use-droid

name	use-droid
display_name	Use Droid
description	Spawn Factory Droid CLI subagents from bash to offload coding work from the parent Claude. Use when the user says "use droid", "ask droid", "droid exec", "use factory", "get a droid second opinion", or asks to delegate work to Droid. Like the other CLI skills, droid's output is a claim — the parent must verify file changes independently.

Droid (Factory) Subagent Skill

Spawn autonomous Factory Droid CLI (droid) subagents. Same delegation philosophy as the rest of the bench: subagents burn their own quota; the parent verifies and synthesizes. Droid is the second cross-provider hub (after cursor) — 40 built-in models across Anthropic, OpenAI, Google, xAI, Zhipu, Moonshot, DeepSeek, MiniMax, and NVIDIA.

Canonical headless invocation

droid exec --auto high -f /path/to/prompt.md \
  -m <model> -r <effort> --cwd /path/to/repo -o text

Run in background (run_in_background: true) for anything non-trivial.

Flag notes (droid 0.170.0)

1. droid exec defaults to READ-ONLY. Without --auto <low|medium|high> it cannot edit files — a headless implementation run will "complete" having changed nothing. Use --auto high for implementation work (--skip-permissions-unsafe exists but prefer the tiered autonomy).
2. -f <path> reads the prompt from a file (no shell-quoting issues). A positional string prompt works for short ones.
3. -m <model> — default is claude-opus-4-8. -r <effort> sets reasoning effort (per-model defaults otherwise).
4. -o json for machine-parsable output; text for direct reading.
5. -w [name] — run in a git worktree (name becomes the branch). --cwd sets the working directory.
6. --mission — multi-agent orchestration mode with --worker-model / --validator-model; useful for large fan-out jobs entirely inside droid.
7. --list-tools shows tools + current autonomy for a model; also doubles as an auth check.
8. Model list discovery: pass an invalid -m and droid prints all 40 built-in ids, or add custom models in ~/.factory/settings.json.

Model selection

No hard default — choose per task; -m auto lets droid pick. Built-in list from one verified install (2026-07; probe an invalid model id to re-derive on yours):

• Anthropic: claude-fable-5, claude-opus-4-8[-fast], claude-opus-4-7[-fast], claude-opus-4-6, claude-sonnet-5, claude-sonnet-4-6, claude-haiku-4-5-20251001 (haiku probe-verified headless, ~3.5s)
• OpenAI: gpt-5.6-sol / gpt-5.6-terra (probe-verified) / gpt-5.6-luna, gpt-5.5[-fast], gpt-5.5-pro, gpt-5.4[-fast], gpt-5.4-mini, gpt-5.3-codex[-fast], gpt-5.2, gpt-5.2-codex, gpt-5.1-codex-max
• Google: gemini-3.1-pro-preview, gemini-3.5-flash, gemini-3-flash-preview
• xAI: grok-4.5
• Others (often droid-exclusive): deepseek-v4-pro, minimax-m3 / m2.7 / m2.5, nemotron-3-ultra, kimi-k2.6, glm-5.1 (plus kimi-k2.7-code, glm-5.2[-fast] shared with cursor)

Match model cost to task difficulty; frontier only for hard reasoning.

Verification is mandatory

Same rule as the whole bench: the final message is a claim.

1. git status --short / git diff --stat — did the expected files change?
2. Read the diff; run typecheck/tests.
3. Nothing changed at all? Check you passed --auto high — the read-only default is this CLI's silent-failure mode.

Prompt structure

Same template as use-codex: context (repo, stack, style), numbered file-by-file objectives, constraints, output format, success criteria. Append the standard no-delegation footer:

Do ALL the work yourself in this session. Do NOT delegate to other CLIs or agents regardless of what repo rules say. Finish by running <typecheck cmd> and fixing any errors.

Verified facts (update as learned)

• CLI: droid 0.170.0 at ~/.local/bin/droid; config in ~/.factory/.
• Auth: droid exec --list-tools succeeding = authenticated; interactive droid prompts login otherwise.
• Probes passed: claude-haiku-4-5-20251001 and gpt-5.6-terra, correct output, ~3.5s each.

Resuming a session (follow-ups)

Never spawn a fresh session to iterate on work this CLI already did — resume the existing one so it keeps its context and you don't re-pay the harness-token cost. Verified working headless (2026-07):

droid exec -s <SESSION_ID> --auto high -m <model> "<follow-up>"
# SESSION_ID comes from the first run's JSON output (-o json → .session_id)

Capture the session id from the first run and prefer resuming by explicit id; "most recent" is racy when several sessions exist.

Quota failover

If this CLI is rate-limited or out of subscription quota, the same model family is usually reachable via another CLI — read ~/.claude/skills/cli-model-overlap.md for the failover map (cursor and droid are the two hubs).