10 New Features Devs Need to Know

The June 11 session persistence fix for Claude Code has already been widely documented. What has received far less attention is the broader set of Claude Code June 2026 features shipped alongside it and in the weeks following. These capabilities extend Claude Code’s single-agent coding assistant model with multi-agent orchestration primitives.

Table of Contents

Prerequisites: The configurations in this article assume Claude Code CLI ≥ 2026.6 (the June 2026 release train). Run claude --version to confirm your installed version before using any of the commands or configuration schemas below. Features, flags, and YAML/JSON schemas are version-specific and may not work on earlier or later releases.

Nested sub-agents, fallback model chains, a community tool marketplace, per-agent cost attribution, and scoped permissions are now available, though many remain scattered across changelogs. This article covers all 10 features with configurations based on the June 2026 CLI release that developers can adapt. Verify compatibility with claude --version before use.

1. Nested Sub-Agents with 3-Level Depth

How Layered Task Decomposition Works

Claude Code now supports hierarchical agent spawning where a parent agent can create child agents, and each child can spawn its own children, up to three levels deep. This architecture enables layered task decomposition for operations spanning multiple modules or files. A top-level agent handling a large codebase migration, for instance, can delegate to module-level agents, which in turn delegate to function-level agents for targeted rewrites. You can assign each agent in the tree a distinct role, model, and constraint set, and each operates with its own context window.

Configuration and Implementation

Define the hierarchy in .claude/agents.yaml:

max_depth: 3
agents:
  - agent_role: migration_coordinator
    description: "Top-level agent for Python 2 to 3 migration"
    model: claude-sonnet-4
    delegate_to:
      - agent_role: module_migrator
        description: "Handles per-module migration tasks"
        model: claude-sonnet-4
        delegate_to:
          - agent_role: function_refactorer
            description: "Rewrites individual functions for Python 3 compatibility"
            model: claude-haiku

Invoke the top-level agent and observe the delegation chain in real time:

claude agent run migration_coordinator --verbose-agents

The --verbose-agents flag outputs each time one agent delegates to another, including which child agent it spawned, what role it assigned, and the task fragment it received. This is essential for debugging hierarchies where failures at lower levels can be difficult to trace without structured visibility.

You can assign each agent in the tree a distinct role, model, and constraint set, and each operates with its own context window.

2. fallbackModel Configuration

Setting Up Model Fallback Chains

Rate limits, transient model outages, and cost management all create scenarios where a single model target is insufficient. The fallbackModel configuration addresses this by allowing an ordered list of models that Claude Code attempts sequentially. If the primary model returns a rate-limit error or is unavailable, Claude Code automatically tries the next model in the chain.

Practical Configuration

Save the following as .claude/settings.json:

{
  "model": "claude-sonnet-4",
  "fallbackModel": [
    {
      "model": "claude-haiku",
      "maxTokens": 4096,
      "costCeiling": 0.25
    },
    {
      "model": "local/codellama-34b",
      "maxTokens": 4096,
      "costCeiling": 0.00,
      "timeoutSeconds": 30
    }
  ]
}

The fallbackModel array should list only alternative models. Do not repeat the primary model as a fallback entry, since a model that has already failed due to a rate limit or outage will not succeed on immediate retry. Each entry supports per-model maxTokens and costCeiling parameters, giving you per-model limits on output size and spend as the system falls through the chain. The costCeiling parameter sets the maximum spend (as a numeric dollar amount) allowed on that model before advancing to the next fallback. The local model option with a zero cost ceiling serves as a last-resort fallback that guarantees availability.

Local model support: The local/ prefix requires a running local inference server (e.g., Ollama). Configure the endpoint in your Claude Code settings before using local/-prefixed models. Without a running inference server, this fallback entry will fail.

Browsing, Installing, and Publishing Tools

Skip this section if your organization restricts third-party tool installation. For everyone else: the marketplace is a registry of community-contributed Claude Code tools, including linters, formatters, custom slash commands, and specialized refactoring utilities. It operates through the CLI:

claude marketplace search "react"



claude marketplace install @community/react-refactor


claude marketplace publish

⚠️ Security notice: Verify your authentication credentials before publishing (claude auth status). Community-published tools are not reviewed or guaranteed by Anthropic. Before installing community tools, review the tool’s source and permissions. Exercise the same caution you would with any third-party package from a public registry.

The publish command walks through a guided flow that packages the tool definition, validates its manifest, and submits it to the registry. You can invoke installed tools as slash commands or agent-callable tools within the current project.

4. Usage Attribution and Cost Tracking

Per-Agent and Per-Task Cost Breakdown

The new --attribution flag generates a detailed .claude/attribution.json file after each session, breaking down token usage, cost, and model selection by sub-agent, task, and timestamp. For teams billing AI usage to specific projects or clients, this provides the granularity needed for accurate chargebacks.

