when the source emits events directly.

Foundation integration: Monitor is useful in /qa:qa-loop, /ops:ops-ci-fix, and long-running /loop workflows where a bash sleep loop would be the alternative.

/autofix-pr (CLI 2.1.92+)

Enables PR auto-fix on Claude Code Web from the terminal for the PR of the current branch. After push, Claude monitors the CI and review comments and pushes fixes until green without requiring an active local session.

git push -u origin feature/auth
/autofix-pr

When to use	Description
Long CI cycle	Lints, tests, type-check looping on small fixes
PR with many review nits	Renames, formats, docstrings requested in review
Asynchronous work	You want to leave the terminal and let Claude finish

Complement to /work:work-pr: /work:work-pr creates the PR, /autofix-pr makes it converge autonomously. Requires Claude Code on the web (Pro/Max/Team/Enterprise).

/powerup Command

Interactive lessons and animated demos to discover Claude Code's features. Useful for onboarding new users.

/less-permission-prompts (CLI 2.1.111+)

Scans session transcripts and proposes optimized permission allowlists. Reduces the number of permission prompts without compromising security.

Useful for: onboarding (generating initial permissions), sessions with too many prompts, team configuration optimization.

Since CLI 2.1.166, deny rules accept glob patterns in the tool-name position (e.g., "*" denies all tools, mcp__github__* denies every GitHub MCP tool) — useful to deny whole tool families instead of listing each tool.

Advanced Prompt Caching (CLI 2.1.108+)

Variable	TTL	Description
ENABLE_PROMPT_CACHING_1H	1 hour	Extended prompt cache for long sessions (API key, Bedrock, Vertex, Foundry)
FORCE_PROMPT_CACHING_5M	5 minutes	Forces 5 min TTL (useful if telemetry is disabled)

Enable in .claude/settings.local.json (not committed):

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

Advanced Environment Variables

Variable	Description
CLAUDE_CODE_NO_FLICKER=1	Alt-screen rendering without flicker (virtualized scrollback)
CLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1	Strips credentials from subprocess env variables
MCP_CONNECTION_NONBLOCKING=true	Skip waiting for MCP connection in -p mode (headless/CI)
ENABLE_PROMPT_CACHING_1H=1	1-hour prompt cache (significant savings)
FORCE_PROMPT_CACHING_5M=1	Force 5-minute prompt cache
CLAUDE_CODE_USE_POWERSHELL_TOOL	Opt-in/out of the PowerShell tool on Windows (CLI 2.1.111+)
CLAUDE_CODE_ENABLE_AWAY_SUMMARY=1	Forces the session recap even if telemetry is disabled (CLI 2.1.108+)
CLAUDE_CODE_PERFORCE_MODE=1	Edit/Write fail on read-only files with p4 edit hint (CLI 2.1.98+)
CLAUDE_STREAM_IDLE_TIMEOUT_MS	Configures the streaming inactivity watchdog (CLI 2.1.84+)
OTEL_LOG_RAW_API_BODIES=1	Emits full API request/response bodies via OpenTelemetry (CLI 2.1.113+)
MAX_THINKING_TOKENS=0	Disables thinking, including on models that think by default — same effect as --thinking disabled or the per-model toggle (CLI 2.1.166+)

