| name | claude-runtime |
| description | Claude Code runtime/session control: use Claude Code slash commands and CLI flags to manage sessions, permissions, model selection, and restarts in a worker terminal. Trigger when the controller needs to operate the Claude Code CLI like a human. |
Claude Runtime
Overview
Operate the Claude Code CLI safely: manage session state, permissions, models, and restarts from a worker terminal.
Session Safety
- Confirm idle state
- Snapshot and/or
status the worker; do not intervene mid-run.
- Only proceed when the worker is at a prompt or explicitly idle.
- Prefer in-session resets
- Use
/clear to reset the current conversation.
- Use
/compact to summarize and shrink context.
- Full reset (only between tasks)
- Exit the CLI with
Ctrl+D (EOF) when permitted by controller rules/human request.
- Relaunch with appropriate flags when a fresh context is needed.
Core Slash Commands (Claude Code)
Use /help to see the full list. Common commands:
/model change the active model./permissions edit tool permissions./clear reset the current conversation./compact summarize the conversation./add-dir add a working directory./agents view or manage agents./config update configuration./memory edit memory files./mcp manage MCP servers/tools./cost show usage cost./doctor check installation health./login and /logout manage auth./init initialize in the current repo./bug submit a bug report./pr_comments read PR comments in the repo.
Resume / Continue (Shell)
claude -c / claude --continue continue the most recent conversation.claude -r <SESSION_ID> / claude --resume <SESSION_ID> resume a specific session.
Launch Flags (Shell)
Use these when starting Claude from the shell (not inside the CLI):
--model choose a model at launch.--permission-mode set tool permission mode.--add-dir grant additional directories.--dangerously-skip-permissions bypass permission prompts (use only when explicitly allowed).
Guardrails
- Do not restart mid-run.
- Prefer
/clear or /compact over full restarts when possible.
- Use
Ctrl+D only when allowed by controller rules or explicitly requested by the human.
- If the worker is not Claude Code, switch to the model-specific runtime skill instead.
