Anthropic Claude Code CLI

Scope

• System: Anthropic Claude Code
• Role: Harness
• Interface/surface: Claude Code CLI, including native settings/hooks and the CLI --print automation surface
• Version: 2.1.282
• Release date: September 24, 2026
• Deployment mode: local CLI process
• Reviewed at: September 25, 2026
• Evidence status: first-party documentation and release notes; live Harness Operations tests pending

Claude Agent SDK, Claude Desktop, hosted/cloud sessions, remote-control-specific behavior, IDE-specific behavior, and the redesigned Claude Projects beta are distinct surfaces and are not silently folded into this row. See Claude Projects (beta) for the coordinator/thread orchestration surface.

Primary sources

• https://github.com/anthropics/claude-code/releases/tag/v2.1.282
• https://code.claude.com/docs/en/cli-reference
• https://code.claude.com/docs/en/permissions
• https://code.claude.com/docs/en/hooks
• https://code.claude.com/docs/en/features-overview
• https://code.claude.com/docs/en/env-vars
• https://code.claude.com/docs/en/commands
• https://code.claude.com/docs/en/security
• https://code.claude.com/docs/en/monitoring-usage

Native operational concepts

Native concept	Native meaning	Harness Operations relationship	Notes
Session	Claude Code conversation/execution context with a session ID and persisted transcript	Agent Session; may participate in a broader Run	Preserve native session identity and resume semantics.
Tool request	Proposed Bash/file/MCP/etc. action	Capability use inside an Agent Session	Permission evaluation can gate the call before execution.
Permission rule/mode	Declarative allow/deny/ask behavior for tool use	Policy + execution control	Not identical to organizational Authority.
PreToolUse hook	External hook invoked before matching tool use and able to block/modify decisions	Enforcement/control point	Stronger than prompt-only instructions.
Subagent	Isolated worker with its own context that reports to the main agent	Delegated Agent Session/work unit	Native Claude Code semantics remain authoritative.
Agent team	Experimental independent Claude Code teammates with shared tasks/message exchange	Multi-session coordination	Experimental and disabled by default.
Sandbox	Optional OS-level sandbox mode on supported platforms	Execution Environment / enforcement mechanism	Distinct from permission prompts.
Transcript / OTel	Persisted session history and optional telemetry export	Evidence inputs	Retention/redaction configuration matters.

Lifecycle

Claude Code exposes stable session identifiers and resume/continue behavior. The CLI supports interactive sessions and a non-interactive --print mode for scripting/automation.

The current release also has explicit current-turn interruption behavior. Interruption is not rollback: already completed tool or external side effects may remain.

Coordination

Claude Code supports native subagents for isolated delegated work. Agent teams provide a broader multi-session coordination model with peer-to-peer communication and shared task coordination, but Anthropic documents them as experimental and disabled by default.

This reference entry treats native subagent delegation as available without claiming A2A or cross-vendor handoff compatibility.

Governance and enforcement

Claude Code’s permission evaluation includes hooks, deny rules, permission modes, allow rules, and interactive/runtime decisions. PreToolUse hooks and deny rules can prevent matching tool execution before the side effect.

This distinction matters operationally: CLAUDE.md or a skill saying “do not do X” is model guidance, while a blocking PreToolUse hook or permission rule is an execution control.

The scoped public docs do not establish durable named-human Principal attribution for every approval decision, so attributed decision remains partial.

Isolation and credentials

Claude Code documents optional OS-level sandboxing for shell execution on supported platforms. Permission modes and sandboxing are distinct controls.

Credential mediation remains unknown under the comparison definition’s strong definition. Claude Code supports several authentication/credential configuration paths, but Anthropic also warns that a tool reading a credential file or printing a secret can cause that value to appear in session history. That is not sufficient evidence to claim credentials are consistently brokered away from model/tool context.

Resources

Claude Code documents usage/monitoring surfaces and OpenTelemetry export. It also documents hard configurable bounds such as maximum tool/subagent concurrency and output-token limits.

Those are genuine enforced limits, but they do not amount to a universal monetary or all-resource budget system.

Evidence

Claude Code persists session transcripts unless configured otherwise and exports session IDs to hooks/subprocesses. Optional OpenTelemetry can emit metrics/logging, and recent releases can emit managed-settings resolution evidence.

This is useful operational evidence but not enough to claim a complete Harness Operations Audit Record for all policy, identity, credential, environment, and external side-effect facts. Resolved execution facts are therefore partial.

Capability summary

Documentation-backed available findings:

• interactive CLI operation;
• machine-oriented --print automation;
• stable session identity;
• resume;
• interruption;
• native parallel work through subagents;
• native delegation to subagents;
• configured pre-execution permission/hook gate;
• optional sandbox isolation;
• usage/telemetry reporting;
• configured resource limits;
• session/telemetry event evidence.

Partial:

• attributed decision;
• resolved execution facts.

Unknown:

• credential mediation under the stronger comparison definition.

Further validation

Useful live tests include:

1. start session and capture session ID;
2. resume the same session and verify continuity;
3. run non-interactive --print with structured output;
4. configure a PreToolUse hook that denies a disposable write/command and verify the side effect does not occur;
5. exercise sandbox behavior on a disposable path;
6. spawn a subagent and capture delegation/result behavior;
7. enable a safe OTel test collector or inspect a redacted local transcript and compare documented evidence fields.

Live-test evidence must be added separately from documentation evidence.

Limitations

This reference entry is descriptive, not certification or endorsement. Current documentation can be superseded by later Claude Code releases. A successful live test proves only the tested version/interface/configuration and operation.