MindPattern
← The Wire
Source trail

Anthropic

Public MindPattern findings, entities, and graph evidence that cite this source.

Findings
40
All-time hits
39
High value
16
Last seen
2026-07-01

Connected entities

Claude Code Docs — Orchestrate teams of Claude Code sessionsRun a Claude Code 'agent team' by flipping one settings flag — and put all shared context in the spaClaude Code DocsClaude Code v2.1.194–195 Adds Trusted Devices Gating and CLAUDE_CODE_DISABLE_MOUSE_CLICKSClaude Code v2.1.193 Adds Shell-Classifier Routing, Assistant-Response Telemetry, and JetBrains TermClaude Code Docs — Changelog (June 2026)Nested subagents can now spawn their own subagents — background trees capped at 5 levels deepClaude Code ChangelogSpawn Claude Code teammates directly via the Agent tool's name param — no TeamCreate step

Related findings

  1. 2026-07-02 / SKILLSScope Claude Code hooks inside skill/subagent frontmatter, not just settings.jsonBeyond global hooks in settings.json, Claude Code lets you define hooks directly in a skill's or subagent's frontmatter, and those hooks are scoped to that component's lifecycle — they only fire while that skill/subagent is active. This means you can attach deterministic guards (block a destructive command, enforce a lint/test, redact secrets) to a specific workflow without imposing them on every session. For builders, it's the clean way to ship a self-contained skill that carries its own safety rails instead of relying on the model to remember instructions or on a bloated global hook config.
  2. 2026-07-02 / TOOLSTip: Claude Code's New /dataviz Skill Enforces Chart and Palette DisciplineClaude Code added a `/dataviz` skill that provides chart and dashboard design guidance plus a runnable color-palette validator, so agents produce accessible, system-consistent visualizations in light and dark instead of ad-hoc colors. Invoke it before writing any chart code (matplotlib, Recharts, d3, or inline SVG) to catch contrast and categorical-color issues that models otherwise get wrong.
  3. 2026-07-01 / TOOLSPattern: Grader Loops — Subagents That Revise Until Output Meets a RubricThe emerging production pattern pairs fan-out subagents with a separate grader that sends each subagent back to revise until its result satisfies a rubric, rather than accepting the first completion. This 'Performance Outcomes' style loop turns probabilistic agent output into something closer to test-gated work. Adversarially verifying or rubric-grading each subagent's result before merging is the reliability lever for large fan-outs.
  4. 2026-07-01 / TOOLSTip: 'Skill Teaches the How, Hook Enforces the Rule, Subagent Isolates the Work'Anthropic's steering guidance codifies a decision rule for production Claude Code setups: use a Skill to teach a workflow, a Hook to deterministically enforce a constraint, and a Subagent to isolate work off the lead agent's context — and use all three together, not interchangeably. It's a clean heuristic for deciding which mechanism a given requirement belongs to. Reaching for a skill when you needed a hook (or vice versa) is the common mistake.
  5. 2026-07-01 / TOOLSTip: Use SubagentStop's additionalContext to Extend a Subagent Instead of Failing ItThe SubagentStop hook fires when a subagent finishes and before its summary reaches the lead agent. Its payload now includes hookSpecificOutput.additionalContext, so you can inject new context and extend the subagent's turn rather than treating completion as a binary stop/error. This turns SubagentStop into a verification-loop primitive — re-feed a rubric or missing requirement and let the subagent continue.
  6. 2026-06-30 / SKILLSRun a Claude Code 'agent team' by flipping one settings flag — and put all shared context in the spawn promptClaude Code's research-preview Agent Teams (enable CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS in settings.json) let one orchestrator spawn parallel teammates that coordinate through the shared file system, ideal for parallel research, competing-hypothesis debugging, and cross-layer features. The non-obvious gotcha: teammates start fresh from CLAUDE.md/MCP only and have zero knowledge of your prior chat, so anything they need must be written into the spawn prompt. Tokens scale per teammate, so reserve teams for research/review/new modules and keep routine work single-session.
  7. 2026-06-30 / TOOLSClaude Code Promotes Its Agent-Orchestration Bundle (Nested Subagents, Scoped Permissions, Usage Attribution) to GAClaude Code's June 2026 release line moved a cluster of orchestration features to general availability: nested sub-agents up to 3-level depth for layered task decomposition, fallbackModel chains for resilience, scoped permissions enforcing least-privilege on sub-agents, per-agent/per-task usage attribution, streaming agent logs (beta), and a community tool marketplace. Agent-tree checkpointing remains beta. Together these turn Claude Code from a single-agent loop into a governed multi-agent runtime—worth re-architecting deep task pipelines around.
