Claude Code Complete Guide

September 12, 2026 release refresh

Updated for Claude Code v2.1.263–v2.1.269 (September 6–11). npm checked on September 12, 2026 KST: latest/next=2.1.269, stable=2.1.236. Keep latest-feature evaluation separate from a team's stable-channel rollout.

• Plugins: claude plugin eval produces JSON and HTML evaluation reports; install, uninstall, update, enable, and disable accept --json. --plugin-dir can point to a parent directory containing multiple plugins.
• Commands and settings: /output-style [name] works over Remote Control and in cloud/headless sessions. maxEffortLevel caps effort globally or per model under modelSettings. bashEditDiffEnabled adds file-edit diffs to Bash results.
• Recovery and compatibility: 2.1.266 fixes forced gateway sign-in; 2.1.268 fixes HTTP 400s on Anthropic-compatible endpoints introduced in 2.1.265. /goal retries with backoff or pauses with a reason after API, network, or usage-limit failures.
• Permissions and operations: fixes cover symlink deny/ask rules, Bash tee write boundaries, and secrets in MCP/plugin errors. WebFetch has a default 300-second deadline. /ultrareview --post posts the PR comment directly when findings arrive.

Sources: official changelog, npm dist-tags.

Previous verification snapshot (September 5, 2026)

Claude Code is a terminal-based AI coding agent. This handbook is for advanced users who want to run it as an operating system for software work: multiple sessions, explicit plans, automated checks, project memory, subagents, hooks, MCP tools, and repeatable team workflows.

Latest Update Baseline

• Claude Fable 5.1 (claude-fable-5-1) shipped on 2026-09-01 and became the default Fable model in Claude Code v2.1.257 (minimum v2.1.255). fable and best resolve to Fable 5.1 unless ANTHROPIC_DEFAULT_FABLE_MODEL says otherwise; Fable 5 is now /model claude-fable-5. Pricing ($10/$50 per MTok), the 1M context window, 128K output, and always-on adaptive thinking match Fable 5, while prompt-cache reads drop to $0.25/MTok. Neither Fable model is the default on any plan.

• Classifier fallback is now category-based: Fable 5.1/5 biology flags rerun on Opus 5 and cybersecurity flags on Opus 4.8; Opus 5 itself reruns cybersecurity flags on Opus 4.8 and refuses biology flags. switchModelsOnFlag: false disables the switch, and third-party providers need ANTHROPIC_DEFAULT_OPUS_MODEL for fallback to work.

• Fable 5.1 has three Messages API breaking changes for custom harnesses: forced tool_choice (any/tool) returns 400, its thinking blocks are readable only by Fable 5.1, and editing earlier turns invalidates later thinking blocks (enforced for accounts created on or after 2026-08-31). Per-message effort, turn-scoped system messages, thinking.display: "updates", and content provenance are additive. Claude Code keeps history append-only, so interactive sessions are unaffected.

• Treat v2.1.252~v2.1.261 as the Fable 5.1 rollout: Claude apps gateway sessions keep fable/best on Fable 5 until gateways accept 5.1 (v2.1.257); /effort gains a session-only s key and CLAUDE_CODE_SUBAGENT_MODEL_FORCE arrives; v2.1.260 fixes the picker hiding Fable 5.1, Fable 5.1 prompt caching after tool results, and the [1m] tag on ANTHROPIC_DEFAULT_FABLE_MODEL, moves Opus/Fable 1M auto-compaction to just before the limit, and keeps the prompt cache when effort changes on Fable 5.1. /diff, /skill-doctor, managedMcpServers, --permission-prompts none, and timeFormat are the notable non-model additions.

• Treat v2.1.239~v2.1.247 as an operations and reliability refresh. /usage adds per-loop accounting; v2.1.242+ documents user/managed modelPicker and separate main/subagent cache TTLs. The changelog announces managed modelPricing, but the current Settings reference does not yet publish its schema. /permissions adds an Auto-mode rules tab; /claude-api cost-optimize profiles API spend one measured change at a time, although the current Commands signature still lags the v2.1.247 release note. feedbackDrafts accepts notify, quiet, or off at user/managed scope, with off removing SendFeedback.

• /cd now applies the destination's settings, hooks, .mcp.json, skills, and agents immediately. Subagents that hit maxTurns return an explicitly partial result, a first-call model 404 uses the session fallback chain, and non-interactive responses automatically continue after interrupted streams, connection loss, or stalls.

• Sonnet 5 sessions using the 1M window now auto-compact near 967K tokens. Marketplace text, terminal links, wildcard Bash allow rules, gateway credentials, and managed-settings startup boundaries received additional hardening.

• Treat v2.1.221~v2.1.237 as a session, automation, and containment refresh. Interactive sessions now enable inherited-context fork mode and background subagent spawning by default; the 200-per-session subagent cap is gone, while the default 20-concurrent and three-level boundaries remain. Cross-session SendMessage/ListAgents, @ mentions of other Claude sessions, GitLab MR workflows, self-hosted runners, and Remote Control were expanded and hardened.

