giasip-dispatch

Files (8)

• SKILL.md
• model-roster.md · references/
• api-dispatch.sh · scripts/
• codex-appserver.mjs · scripts/
• dispatch-persist.mjs · scripts/
• gemini-supervisor.sh · scripts/
• kimi-dispatch.sh · scripts/
• stop-review-gate.mjs · scripts/

SKILL.md (first lines)

> ✦ A **GiaSip** skill · part of the `giasip` toolkit · github.com/GiaSip

# /dispatch — Multi-Model Dispatcher

Sends a task or prompt to another AI model for execution and retrieves the result. This skill **only provides the dispatch capability** — which model to pick, whether to fan out to multiple models, is decided by you (or the current Claude) based on the task at hand. No built-in model preference.

## Script Directory

**Important**: All scripts live in the `scripts/` subdirectory of *this* skill folder.

**Agent setup — do this ONCE before running any command below**: determine the absolute path of the directory that contains this `SKILL.md`, and export it as `BASE_DIR`:

```bash
# Global install (most common); use the plugin cache path instead if installed as a plugin
export BASE_DIR="$HOME/.claude/skills/giasip-dispatch"
```

Every command below references scripts as `$BASE_DIR/scripts/<script-name>`. `BASE_DIR` is a shell variable **you** set in the session — there is **no** `CLAUDE_SKILL_DIR` environment variable injected by the runtime, so set `BASE_DIR` first or the script paths will resolve to nothing.

## Dispatch Channels

| Channel | Models | Prerequisite |
|---------|--------|-------------|
| **Aggregator API** ★ easy path — one key | Overseas: OpenRouter (DeepSeek/Qwen/GLM/Kimi/MiniMax + Claude/GPT/Gemini). China: SiliconFlow 硅基流动 (DeepSeek/Qwen/GLM/Kimi/MiniMax) | **One** key file — `openrouter.env` or `siliconflow.env` in `~/.config/ai-keys/`; set `DISPATCH_PROVIDER` once |
| **API direct call** (curl, per-vendor) | DeepSeek / Qwen / GLM / Doubao / MiniMax | A separate `.env` **per vendor** in `~/.config/ai-keys/` (advanced) |
| **CLI invocation** (agentic) | Codex / Gemini / Kimi | Local install + login per CLI — needed for agentic capability (code write, native vision) |
| **Internal SubAgent** | Claude Haiku / Sonnet | Built into Claude Code, no external dependency |

> **Easy path (recommended for most users):** grab **one** aggregator key and every alias your chosen provider supports works — no per-vendor signup. Overseas → OpenRouter; China → SiliconFlow. See README "giasip-dispatch — Dependencies" for setup.
>
> **When you still need the per-vendor / CLI paths:** direct per-vendor keys (if you already have them, or to avoid OpenRouter's ~5.5% credit-top-up fee); CLI channels for **agentic** work the chat API can't do — Codex write-mode (edits files), Gemini native PDF/image vision. Aggregators cover all pure-analysis + multi-dispatch use.
>
> For pure thinking/analysis tasks (no file I/O, no command execution, no code changes), prefer the **Aggregator / API direct** call — roughly 10x faster than CLI. Use CLI only when the task needs agent capabilities (file system access, command execution, code changes, native vision).

---

## General Discipline (must-read for all CLI calls)

1. **Use heredoc for prompts** — avoids quoting / special character issues.
2. **Append `</dev/null` to all CLI calls** — Claude Code's Bash environment has a never-closing stdin pipe; without this, CLIs hang waiting for input.
3. **When constructing prompts**: keep them concise and clear; don't inject Claude Code-specific concepts (skills / hooks / SubAgent); use Chinese prompts for Chinese tasks, English for English; include necessary project context when relevant.

---

## Complexity Routing

Automatically select the dispatch strategy based on task nature:

```
Visual task? (PDF catalog / scanned doc / screenshot / image parsing)
├── Yes → [Gemini CLI] — native PDF + image visual analysis
│   Trigger: user provides image/PDF path + task involves "reading" content;
│   or markitdown output is empty/abnormally small
│   → Do NOT try markitdown first — route directly to Gemini
│
├── Pure thinking/analysis (business analysis, strategy, text understanding, translation)
│   → [API direct call] — no agent capability needed, curl the API directly
│   10x faster; supports models without CLI (DeepSeek/Qwen/GLM etc.)
│
├── Simple + reversible (install packages, format, search, generate templates)
│   → [Single dispatch] — prefer Claude SubAgent (Haiku) for speed
│   Use external CLI only when the task needs Chinese-native or specific AI capabilities
│
├── Code execution (bug fix, write tests, small refactor, generate code files)
│   → [Codex write mode] — sandbox=full, Codex modifies code directly
│   Condition: clear task, in a git-managed project, controllable blast radius
│   Always remind user to `git diff` afterward
│
├── Medium complexity (feature dev, document analysis, code analysis)
│   → [Single dispatch] — send to the best-fit AI
│
├── Complex + irreversible (architecture design, tech selection, major refactor)
│   → [Multi-dispatch] — auto-escalate, send to 2-3 AIs, compare outputs
│
└── User specifies AI ("run this with Kimi")
    → [Direct assignment] — send to specified AI, skip matching
```

**Multi-dispatch triggers** (any one auto-escalates):
- User says "important", "critical decision", "can't be wrong"
- Task involves irreversible operations (database migration, production deployment)
- Task is tech selection or architecture design
- Estimated blast radius > 10 files or 3 modules
- Chinese business/strategic analysis tasks (default: three-way parallel)

→ See `references/model-roster.md` for multi-dispatch lineup recommendations by task type.

---

## Dispatch Methods

### Aggregator API — ★ easy path (one key, recommended)

The same `api-dispatch.sh` routes through an aggregator when you set a provider. **One** key unlocks all the models below — no per-vendor signup.

```bash
# Set once per session (overseas → openrouter, China → siliconflow), then every supported alias works
export DISPATCH_PROVIDER=openrouter        # or: siliconflow

$BASE_DIR/scripts/api-dispatch.sh --model deepseek "$(cat <<'EOF'
prompt content
EOF
)"

# Per-call override without touching the env
$BASE_DIR/scripts/api-dispatch.sh --via siliconflow --model kimi "prompt"

# Escape hatch — pass any aggregator model ID verbatim (alias table can't cover everything)
$BASE_DIR/scripts/api-dispatch.sh --via openrouter --model-id anthropic/claude-3.7-sonnet "prompt"
```

Provider resolution order: `--via` flag > `$DISPATCH_PROVIDER` env > `direct` (default).

| Provider | Key File (`~/.config/ai-keys/`) | Aliases | Note |
|----------|-------------------------------|---------|------|
| `openrouter` | `openrouter.env` (`OPENROUTER_API_KEY`) | deepseek / qwen / glm / kimi / minimax / **claude / gpt / gemini** | Overseas; needs a VPN in mainland China |
| `siliconflow` | `siliconflow.env` (`SILICONFLOW_API_KEY`) | deepseek / qwen / glm / kimi / minimax | China direct; open-source/domestic only (no Claude/GPT/Gemini). Intl users: `export SILICONFLOW_BASE_URL=https://api.siliconflow.com/v1` |

> Aggregator model IDs are the volatile part — alias → model-ID maps live in `references/model-roster.md`; if a call 404s on the model, verify on the vendor's models page or pass the correct ID via `--model-id`.

From GiaSip/giasip-skills, MIT.

More from GiaSip/giasip-skills