  8. 2026-06-29 / TOOLSClaude Code v2.1.194–195 Adds Trusted Devices Gating and CLAUDE_CODE_DISABLE_MOUSE_CLICKSAnthropic shipped Claude Code past v2.1.193 with a new 'Trusted Devices' control letting Team/Enterprise admins require device verification before a member can view or steer local Claude Code sessions remotely. v2.1.195 adds CLAUDE_CODE_DISABLE_MOUSE_CLICKS (kills fullscreen click/drag/hover while keeping wheel scroll), fixes hook matchers with hyphenated identifiers (e.g. code-reviewer, mcpbrave-search), and improves background-agent reliability and Linux voice detection. These are net-new beyond the v2.1.185–193 batch covered last week.
  9. 2026-06-28 / TOOLSTip: Use `claude mcp login` to Authenticate MCP Servers From the Shell Instead of Hand-Editing OAuth ConfigClaude Code now ships a `claude mcp login` subcommand that runs the OAuth/auth flow for a configured MCP server directly from the terminal, plus a startup notice that flags which servers still need authentication. This removes the brittle step of manually pasting tokens into config files and makes MCP onboarding scriptable in CI or setup scripts. Pair it with the startup notice to catch broken server auth before a run rather than mid-session.
  10. 2026-06-27 / SOURCESClaude Code's June 25 Update: /rewind Across /clear, Permanent Agent Stop, ~37% Lower Streaming CPUClaude Code's June 25, 2026 release adds /rewind to resume a conversation from before /clear was run, makes stopping a background agent from the tasks panel permanent (no more agents resurrecting), and auto-reconnects MCP tools when a call returns 401/403. It also cuts CPU use during streaming by about 37% and improves OAuth and sandbox-prompt handling. For daily Claude Code users, the permanent-stop fix and MCP auth auto-recovery remove two long-standing footguns in long agent sessions.
  11. 2026-06-27 / TOOLSTip: Capture Model Output in Your Observability Stack via the New claude_code.assistant_response OTel EventClaude Code 2.1.193 adds a claude_code.assistant_response OpenTelemetry log event that contains the model's actual response text, not just token counts and timing. If you already export Claude Code OTel data, you can now log, diff, and audit what agents actually said — useful for regression-tracking prompt changes and for after-the-fact review of autonomous runs.
  12. 2026-06-27 / TOOLSTip: Route Every Shell Command Through the Auto-Mode Classifier With autoMode.classifyAllShellNew in Claude Code 2.1.193, setting autoMode.classifyAllShell forces all Bash and PowerShell commands through the auto-mode classifier instead of only a subset, so risky commands are evaluated uniformly. Pair it with the newly added denial reasons (now in the transcript and denial toast) to debug exactly why a command was blocked rather than guessing at permission config.
  13. 2026-06-27 / TOOLSClaude Code v2.1.193 Adds Shell-Classifier Routing, Assistant-Response Telemetry, and JetBrains Terminal FixesThe June 25 release adds an autoMode.classifyAllShell setting that routes every Bash/PowerShell command through the auto-mode classifier, surfaces auto-mode denial reasons in the transcript and denial toast, and emits a new claude_code.assistant_response OpenTelemetry log event carrying the model's response text. It also fixes flickering in JetBrains 2026.1+ terminals by enabling synchronized output and repairs Shift+non-ASCII and Kitty keyboard-protocol handling in WezTerm, Ghostty, and kitty.
  14. 2026-06-26 / SKILLSRun background Claude Code agents in isolated git worktrees with automatic cleanupLaunching an agent with run_in_background: true plus isolation: "worktree" gives it its own working directory and branch sharing the same repo history, so parallel agents never touch each other's files and you keep working in the main session until you're notified — no polling. Set isolation: worktree in a subagent's frontmatter to make it permanent. Cleanup is automatic: an unchanged worktree (no commits, no untracked files, no uncommitted edits) is removed on exit, so disposable parallel work leaves no residue.