Reading Attribution Reports

The following is an illustrative example. Actual cost values will vary based on current Anthropic pricing for each model tier:

{
  "schema_version": "1.0",
  "session_id": "ses_550e8400-e29b-41d4-a716-446655440000",
  "total_cost": 0.87,
  "total_tokens": 142350,
  "agents": [
    {
      "agent_role": "migration_coordinator",
      "model": "claude-sonnet-4",
      "tokens_in": 12000,
      "tokens_out": 4500,
      "cost": 0.12,
      "timestamp": "2026-06-18T14:22:01Z",
      "tasks": ["coordinate migration plan"]
    },
    {
      "agent_role": "module_migrator",
      "model": "claude-sonnet-4",
      "tokens_in": 85000,
      "tokens_out": 32000,
      "cost": 0.63,
      "timestamp": "2026-06-18T14:23:15Z",
      "tasks": ["migrate auth module", "migrate payments module"]
    },
    {
      "agent_role": "function_refactorer",
      "model": "claude-haiku",
      "tokens_in": 6500,
      "tokens_out": 2350,
      "cost": 0.12,
      "timestamp": "2026-06-18T14:25:44Z",
      "tasks": ["refactor parse_token()", "refactor validate_session()"]
    }
  ]
}

Prompt tokens (tokens_in) and completion tokens (tokens_out) are tracked separately, which matters because pricing differs between the two. Each tasks array ties costs to specific work units, so you can spot a subtask that consumed, say, over half the session’s total cost and investigate why. Cost values are stored as numeric types (e.g., 0.87) for direct programmatic consumption without string parsing.

5. Scoped Permissions for Sub-Agents

Principle of Least Privilege in Agent Trees

You can now restrict child agents to specific directories, tools, and write capabilities. This prevents a function-level refactoring agent from accidentally modifying configuration files or invoking destructive tools outside its intended scope.

- agent_role: function_refactorer
  model: claude-haiku
  permissions:
    allow_paths:
      - "${PROJECT_ROOT}/src/utils"
    deny_tools:
      - shell_execute
      - file_delete
    can_write: true

The allow_paths field restricts the agent’s file system access. Use ${PROJECT_ROOT} to anchor paths to the project root (the directory containing .claude/), ensuring paths resolve to absolute canonical locations and cannot escape the intended scope via .. segments or symlinks. The deny_tools array explicitly blocks specific tools even if they are globally available. Tool names must match the identifiers listed in claude tools list. Setting can_write: false creates a read-only agent useful for analysis or review tasks.

⚠️ Caution: Sub-agents with can_write: true and broad allow_paths values can modify files across their entire permitted scope. Scope paths as narrowly as possible, especially in configurations where agents operate across multiple repositories.

6. Streaming Agent Logs

Beta feature. Behavior and schema may change in future releases. Not recommended for production pipelines without pinning your CLI version.

Pipe real-time, structured JSON logs from all active sub-agents into monitoring dashboards, CI/CD systems, or log aggregation platforms using the --stream-logs flag.

(Requires jq: brew install jq / apt install jq)

claude agent run migration_coordinator --stream-logs \
  | jq 'select((.agent_level | tonumber) <= 2 and .status == "error")'

Each log line is a JSON object containing fields for agent_role, agent_level, status, message, and timestamp. The jq filter above uses tonumber to safely coerce agent_level (which may be emitted as a string) before comparison, and narrows output to errors from the top two levels of the agent tree. This is useful for surfacing critical failures without drowning in verbose output from leaf agents.

7. Agent Checkpointing and Resume

Beta feature. Behavior and schema may change in future releases. Not recommended for production pipelines without pinning your CLI version.

Saving and Restoring Agent State Mid-Task

This feature is distinct from session persistence. Session persistence saves conversation history. Agent checkpointing additionally saves the entire agent tree state, including each sub-agent’s progress, intermediate outputs, and pending task queue. This allows a multi-hour migration to be paused and resumed without restarting from scratch.

claude checkpoint save migration-v2



claude checkpoint resume migration-v2

The checkpoint includes the pending task queue, so resumed sessions continue exactly where they left off rather than re-evaluating completed work.

Important: If files modified by the agent tree have changed externally since the checkpoint was saved, review the differences before resuming to avoid conflicts with stale intermediate state. Checkpoint files may contain sensitive intermediate outputs (code fragments, token context). Add the checkpoint storage directory to .gitignore and restrict directory permissions (e.g., chmod 700) to avoid committing or exposing this data.

8. Inline Cost Budgets per Agent

Hard dollar or token caps can be set on individual sub-agents to prevent runaway costs. When a budget is exceeded, the agent halts and reports an error to its parent.

