智能体上手手册 Start

智能体上手手册 Start

English · 中文

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

License: AGPL-3.0-or-later.

What this skill teaches

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

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

---

Phase 0 — Install & auth

# Install once
npm install -g @dlazy/cli

# Authenticate (device-code flow; works in remote shells)
dlazy auth login

Alternate auth: dlazy auth set YOUR_API_KEY, or set the DLAZY_API_KEY env var. Config lives at ~/.dlazy/config.json (Windows: %USERPROFILE%\.dlazy\).

Global flags every command accepts: --api-key, --base-url, --verbose, --format , --refresh-manifest, -l/--lang .

---

Phase 1 — Discover capabilities

dlazy --help                         # top-level command surface
dlazy tools list                     # registered tools with type + cost shape
dlazy tools describe           # input/output JSON schema, hasCosts, examples

Optional local runtimes need a one-time install:

dlazy doctor remotion                # report Remotion composer state
dlazy doctor remotion --install      # ~50s, installs the bundled composer

dlazy doctor yt-dlp --install        # for video_downloader on YouTube et al.
dlazy doctor yt-dlp --install --proxy http://127.0.0.1:1087

Some sandboxes restrict the tool surface via DLAZY_DISABLED_TOOLS=; disabled tools are hidden from dlazy --help and refuse invocation with a clear tool_disabled error.

---

Phase 2 — Invoke a tool

Every tool is a top-level subcommand:

# Inline flags (mirrors the input schema)
dlazy gpt-image-2 --prompt "cyberpunk cat at dusk"

# JSON input file (preferred for complex shapes)
dlazy video_compose --input @work/compose.json --format json

# Dry-run for validation only (no remote call, no credit consumption)
dlazy seedance-2-0 --input @plan.json --dry-run

Per-tool help is generated from the schema:

dlazy  --help

Output modes:

• --format json (default) — machine-readable envelope; parse with jq
• --format url — bare URL when the tool produces a single asset
• --format text — human-readable text payload
• --save — download the asset straight to disk (mkdir + retry handled for you)

---

Phase 3 — Poll async cloud tasks

Long-running generations return a generateId instead of the final asset:

dlazy status 
dlazy status  --format json

Repeat until status is succeeded (then the asset URL is in the payload) or failed (with error.code + error.message).

---

Phase 4 — Common failure recovery

dlazy doctor remotion --install fails on npm install:

• Check Node ≥ 18 (node --version).
• Behind a corp proxy: set npm_config_proxy / npm_config_https_proxy.

video_downloader returns "Sign in to confirm you're not a bot":

• YouTube anti-bot challenge. Pass "cookies_from_browser": "chrome" (or firefox / safari / edge) in the input JSON.

video_compose returns "render_runtime=hyperframes not yet implemented":

• HyperFrames runtime not shipped. Switch edit_decisions.render_runtime to remotion or ffmpeg, then re-validate via pre_render_validator.

ElevenLabs STT returns an empty words array:

• Pass timestamps_granularity: "word" explicitly.

Need to know a tool's cost before invoking:

• dlazy tools describe exposes hasCosts and the cost shape. Log the estimate to a local file or your audit log before calling the tool.

Unknown command suggestion:

• dlazy returns error: unknown command '' plus a "Did you mean …?" suggestion line based on edit distance. Trust the suggestion only after confirming via dlazy tools list.

---

Anti-patterns

• Calling a tool whose existence you haven't verified via dlazy tools list.
• Memorizing provider names from prior sessions instead of re-checking the registry (tools come and go).
• Silently swapping render runtime mid-pipeline (govern via state_lock_profile
  • post_render_gate parity checks instead).
• Calling paid generation without announcing provider / model / cost first.

---

Reference card

INSTALL       npm install -g @dlazy/cli && dlazy auth login
DISCOVER      dlazy tools list  |  dlazy tools describe 
LOCAL RT      dlazy doctor remotion --install   (or yt-dlp)
INVOKE        dlazy  --input @file.json --format json
DRY RUN       dlazy  --input @file.json --dry-run
POLL          dlazy status 
HELP          dlazy --help  |  dlazy  --help
RECOVER       dlazy doctor   |  dlazy tools describe