| name | using-cline-cli |
| description | Operate Cline CLI safely: core concepts, sandbox rules, provider/model selection, approvals, and common workflows |
| when_to_use | when setting up or driving Cline from the CLI, switching models/providers, handling sandbox/approvals, or troubleshooting blocked commands |
| version | 1.0.0 |
| allowed-tools | ["Read","Write","Execute","Search"] |
| compatibility | {"runtimes":["Codex CLI"],"notes":"assumes workspace-write sandbox; approvals required for escalated commands or network installs"} |
Using Cline CLI
Use this skill to run Cline effectively from the CLI, keep commands safe under sandboxing, and pick inference providers/models.
Core Concepts
- Roles: System/dev instructions set guardrails; user asks; you execute within sandboxed shell unless approved otherwise.
- Sandbox: Default
workspace-write; writing outside writable roots or network installs needs approval (escalated run).
- Approvals: Destructive ops (
rm -rf, git reset --hard), networked installs, GUI apps, or writes outside roots require sandbox_permissions: require_escalated + short justification in tool call.
- Tools:
Execute runs shell commands; keep workdir explicit; prefer rg for search; avoid cd chains—set workdir instead.
- Memory/Context: Keep responses concise; reference paths with backticks; no raw URLs unless asked; cite sources when web-searching.
Quick Commands
- List files:
ls, ls -la.
- Search text fast:
rg pattern path.
- View file:
sed -n '1,160p' file.
- Edit file: use
apply_patch for targeted edits; avoid for autogenerated files.
- Run tests/linters:
npm test, pytest, etc.—ensure workdir set; mention if sandbox blocks writes.
Provider / Model Selection
- Prefer repository or skill-specified model; fall back to default if unavailable.
- If switching providers (e.g., OpenAI ↔ Anthropic ↔ local), note token/env var needs; avoid changing unless requested.
- Be explicit when using non-default model; mention trade-offs (speed vs. quality) briefly.
Working with Tabs and Files
- Respect active/open tabs context; address the active file first when relevant.
- Reference files as
path:line (no ranges, no URIs).
- Don’t dump large files; show relevant slices.
Safety Rules
- Never run destructive git commands unless explicitly asked.
- Do not revert user changes you didn’t make.
- For commands that fail due to sandbox, retry with approval parameters instead of asking first.
- No network installs unless required and approved.
Common Workflows
Edit flow
- Inspect file (
sed -n), 2) Plan change (brief), 3) apply_patch, 4) Re-open snippet to verify.
Search flow
rg 'pattern' src/, 2) narrow with rg -n, 3) edit.
Run/check flow
npm test / pytest with workdir set, 2) summarize failures, 3) propose fixes.
Troubleshooting
- Command not found: ensure correct shell, check PATH; if missing tool, note and propose install (needs approval).
- Permission denied / sandbox: rerun with
sandbox_permissions: require_escalated + justification; keep scope minimal.
- Large outputs: truncate (
| head -n 200); summarize rather than paste.
- Model mismatch: if instructions require a model not available, state fallback used.
Agent Etiquette
- Be concise; avoid redundant summaries.
- Suggest next steps only when natural (tests, commits).
- Keep citations outside code fences; cite up-to-date info when web searching.