  15. 2026-06-26 / TOOLSClaude Code Late-June v2.1.x Fixes: Background-Session Env-Var Bleed, Model-Restriction Leaks, Mouse SelectionClaude Code's late-June v2.1.x updates (changelog through June 24-26) fix background sessions inheriting another session's ANTHROPIC_* provider environment variables, a 1-2s exit pause after interrupting a shell command on macOS/Linux, and org-configured model restrictions leaking into the model picker, --model, /model, and ANTHROPIC_MODEL. Mouse-click selection was also added to fullscreen select menus. These are reliability fixes distinct from the earlier 2.1.191 /rewind release.
  16. 2026-06-26 / SKILLSUse programmatic tool calling so large tool results never enter the model's contextA key 2026 context lever is keeping bulky tool outputs out of the window entirely: have the agent call tools programmatically (code that processes results) and route heavy sub-tasks to isolated subagents that return only 1,000-2,000 token condensed summaries. The model reasons over distilled state, not raw payloads. This is how long-horizon agents avoid the token exhaustion that kills naive 'dump every result into the chat' designs.
  17. 2026-06-26 / SKILLSEnforce must-happen actions with hooks, not prompts — deterministic execution at near-zero context costAnything that must happen reliably (running a linter, blocking a dangerous command, firing a notification) belongs in a `settings.json` hook, which executes deterministically instead of depending on the model deciding to comply. Hooks add negligible context overhead and remove a whole class of 'the agent forgot to run the check' failures. The rule of thumb: use a skill for domain logic, a subagent for isolated work, and a hook to enforce a rule with code.
  18. 2026-06-26 / SKILLSKnow that custom output styles silently override Claude Code's default coding guidanceOutput styles inject system-prompt-level instructions, but custom ones replace (not augment) Claude Code's built-in coding behavior — a quiet footgun that degrades code quality if you forget. Reach for the built-in Proactive/Explanatory/Learning styles first, and use `--append-system-prompt` for one-off tone or formatting tweaks since it's additive and preserves default behavior. Only write a custom style when you genuinely want to override the defaults.
  19. 2026-06-26 / SKILLSKeep root CLAUDE.md under 200 lines and push path-specific guidance into subdirectory CLAUDE.md filesSubdirectory CLAUDE.md files load on-demand only when Claude reads a file beneath that folder, so deep module conventions never inflate context during unrelated work. Reserve the root file for build commands and team-wide conventions (target <200 lines) and let nested files carry the specifics. This turns the CLAUDE.md hierarchy into a lazy-loaded context tree instead of one fat always-on prompt.
  20. 2026-06-25 / SKILLSPackage repeated Claude Code prompts as SKILL.md skills with a config.json and an allowed-tools fenceAny instruction set you retype at the start of relevant sessions ('write tests in Vitest, prefer integration tests') should become a skill: a folder under ~/.claude/skills/<name>/ with a SKILL.md carrying YAML frontmatter (name, description) that Claude auto-discovers and invokes. Two underused refinements — store per-user setup in a config.json so the agent asks for missing values rather than guessing, and add `allowed-tools: Bash Read` to fence a skill out of file-writing as a cheap safety boundary. Because skills follow the cross-platform Agent Skills standard, the same packaged expertise carries across multiple coding agents.
  21. 2026-06-24 / SKILLSKnow the difference between compaction and agentic memory — and use both deliberatelyCompaction trims what's already in the context window (Anthropic productized a server-side compaction API on Opus 4.6 that auto-summarizes older turns into a compressed state mid-session), while agentic memory moves information entirely outside the context as structured notes the agent writes and pulls back on demand. Conflating the two leaves long-horizon agents either losing critical state to summarization or drowning in context they should have offloaded. Use compaction to manage window pressure and note-taking memory to preserve progress across dozens of tool calls and context resets.
  22. 2026-06-23 / SKILLSUse LLM-prompt 'agent hooks' to verify conditions deterministically before a tool runsBeyond shell-command hooks, Claude Code supports agent hooks that fire a subagent (able to read files, grep, and use tools) at lifecycle events to return an allow/deny decision — e.g., a PreToolUse hook that inspects a write and blocks it if it touches secrets. Unlike instructions in a prompt, hooks execute deterministic logic that can't be hallucinated away. Use them to enforce guardrails the model is otherwise free to ignore.
  23. 2026-06-23 / SKILLSUse plan-validate-execute with a verifiable intermediate file for any batch or destructive agent operationFor multi-item changes (e.g., updating 50 fields from a spreadsheet), have the agent first emit a structured plan file (changes.json), run a verbose validation script against it, and only then execute — turning silent mistakes into machine-checkable errors caught before anything is mutated. Make validators verbose ('Field signature_date not found. Available: …') so the agent can self-correct in a loop. This pattern makes open-ended agent work reversible and debuggable.