• v2.1.236 adds ANTHROPIC_DEFAULT_MODEL as a low-precedence default for new sessions, and v2.1.237 adds the built-in Concise output style. Explicit --model, ANTHROPIC_MODEL, settings model, and organization defaults still take precedence.

• From v2.1.233, legacy Task/Todo tools are disabled by default for Opus 4.8, Sonnet 5, Fable 5, Mythos 5, and newer models; set CLAUDE_CODE_ENABLE_TODO_TOOLS=1 only for compatibility. The same release window adds claude self-hosted-runner, claude remote-control --continue, plugin command sources, archive-plus-SHA plugin pinning, and extensive trust, sandbox, OAuth, and worktree fixes.

• Treat v2.1.206~v2.1.220 as a model, session, and containment refresh: Opus 5 becomes the default Opus path; /fork creates a separate background session while /subtask owns the inherited-context subagent path; session and concurrency caps bound subagents; and new controls cover strict network allowlisting, optional filesystem isolation, DirectoryAdded, MCP init errors, and nested subagent forwarding in headless output.

• Treat v2.1.202~v2.1.205 as a workflow, background-session, and diagnostics refresh: dynamic workflow sizing and OTel attributes, a fast single-pass /review, MCP root updates, extensive Agent View/background reliability fixes, headless SessionStart hook streaming, a full /doctor setup checkup, transcript-tampering protection in Auto mode, and corrected --json-schema validation.

• Treat v2.1.197 as the Sonnet 5 baseline: Sonnet 5 is now the default Sonnet path in Claude Code, maps from the Anthropic API sonnet alias, has native 1M context, and supports xhigh effort.

• Treat v2.1.181~v2.1.201 as an operations and safety refresh: sandbox.credentials, destructive-command sandbox auto-approval, Windows PowerShell sandbox fixes, MCP login/logout, MCP idle timeout, Korean paste handling, non-interactive fallback stability, background/Remote recovery, plugin validation fixes, and app/schema bug fixes. v2.1.198 makes subagents run in the background by default (plus Claude in Chrome GA and the new /dataviz skill), v2.1.200 changes the default permission mode to "Manual", and v2.1.201 removes the mid-conversation system role on Sonnet 5 sessions.

• Track model restrictions and warnings: main model restrictions now center on Opus, Sonnet, Haiku, and Fable families; Haiku is not available as the main model but can be used for subagents, and deprecated model warnings are stronger.

• Keep subagent work bounded: v2.1.237 defaults to three nested layers below the main conversation and 20 concurrent subagents. v2.1.224 removed the old 200-per-session total. Fork mode is on by default in interactive sessions and off by default in -p/SDK flows.

• Use Fable 5.1 for the hardest and longest-running autonomous tasks when your organization has access. It requires Claude Code v2.1.255 or later, is selected with /model fable, and is not the default model. Fable 5 (v2.1.170+) remains available as a legacy model via /model claude-fable-5.

• Treat Opus 5 as the default high-capability Opus path for complex architecture, multi-step refactors, and long-running agentic work when Fable is unavailable or not appropriate.

• Check model aliases by provider. best uses the latest Fable (Fable 5.1) where available and otherwise falls back to the latest Opus model; opus maps to Opus 5 on the Anthropic API, Claude Platform on AWS, Bedrock, and Google Cloud Agent Platform. Microsoft Foundry remains on Opus 4.6, so pin a full model name or ANTHROPIC_DEFAULT_*_MODEL when reproducibility matters.

• Plan for Fable-specific fallback. Claude Code automatically reroutes Fable requests flagged by safety classifiers to Opus 5 (biology) or Opus 4.8 (cybersecurity); use --safe-mode when repository context itself may be triggering a classifier.

• Note the Fable 5 redeployment. Fable 5 and Mythos 5 were suspended 2026-06-12 under US export controls (~19 days) and Fable 5 was restored globally on 2026-07-01 with an improved safety classifier targeting the reported vulnerability-finding behavior — the automatic fallback follows it. Pro/Max/Team/select Enterprise include Fable 5 up to 50% of weekly usage limits through 2026-07-07, then via usage credits. Mythos 5 (no classifiers) and its successor Mythos 5.1 (same model as Fable 5.1 with a different safeguard level) are Project Glasswing-only and cannot be selected through /model.

• Use Agent View and background sessions for parallel work: claude agents, /bg, /background, claude --bg, and claude --bg --exec.

• Configure fallbackModel chains for availability failures before long autonomous sessions so overloads fail over deliberately instead of stopping the turn.

• Use /goal for completion-condition driven work and /workflows for dynamic workflows that orchestrate many background agents.

• Prefer /usage-credits over legacy /extra-usage, and use /reload-skills when commands or skills change during a session.

• Use native installation paths first; treat npm installation as legacy where official tooling does.

• Keep every agent session scoped to a concrete working directory, branch, and verification target.

• Prefer worktrees for parallel implementation, review, and experiment tracks.

• Put durable project rules in CLAUDE.md, but keep them short enough that the agent can obey them.

• Use subagents for bounded parallel work, not for urgent blocking decisions.

• Use hooks and MCP only after the manual workflow is clear enough to automate.

• Make verification part of the prompt contract: tests, lint, typecheck, browser checks, or review notes.

Handbook Map