Giasip Dispatch

✦ 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:

# 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

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.

# 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).

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.

API Direct Call — per-vendor (advanced, no aggregator)

If you already hold per-vendor keys (or want to avoid the aggregator's credit-top-up fee), call each vendor directly. Requires a separate .env per vendor:

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

Long text via stdin:

echo "long text content" | $BASE_DIR/scripts/api-dispatch.sh --model <model> --stdin

Supported models — see references/model-roster.md for the full roster with per-model strengths and multi-dispatch lineup recommendations.

Model names evolve with vendor updates — check vendor docs before calling.

Codex CLI (OpenAI) — App Server protocol, no cold start

Read-only mode (analysis / review / research):

node $BASE_DIR/scripts/codex-appserver.mjs --effort xhigh "$(cat <<'EOF'
prompt content
EOF
)" </dev/null

Write mode (code changes / bug fixes / test generation / file creation):

node $BASE_DIR/scripts/codex-appserver.mjs --effort xhigh --sandbox full "$(cat <<'EOF'
prompt content
EOF
)" </dev/null

Write mode safety rules:

• Only use in git-managed project directories (--cwd /path/to/project)
• After execution, always remind the user to git diff to review changes; git checkout . to revert if unsatisfied
• For production code / databases / deployment scripts, downgrade to read-only mode + output recommendations instead

Notes:

• --sandbox defaults to read-only; write mode uses full (the script also supports workspace-write / danger-full-access for advanced use)
• The script has built-in non-ASCII path auto-symlink workaround (--cwd with CJK characters auto-creates a temp symlink in /tmp, cleaned up on exit)
• Do not add 2>/dev/null — the script outputs structured progress and error info on stderr
• Long text can use stdin: echo "long text" | node $BASE_DIR/scripts/codex-appserver.mjs --stdin --effort xhigh

Gemini CLI (Google) — supervisor script recommended

The supervisor has built-in smart retry / fallback chain / circuit breaker / timeout / logging:

$BASE_DIR/scripts/gemini-supervisor.sh --cwd "/path/to/work/dir" "$(cat <<'PROMPT_END'
prompt content
PROMPT_END
)"

Supervisor default behavior:

• Fallback chain: flagship model → GA stable → flash (auto-skips if unavailable)
• Error classification: 429 short-term congestion → backoff + jitter retry; daily quota exhaustion → skip to next model; 503 → same backoff path
• Global attempt budget: 6; per-model hard timeout: 600s (configurable via GEMINI_MODEL_TIMEOUT)
• Circuit breaker: 3 consecutive failures per model → 30-minute cool-down
• Logs: ~/.cache/dispatch/gemini.log (JSONL) + state: ~/.cache/dispatch/gemini-state.json

Specify a single model (skip fallback): $BASE_DIR/scripts/gemini-supervisor.sh --model <model-id> "prompt" stdin mode (recommended for long prompts): cat prompt.txt | $BASE_DIR/scripts/gemini-supervisor.sh --stdin --cwd "/work/dir"

Gemini vision / PDF parsing (Gemini natively supports PDF + image visual analysis — the standard path for scanned PDFs / screenshots):

$BASE_DIR/scripts/gemini-supervisor.sh \
  --cwd "/path/to/files/dir" \
  "$(cat <<'PROMPT_END'
Please fully parse all pages of xxx.pdf and output in markdown format:
- Preserve all data tables (use markdown table syntax)
- Preserve all specs, technical parameters, model numbers
- Annotate page numbers (## Page 1 / ## Page 2 ...)
- Do not omit any technical details
PROMPT_END
)" > output.md

Use cases: PDF catalogs (no text layer), scans, product datasheets, screenshot analysis, image OCR, chart data extraction. Pipe to file (... > output.md) for large outputs to avoid stdout truncation.

Kimi CLI (Moonshot) — wrapper script recommended, auto endpoint routing

Thinking model discipline: Kimi K2.6 is a thinking model — reasoning can take minutes for complex prompts. Bash timeout must be ≥600000 (10 min) for complex tasks. The script has built-in SSE streaming + idle guard (120s no-byte threshold) + 900s hard cap. Do NOT kill mid-run or substitute with hand-written curl. For fast mode: prefix KIMI_NO_THINK=1 (injects {"thinking":{"type":"disabled"}}, ~4s response, but quality drops — only for non-reasoning tasks).