  24. 2026-06-23 / SKILLSBuild 3 evaluations and a no-Skill baseline BEFORE writing the Skill (evaluation-driven Skill development)Anthropic recommends running Claude on representative tasks without a Skill first, documenting the specific failures, then writing three eval scenarios and a baseline measurement — and only then authoring the minimal instructions needed to beat that baseline. This stops you from documenting imagined problems and keeps Skills lean. The companion workflow uses two Claude instances: 'Claude A' to author/refine and a fresh 'Claude B' to test, feeding observed gaps back to A.
  25. 2026-06-23 / SKILLSMatch a Skill's 'degrees of freedom' to task fragility — exact scripts for narrow bridges, prose for open fieldsAnthropic frames Skill authoring as setting freedom level: low freedom (one exact command, 'do not modify') for fragile sequential operations like migrations, medium freedom (parameterized pseudocode) where a preferred pattern exists, and high freedom (heuristic prose) where many paths succeed, like code review. Over-specifying open-ended tasks wastes tokens and handcuffs the model; under-specifying fragile ones invites silent breakage. Decide freedom level per section before writing instructions.
  26. 2026-06-23 / SKILLSKeep Skill file references exactly one level deep from SKILL.md — nested links cause partial readsWhen a reference file links to yet another file, Claude tends to preview with head -100 rather than read the whole thing, so deeply nested references (SKILL.md → advanced.md → details.md) silently deliver incomplete information. Link every reference file directly from SKILL.md so each is read in full when triggered, and add a table-of-contents to any reference over 100 lines so partial previews still reveal full scope. Keep the SKILL.md body under 500 lines and push detail to these one-hop files.
  27. 2026-06-23 / SKILLSWrite Skill descriptions in third person with explicit 'what + when' — it's the discovery key among 100+ skillsClaude selects a Skill purely from its name and description at startup, so the description must state both what it does and the specific triggers/contexts for using it, always in third person ('Extracts text and tables from PDFs… Use when…'), never first/second person, which causes discovery failures. Pair this with gerund-form names (processing-pdfs, analyzing-spreadsheets) for a consistent, greppable library. A vague 'helps with documents' description means the Skill silently never fires.
  28. 2026-06-23 / TOOLSTip: Set fallbackModel in Claude Code to Chain Up to Three Models for ResilienceRecent Claude Code releases added a fallbackModel configuration that tries up to three models in order when the primary is unavailable or rate-limited. For long-running autonomous loops this prevents a single provider hiccup from killing a session mid-task. Pair it with a cheap model in the chain so degraded runs continue instead of erroring out.
  29. 2026-06-23 / SKILLSAuto mode now sends subagent spawns through the permission classifier before launchA gap let a subagent request a blocked action without review because spawns weren't classified pre-launch; the June update closes it by evaluating every subagent spawn through the classifier before it runs. Combined with the new Tool(param:value) rules, you can now actually enforce policies like blocking Opus subagents in auto mode. The lesson for anyone running unattended agent fleets: verify your spawn-time guardrails, since pre-classification changes what auto mode will and won't permit.
  30. 2026-06-23 / SKILLSNested subagents can now spawn their own subagents — background trees capped at 5 levels deepClaude Code's June 2026 update lets a subagent spawn its own subagents, with a live panel showing the full descendant tree and a path back to main. Background trees are capped at five levels to prevent runaway concurrency; foreground chains are self-limiting at any depth. For builders this turns flat fan-out into true hierarchical delegation — a supervisor agent can decompose a task and let each child decompose further without you wiring a custom orchestrator.
  31. 2026-06-22 / SKILLSPin Claude Code version range and model allowlist via managed settingsTwo managed settings landed in June: `requiredMinimumVersion`/`requiredMaximumVersion` (v2.1.163) make Claude Code refuse to start outside an approved version band and point users to the right build, and `enforceAvailableModels` (v2.1.175) constrains the Default model to an `availableModels` allowlist, falling back to the first allowed model. For teams standardizing agent behavior across machines, these prevent silent drift from version or model differences. Set them in managed config rather than relying on everyone updating manually.