- agent_role: function_refactorer
  model: claude-haiku
  budget:
    max_cost: 0.50
    max_tokens: 50000

When exceeded, the output reads: Error: Agent 'function_refactorer' exceeded budget. Used $0.51 of $0.50 limit. Escalating to parent agent. The parent agent can then decide whether to allocate additional budget, reassign the task, or abort. The max_cost value is a bare numeric dollar amount (not a string). Use 0.50 rather than "$0.50" to ensure the CLI correctly enforces the budget cap.

9. Custom Agent Templates

You can pre-define reusable agent configurations as templates, encapsulating role, model, permissions, and budget into a single file invocable by name. Templates are stored in a templates/ directory at the project root, and the templates_dir key in agents.yaml must point to this directory for --template to resolve correctly. Use ${PROJECT_ROOT}/templates to ensure correct resolution regardless of working directory.

agent_role: code_reviewer
description: "Reviews code for style, bugs, and security issues"
model: claude-sonnet-4
permissions:
  can_write: false
  deny_tools:
    - shell_execute
budget:
  max_cost: 0.25
  max_tokens: 30000

Run from the project root where templates/ exists. The --template flag accepts the base filename without extension:

claude agent run --template code-reviewer

This resolves to ./templates/code-reviewer.yaml. Templates standardize agent behavior across teams and projects, preventing inconsistencies like mismatched budget caps or permission scopes when multiple developers define similar agents independently.

10. Multi-Repo Orchestration

Beta feature. Behavior and schema may change in future releases. Not recommended for production pipelines without pinning your CLI version.

Sub-agents can now operate across multiple local repositories within a single session. This enables cross-repo tasks like synchronized dependency updates, API contract validation, or coordinated version bumps.

agents:
  - agent_role: dependency_updater
    model: claude-sonnet-4
    repos:
      - "${PROJECT_ROOT}/../frontend-app"
      - "${PROJECT_ROOT}/../backend-api"
      - "${PROJECT_ROOT}/../shared-libs"
    tasks:
      - "Update axios to 1.8.0 across all repos and fix breaking changes"

The repos array accepts ${PROJECT_ROOT}-relative paths to sibling directories. All listed repos must exist at the specified paths before invocation; missing paths will cause an error. Each sub-agent gains file system access scoped to the listed repositories, and the parent agent can coordinate changes that span repository boundaries.

⚠️ Caution: Multi-repo agents with write access can modify files across all listed repositories simultaneously. Test on copies or branches of your repositories before running destructive or large-scale operations.

Putting It All Together: Full Configuration Reference

A combined configuration using all 10 features:

Save the following as .claude/settings.json:

{
  "model": "claude-sonnet-4",
  "fallbackModel": [
    {
      "model": "claude-haiku",
      "maxTokens": 4096,
      "costCeiling": 0.25
    }
  ]
}

max_depth: 3
templates_dir: "${PROJECT_ROOT}/templates"
agents:
  - agent_role: migration_coordinator
    model: claude-sonnet-4
    budget:
      max_cost: 5.00
      max_tokens: 500000
    repos:
      - "${PROJECT_ROOT}/../backend-api"
      - "${PROJECT_ROOT}/../shared-libs"
    delegate_to:
      - agent_role: module_migrator
        model: claude-sonnet-4
        budget:
          max_cost: 2.00
          max_tokens: 200000
        permissions:
          allow_paths:
            - "${PROJECT_ROOT}/src"
          deny_tools:
            - file_delete
          can_write: true
        delegate_to:
          - agent_role: function_refactorer
            model: claude-haiku
            budget:
              max_cost: 0.50
              max_tokens: 50000
            permissions:
              allow_paths:
                - "${PROJECT_ROOT}/src/utils"
              deny_tools:
                - shell_execute
                - file_delete
              can_write: true

Solo developers should start with fallback models and cost budgets, which provide immediate value without organizational overhead. Teams should prioritize usage attribution and scoped permissions first, then layer in templates and multi-repo orchestration as workflow complexity grows.

Summary and What’s Next

The 10 Claude Code June 2026 features covered here:

• Nested sub-agents with up to 3-level depth for layered task decomposition
• fallbackModel configuration for resilient model chains
• Community tool marketplace for discovering and sharing tools
• Usage attribution with per-agent, per-task cost breakdowns
• Scoped permissions enforcing least privilege on sub-agents
• Streaming agent logs for real-time structured monitoring (beta)
• Agent checkpointing for saving and resuming agent tree state (beta)
• Inline cost budgets for per-agent spend caps
• Custom agent templates for reusable agent definitions
• Multi-repo orchestration for cross-repository tasks (beta)

Nested sub-agents, fallback models, and the marketplace are generally available as of the CLI 2026.6 release train. Agent checkpointing, streaming logs, and multi-repo orchestration remain in beta, and their schemas and behavior may change. Anthropic’s changelog tracks the latest status of each feature. For coverage of the session persistence fix that preceded this batch, see SitePoint’s prior article on that topic.