Advanced Settings

Setting	Description
fallbackModel	Up to three fallback models tried in order when the primary is overloaded or unavailable; unexpected API errors retry on the fallback automatically (CLI 2.1.166+)
disableSkillShellExecution	Disables inline shell execution in skills, commands and plugins
managed-settings.d/	Drop-in directory for policy fragments (Team/Enterprise)
sandbox.network.deniedDomains	Blocks specific domains even under a wildcard allowedDomains (CLI 2.1.113+)
sandbox.failIfUnavailable	Exit with error if sandbox enabled but unavailable (CLI 2.1.83+)
modelOverrides	Maps picker entries to custom model IDs (Bedrock Application Inference Profile ARNs, etc.) (CLI 2.1.84+)
worktree.sparsePaths	Sparse-checkout for large monorepos with claude --worktree (CLI 2.1.76+)
autoScrollEnabled	Disables auto-scroll in fullscreen mode (CLI 2.1.110+)
showThinkingSummaries	Generates extended thinking summaries (default now false — CLI 2.1.108+)
disableDeepLinkRegistration	Prevents registration of the claude-cli:// protocol handler (CLI 2.1.83+)
feedbackSurveyRate	Admin sample rate of the session quality survey (CLI 2.1.76+)
forceRemoteSettingsRefresh	Blocks startup until remote managed settings are refreshed (policy)
Theme "Auto (match terminal)"	Automatically follows the terminal's dark/light mode (CLI 2.1.111+)
attribution.sessionUrl	Set false to omit the claude.ai session link from commits and PRs (web / Remote Control sessions) — the native toggle for the "no AI attribution in commits/PRs" convention (CLI 2.1.183, June 2026)
sandbox.credentials	Blocks sandboxed commands from reading credential files and secret env vars — defense-in-depth for the secrets-management posture (CLI 2.1.187, June 2026)
respondToBashCommands	Set false so ! <cmd> shell-mode output is not auto-explained by the model (saves a prompt per ! call) (CLI 2.1.186, June 2026)

Permission parameter matching — Tool(param:value) (CLI 2.1.178, June 2026)

permissions.deny / permissions.ask rules can now match a tool's input parameters, not just its name, with a * wildcard. This gives the foundation's security posture a finer lever than an all-or-nothing tool block:

// .claude/settings.json
{
  "permissions": {
    "ask": [
      "Agent(model:opus)",        // confirm before spawning a costly Opus sub-agent
      "Agent(isolation:remote)"   // confirm before a remote/cloud agent run
    ],
    "deny": [
      "Agent(model:fable*)"        // block the (export-restricted) Fable tier outright
    ]
  }
}

Use it to cap sub-agent cost (gate model:opus/model:fable*), to require confirmation before isolation:worktree/remote agents, or to scope any tool by a sensitive parameter. The foundation ships none of these by default (it stays cost-neutral) — they are opt-in hardening a project can add.

LSP (Language Server Protocol)

Semantic code navigation via .lsp.json. Activation: export ENABLE_LSP_TOOL=1.

12 supported languages (TypeScript, Python, Go, Rust, Java, C/C++, C#, PHP, Kotlin, Ruby, HTML, CSS).

LSP for: symbol definitions, references, diagnostics. Grep for: textual searches. See .claude/rules/lsp.md for detailed rules.

Marketplace Curation Engine

A deterministic, billing-safe system that keeps the recommended vendor-skill list current — observe-and-propose only, never auto-install. Two scheduled jobs:

• Nightly rot-watch (scripts/curation-watch.sh) — LLM-free → $0 tokens. Re-verifies every recommended/pointed skill (archived / abandoned / sustained popularity-collapse / license-change / content-drift vs the pinned ref) and emits ONE digest. Opt-in, fail-safe GitHub emission: --emit-issue (propose-only) and --emit-pr --draft (low-risk re-pin, gated by a pin-time safety screen).
• Monthly discovery (scripts/curation-discover.sh) — model-using under a hard token budget + fail-safe (the 2026-06-15 agentic-billing change). Trust + safety gates run first (LLM-free); only survivors reach the advice-neutrality + fit judge. Flags moat-encroachment as a strategic signal, never an auto-candidate.

Deploy both as cron/systemd timers — the monthly job uses a dedicated, capped API key in its own env, never mixed with the $0 nightly path. See docs/recipes/curation-bot-deploy.md. Policy: advice-neutrality + provenance (not publisher-veto); foundation-vs-vendor precedence in .claude/rules/vendor-precedence.md.