  32. 2026-06-22 / SOURCESClaude Code June Release: Doubled Rate Limits, Nested Skills, and Streamlined Agent TeamsAnthropic's June Claude Code updates doubled rate limits on Claude Code and raised Opus API limits, and shipped nested skills, streamlined agent teams, and tighter permissions/auto-mode review. The added headroom and nested-skills support matter for anyone running Claude Code heavily in automated agentic loops. Distinct from the already-covered Artifacts feature.
  33. 2026-06-21 / SKILLSGate subagents by parameter with Tool(param:value) permission rules (e.g. Agent(model:opus))Claude Code added parameter-matching to permission rules: Tool(param:value) syntax with * wildcards lets you allow or deny a tool based on its input arguments, not just its name. For example, Agent(model:opus) blocks any subagent that requests the Opus model, capping spend or enforcing a cheaper tier for delegated work. This is a precise, declarative cost-and-safety control that previously required custom wrapper logic.
  34. 2026-06-21 / SKILLSSpawn Claude Code teammates directly via the Agent tool's name param — no TeamCreate stepClaude Code's June 15, 2026 update removed the TeamCreate/TeamDelete tools; with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 set, every session now has one implicit team and you spawn named teammates directly through the Agent tool. This collapses the old two-step flow (create a team, then add members) into a single inline call. For builders running agent fleets, parallel teammates can now be launched mid-task without pre-declaring the team.
  35. 2026-06-20 / SKILLSBisect a broken agent config with --safe-mode and disableBundledSkillsWeek 24 adds --safe-mode (or CLAUDE_CODE_SAFE_MODE) to launch with all customizations off — CLAUDE.md, skills, plugins, hooks, MCP servers, and custom commands/agents do not load, while auth, model selection, built-in tools, and permissions still work. A companion disableBundledSkills setting hides bundled skills and built-in commands from the model. If a misbehavior vanishes in safe mode, you've isolated the offending surface in one step instead of disabling things one at a time.
  36. 2026-06-20 / SKILLSConfigure fallbackModel chains so overloaded-primary errors auto-retry on a backup modelWeek 24 lets you set fallbackModel (or --fallback-model, now also for interactive sessions) to a chain of up to three models tried in order when the primary is overloaded or unavailable. This turns a hard 'model overloaded' failure into a transparent degrade-and-continue, which matters for unattended/cron-driven agent runs that otherwise die on capacity blips. Pin a cheaper or different-provider model as the last link to guarantee the run finishes.
  37. 2026-06-18 / TOOLSTip: Enable sandbox.allowAppleEvents to Let Sandboxed Claude Code Drive macOS AppsThe new `sandbox.allowAppleEvents` opt-in lets commands running inside Claude Code's macOS sandbox send Apple Events — the scripting bridge that controls native apps (Finder, Safari, Mail, etc.). It's off by default for safety; turning it on enables AppleScript/osascript-style automation from within a sandboxed agent run without dropping the sandbox entirely, useful for desktop-automation workflows that previously required unsandboxed execution.
  38. 2026-06-18 / TOOLSClaude Code Subagent Panel Gets Auto-Hide and Scroll Caps for Long-Running FleetsThe June 17 changelog also reworks the subagent panel: idle subagents auto-hide after 30 seconds, the list caps at 5 rows with scroll hints, and keyboard hints now render in the footer. Streaming of long paragraphs was changed to appear line-by-line instead of buffering to the first break. These are quality-of-life fixes aimed squarely at sessions running many parallel subagents, where the panel previously became unreadable.
  39. 2026-06-18 / TOOLSTip: Point CLAUDE_CLIENT_PRESENCE_FILE at a Marker to Auto-Silence Mobile Push While at the KeyboardClaude Code's new `CLAUDE_CLIENT_PRESENCE_FILE` env var suppresses mobile push notifications whenever the marker file it points to exists, so you only get phone alerts when you've actually stepped away. Wire the file's presence to your terminal focus (or a tmux/IDE hook) to avoid duplicate notifications during active sessions — a small but real ergonomics win for the increasingly common laptop-plus-phone monitoring setup.
  40. 2026-06-18 / TOOLSTip: Use /config key=value to Flip Any Claude Code Setting Without Leaving the PromptThe new `/config key=value` syntax lets you toggle any setting mid-session — e.g. `/config thinking=false` to disable extended thinking for a quick task, then flip it back. Because it works in `-p` and Remote Control, you can script per-invocation overrides instead of maintaining separate settings files. Pairs well with the new `CLAUDE_CLIENT_PRESENCE_FILE` marker to silence phone alerts while you're actively driving the terminal.
Open latest cited source