# Default: Moonshot general endpoint (api.moonshot.cn/v1, MOONSHOT_API_KEY)
$BASE_DIR/scripts/kimi-dispatch.sh "$(cat <<'EOF'
prompt content
EOF
)"

# Opt-in coding endpoint (Kimi CLI + api.kimi.com/coding/v1, KIMI_API_KEY)
KIMI_FOR_CODING=1 $BASE_DIR/scripts/kimi-dispatch.sh "prompt"

# Fast mode (disable thinking, ~4s response)
KIMI_NO_THINK=1 $BASE_DIR/scripts/kimi-dispatch.sh "simple task"

Claude Code Internal SubAgent (no external CLI)

For simple tasks use Haiku, for standard tasks use Sonnet — spawn a SubAgent via the Agent tool (specify model: haiku or model: sonnet). Faster and cheaper than headless CLI, with no external dependencies.

Environment Check

Quick availability check before dispatching:

# CLI tools
command -v codex; command -v gemini; command -v kimi

# API keys (for API direct calls)
ls ~/.config/ai-keys/*.env 2>/dev/null

Only dispatch to available models. If a CLI is unavailable, fall back to API direct call or switch models; an existing API key file means that model is callable.

Multi-Model Parallel (cross-validation / multi-perspective)

When you need multiple models to give independent perspectives on the same question, fire multiple Bash calls in parallel and have Claude synthesize the results. Typical scenarios: important decisions, tech selection, pre-flight check before irreversible actions, low confidence in a single model's output.

# Three-way parallel example (same prompt to three models)
$BASE_DIR/scripts/kimi-dispatch.sh "analysis task" &
$BASE_DIR/scripts/api-dispatch.sh --model deepseek "analysis task" &
$BASE_DIR/scripts/api-dispatch.sh --model doubao "analysis task" &
wait

Selection principle: cognitive diversity > quantity — pick models with different training data / architecture to get genuinely different perspectives; always use each model's highest tier; control cost by controlling frequency, not by downgrading per-call quality.

Execution Parameters

• Bash timeout:
  • Codex deep reasoning (xhigh): 600000 (10 minutes, Bash tool ceiling, aligned with script default 600s)
  • Gemini single dispatch: 240000 (4 minutes); multi-dispatch per route 300000 (5 minutes)
  • Kimi (thinking model): single/multi-dispatch always ≥600000 (reasoning tail is long and unpredictable); only KIMI_NO_THINK=1 fast mode can use 240000
• Single dispatch = one Bash call; multi-dispatch = multiple Bash calls fired in parallel
• Codex uses App Server protocol with no cold start; xhigh deep reasoning typically takes 3-8 minutes
• Gemini / Kimi use CLI headless mode; keep 2>/dev/null

Fallback Chain

Primary channel fails (timeout / error)
→ Try alternative model (similar capability)
→ Alternative also fails
→ Downgrade to "recommendation only" mode: output suggested approach but don't execute, hand back to user

Output Format

Single dispatch:

## Task Result
**Executor:** [model name]  **Task:** [one-line recap]

### Result
[model output]

### Execution Info
- Duration: [X seconds]  Status: [success / partial / failed]

Multi-dispatch:

## Task Result (Multi-Dispatch)
**Task:** [one-line recap]

### [Model 1]'s Take
[3-5 key points]

### [Model 2]'s Take
[3-5 key points]

### Synthesis
- **Consensus:** [what all parties agree on]
- **Divergence:** [differences, with each party's position noted]
- **My judgment:** [Claude's independent assessment as the orchestrator]

Response Logging

Use dispatch-persist.mjs to persist complete first-hand responses to ~/.cache/dispatch/responses/YYYY/MM/DD/<response_id>.md (YAML frontmatter + prompt + response) and append to ~/.cache/dispatch/index.jsonl (consumption entry point). Pipe dispatch output to this script, or hook it into your dispatch scripts to avoid losing responses to volatile /tmp or session logs.

For multi-dispatch runs, set DISPATCH_BATCH_ID to group responses from the same batch:

batch_id="$(uuidgen)"
DISPATCH_BATCH_ID="$batch_id" $BASE_DIR/scripts/kimi-dispatch.sh "task" &
DISPATCH_BATCH_ID="$batch_id" $BASE_DIR/scripts/api-dispatch.sh --model deepseek "task" &
wait

Implementation: see $BASE_DIR/scripts/dispatch-persist.mjs.

Script Inventory

For installation, dependencies, and API key setup, see README.md.