Agent View Ops

Operate the official Claude Code agent view as a local session-control surface for independent background Claude Code sessions. Keep the skill focused on claude agents, shell management commands, session state, isolation, and troubleshooting. [DOC]

Treat current Claude Code documentation as the behavior authority. If the installed Claude Code version is unknown or not locally verifiable, include coverage_gap: active Claude Code version not verified. [DOC][INFERENCIA]

Use these supporting resources as needed:

• references/official-agent-view.md - Command surface, feature gates, dispatch paths, states, isolation, storage, and shell controls. [DOC]
• references/decision-guide.md - When to choose agent view versus subagents, agent teams, dynamic workflows, worktrees, or background shell commands. [DOC][INFERENCIA]
• references/troubleshooting.md - Symptom-based recovery for unavailable agent view, stalled supervisor, macOS protected folders, failed sessions, and worktree cleanup. [DOC]
• assets/agent-view-runbook.md - Portable operator checklist for dispatch, monitoring, review, and closeout. [CONFIG]
• assets/session-state-matrix.json - Machine-readable state and command matrix for audits and examples. [CONFIG]

Inputs Expected

• User goal: explain, operate, plan, audit, or troubleshoot agent view.
• Local context when operating: Claude Code version, project directory, session IDs, desired model, permission mode, effort level, and isolation constraints.
• Risk posture: read-only monitoring, write-capable dispatch, production-impacting edits, parallel pull requests, or cleanup.
• Existing evidence when available: claude agents --json, claude daemon status, claude logs <id>, git worktree list, or error text.

Outputs Expected

• A concise runbook, command plan, explanation, or troubleshooting report.
• Explicit feature gate and coverage_gap notes for unverified Claude Code version or unavailable local state.
• Session-state interpretation using official states and commands.
• Isolation guidance covering directories, worktrees, permissions, and cleanup risk.
• Recommendation on agent view versus subagents, agent teams, dynamic workflows, or worktrees when the user is choosing an approach.

Procedure

Step 1 - Confirm Scope

Classify the request as one of: launch agent view, dispatch sessions, monitor sessions, reply or attach, troubleshoot, clean up, compare orchestration approaches, or script against JSON output. Do not conflate claude agents with these adjacent control surfaces: /agents is the interactive subagent library, /tasks is the in-session task engine for ordered work items, and /workflows runs dynamic multi-step workflows. Only claude agents and its shell commands operate independent background sessions; the others coordinate work inside one session. [DOC]

Step 2 - Check Feature Gates

For current operations, run or request claude --version when feasible. Agent view is research preview and requires Claude Code v2.1.139 or later. claude agents --cwd <path> requires v2.1.141 or later, and agent-view configuration flags such as --settings, --add-dir, --plugin-dir, and --mcp-config require v2.1.142 or later. Voice dictation in the dispatch or reply input requires v2.1.145 or later with voice dictation enabled, and parallel work-item counts in summaries and peek require v2.1.161 or later. [DOC]

Step 3 - Choose The Control Surface

Use claude agents for an interactive terminal dashboard. Use claude agents --json for scripting, optionally with --all and --cwd <path>. Use claude attach <id>, claude logs <id>, claude stop <id>, claude respawn <id>, claude rm <id>, and claude daemon status for shell control without opening the dashboard. [DOC]

Step 4 - Dispatch Safely

Dispatch from agent view when each prompt should start a new independent background session. Use Shift+Enter only when immediate attach is needed. Use claude --bg "<task>" from the shell for direct background launch, and claude --bg --exec '<command>' or ! <command> in agent view for a background shell job rather than an agent session. Know the dispatch syntax: @<repo> runs the session in a sibling repository, while #<number> or a pull request URL selects an existing session already working on that PR instead of starting a new one. [DOC]

Set defaults deliberately: --permission-mode, --model, --effort, and --agent affect sessions dispatched from agent view. Prefer plan or narrow permissions for uncertain write work; avoid broad write-capable parallelism unless worktrees or clear file ownership boundaries exist. [DOC][INFERENCIA]

Step 5 - Monitor And Intervene

Interpret row states before acting: working, needs input or blocked, idle, completed, failed, and stopped. Use Space for peek, type a reply and press Enter to answer from the peek panel, and use Enter or right arrow to attach for the full conversation. Detach without stopping the session; stop explicitly when the work should end. [DOC]

Step 6 - Protect Isolation

Filter by project with --cwd when needed. Remember that the default view can list background sessions across projects. Treat Claude-created worktrees as disposable only after merging, pushing, or otherwise preserving useful changes. Before deleting a session that edited files, inspect its result, PR status, or worktree path. [DOC][INFERENCIA]

Step 7 - Troubleshoot From Evidence

Load references/troubleshooting.md when an error message or surprising behavior appears. Prefer local diagnostics such as claude daemon status, /doctor, claude logs <id>, git worktree list, and exact error text over speculation. Mark unknown version, provider-specific behavior, and inaccessible protected folders as coverage_gap. [DOC][INFERENCIA]

Quality Criteria

• The output identifies the correct control surface: claude agents, /agents, /tasks, /workflows, shell commands, or worktrees.
• Feature gate and research-preview status are stated when advising current operations.
• Session states and row groupings are not confused.
• Peek, reply, attach, detach, stop, respawn, remove, and logs are used with correct intent.
• Directory, worktree, and cleanup risks are explicit before destructive cleanup.
• Permission mode, model, effort, settings, plugin, MCP, and --add-dir defaults are treated as dispatch/session configuration, not global truth.
• Agent view is recommended only for independent background sessions; alternatives are named when coordination or context isolation requires them.
• Residual risks and coverage_gap items are visible.

Limits

• Do not claim agent view availability without version and setting evidence.
• Do not treat research-preview behavior as stable across versions.
• Do not delete sessions or worktrees without preserving desired work.
• Do not present agent teams, subagents, or dynamic workflows as the same surface as claude agents.

Contract

• Aceptación: sesiones de fondo monitoreables/adjuntables descritas con su ciclo de vida. [EXPLICIT]
• Límites: superficie experimental; confirmar antes de afirmar disponibilidad. [EXPLICIT]
• Casos borde: no disponible en todos los terminales (split-pane excluye algunos). [EXPLICIT]
• Supuestos: feature-gated/research-preview [DOC]. [SUPUESTO]
• Trade-off: sesiones full vs subagente — más visibilidad, más costo de supervisión. [EXPLICIT]

Packet

Capas del packet, cargables bajo demanda (disciplina ICM: una capa por vez, nunca todas juntas): references/ guías de profundidad (cargar UNA por etapa) · knowledge/ cuerpo de conocimiento · prompts/ prompts listos · examples/ salida de ejemplo · agents/ subagentes del packet · templates/ plantilla de output · assets/ recursos estáticos.