\r \r

dlazy-start — Quickstart for AI Agents\r

\r English · 中文\r \r A minimal contract for AI orchestrators using @dlazy/cli. The CLI is a\r tool-dispatch surface: every registered cloud + local tool becomes a top-level\r subcommand. There is no built-in project workspace or pipeline state machine —\r those are agent-side concepts.\r \r License: AGPL-3.0-or-later.\r \r

What this skill teaches\r

\r You drive @dlazy/cli from auth through tool invocation:\r \r

• Cloud tools (40+) — image / video / audio / text providers (Seedream,\r Recraft, MJ, Veo, Seedance, Kling, ElevenLabs, …)\r
• Local tools (40+) — state_lock_profile, video_compose,\r post_render_gate, scene_detect, frame_sampler, audio_mixer,\r audio_probe, transcribe, subtitle, color_grade, extract_segment,\r ffmpeg_run, … (full list via dlazy tools list)\r
• CLI commands: auth, doctor, tools list, tools describe,\r status, plus one top-level subcommand per registered tool.\r \r ---\r \r

Phase 0 — Install & auth\r

\r

# Install once\r
npm install -g @dlazy/cli\r
\r
# Authenticate (device-code flow; works in remote shells)\r
dlazy auth login\r
```\r
\r
Alternate auth: `dlazy auth set YOUR_API_KEY`, or set the `DLAZY_API_KEY`\r
env var. Config lives at `~/.dlazy/config.json` (Windows:\r
`%USERPROFILE%\.dlazy\`).\r
\r
Global flags every command accepts: `--api-key`, `--base-url`, `--verbose`,\r
`--output \x3Cjson|url|text>`, `--refresh-manifest`, `-l/--lang \x3Clocale>`.\r
\r
---\r
\r
## Phase 1 — Discover capabilities\r
\r
```bash\r
dlazy --help                         # top-level command surface\r
dlazy tools list                     # registered tools with type + cost shape\r
dlazy tools describe \x3Cname>          # input/output JSON schema, hasCosts, examples\r
```\r
\r
Optional local runtimes need a one-time install:\r
\r
```bash\r
dlazy doctor remotion                # report Remotion composer state\r
dlazy doctor remotion --install      # ~50s, installs the bundled composer\r
\r
dlazy doctor yt-dlp --install        # for video_downloader on YouTube et al.\r
dlazy doctor yt-dlp --install --proxy http://127.0.0.1:1087\r
```\r
\r
Some sandboxes restrict the tool surface via `DLAZY_DISABLED_TOOLS=\x3Ccsv>`;\r
disabled tools are hidden from `dlazy --help` and refuse invocation with a\r
clear `tool_disabled` error.\r
\r
---\r
\r
## Phase 2 — Invoke a tool\r
\r
Every tool is a top-level subcommand:\r
\r
```bash\r
# Inline flags (mirrors the input schema)\r
dlazy gpt-image-2 --prompt "cyberpunk cat at dusk"\r
\r
# JSON input file (preferred for complex shapes)\r
dlazy video_compose --input @work/compose.json --output json\r
\r
# Dry-run for validation only (no remote call, no credit consumption)\r
dlazy seedance-2-0 --input @plan.json --dry-run\r
```\r
\r
Per-tool help is generated from the schema:\r
\r
```bash\r
dlazy \x3Ctool-name> --help\r
```\r
\r
Output modes:\r
\r
- `--output text` (default) — human-readable envelope\r
- `--output json` — single JSON line per call; parse with `jq`\r
- `--output url` — bare URL when the tool produces a single asset\r
\r
---\r
\r
## Phase 3 — Poll async cloud tasks\r
\r
Long-running generations return a `generateId` instead of the final asset:\r
\r
```bash\r
dlazy status \x3CgenerateId>\r
dlazy status \x3CgenerateId> --output json\r
```\r
\r
Repeat until status is `succeeded` (then the asset URL is in the payload) or\r
`failed` (with `error.code` + `error.message`).\r
\r
---\r
\r
## Phase 4 — Common failure recovery\r
\r
**`dlazy doctor remotion --install` fails on `npm install`:**\r
- Check Node ≥ 18 (`node --version`).\r
- Behind a corp proxy: set `npm_config_proxy` / `npm_config_https_proxy`.\r
\r
**`video_downloader` returns "Sign in to confirm you're not a bot":**\r
- YouTube anti-bot challenge. Pass `"cookies_from_browser": "chrome"`\r
  (or firefox / safari / edge) in the input JSON.\r
\r
**`video_compose` returns "render_runtime=hyperframes not yet implemented":**\r
- HyperFrames runtime not shipped. Switch `edit_decisions.render_runtime` to\r
  `remotion` or `ffmpeg`, then re-validate via `pre_render_validator`.\r
\r
**ElevenLabs STT returns an empty `words` array:**\r
- Pass `timestamps_granularity: "word"` explicitly.\r
\r
**Need to know a tool's cost before invoking:**\r
- `dlazy tools describe \x3Cname>` exposes `hasCosts` and the cost shape. Log\r
  the estimate to a local file or your audit log before calling the tool.\r
\r
**Unknown command suggestion:**\r
- `dlazy \x3Ctypo>` returns `error: unknown command '\x3Ctypo>'` plus a "Did you\r
  mean …?" suggestion line based on edit distance. Trust the suggestion only\r
  after confirming via `dlazy tools list`.\r
\r
---\r
\r
## Anti-patterns\r
\r
- Calling a tool whose existence you haven't verified via `dlazy tools list`.\r
- Memorizing provider names from prior sessions instead of re-checking the\r
  registry (tools come and go).\r
- Silently swapping render runtime mid-pipeline (govern via `state_lock_profile`\r
  + `post_render_gate` parity checks instead).\r
- Calling paid generation without announcing provider / model / cost first.\r
\r
---\r
\r
## Reference card\r
\r
```\r
INSTALL       npm install -g @dlazy/cli && dlazy auth login\r
DISCOVER      dlazy tools list  |  dlazy tools describe \x3Cname>\r
LOCAL RT      dlazy doctor remotion --install   (or yt-dlp)\r
INVOKE        dlazy \x3Ctool-name> --input @file.json --output json\r
DRY RUN       dlazy \x3Ctool-name> --input @file.json --dry-run\r
POLL          dlazy status \x3CgenerateId>\r
HELP          dlazy --help  |  dlazy \x3Ctool-name> --help\r
RECOVER       dlazy doctor \x3Ctarget>  |  dlazy tools describe \x3Cname>\r
```\r