智慧AI工作流

⚡️ Quick Start

Set up LLM provider (uses LiteLLM in this example)

export LITELLM_API_URL=http://localhost:4401/v1 export LITELLM_API_KEY=sk-litellm-master-key

3. Try the example agents

```bash

Interactive example picker

./target/debug/chidori demo

🧪 Examples

See examples/:

• agents/hello.ts — minimal agent, no LLM
• agents/summarizer.ts — LLM summary pipeline
• agents/streaming_progress.ts — labelled prompt progress streams
• agents/webhook.ts — event-driven HTTP handler
• agents/tool_use.ts — tool call example
• sdk_demo.py — Python SDK with checkpointing + replay
• prompts/analysis.jinja — shared prompt template
• tools/web_search.ts — simple tool definition
• legacy-starlark/ — archived Starlark examples kept for migration reference

▶️ Try The Demo

The easiest way to explore Chidori is the interactive demo picker:

cargo build
./target/debug/chidori demo

chidori demo shows a numbered list of runnable examples, including demos that do not need an LLM provider and demos that exercise prompt tracing or streaming when provider environment variables are configured. Choose Hello agent for the fastest no-key path.

That demo runs a TypeScript agent, records a durable host-call log, and returns JSON. The direct command is:

./target/debug/chidori run examples/agents/hello.ts --input name=Colton

Expected output:

{
  "greeting": "Hello, Colton!"
}

What this demonstrates:

- examples/agents/hello.ts exports agent(input, chidori). - The agent calls chidori.log(...), so the runtime records a host call. - The agent returns plain JSON, which is what CLI, server, and SDK users receive. - A checkpoint is written under examples/agents/.chidori/runs/<run_id>/ for trace/replay workflows.

You can inspect the most recent run:

RUN_ID=$(ls -t examples/agents/.chidori/runs | head -1)
./target/debug/chidori trace "$RUN_ID" --dir examples/agents
./target/debug/chidori snapshot "$RUN_ID" --dir examples/agents

Human-In-The-Loop Demo

This demo shows the session API pausing on chidori.input(...) and resuming from the persisted VM state.

Start the server:

./target/debug/chidori serve examples/agents/input_pause.ts --port 8080

In another terminal, create a session:

curl -s http://localhost:8080/sessions \
  -H "Content-Type: application/json" \
  -d '{"input":{"request":"ship the TypeScript runtime"}}'

The response will have "status":"paused", an "id", and "pending_prompt":"Approve this request?". Resume it with:

SESSION_ID=<paste id from the previous response>

curl -s http://localhost:8080/sessions/$SESSION_ID/resume \
  -H "Content-Type: application/json" \
  -d '{"response":"yes"}'

The completed response includes:

{
  "output": {
    "request": "ship the TypeScript runtime",
    "approved": true
  }
}

That flow is the core Chidori loop: TypeScript code runs until a durable host operation pauses, Chidori persists the run, and resume continues from the saved state.

export ANTHROPIC_API_KEY=sk-ant-...

export OPENAI_API_KEY=sk-...

cargo build ./target/debug/chidori run agents/summarizer.ts \ --input document="Rust is a systems programming language..." ```

2. HTTP Server (event-driven + session API)

chidori serve agents/my_agent.ts --port 8080

Exposes: - GET /health — health check - ANY /* — any request is passed to agent(event) as an event dict - POST /sessions — create a session and run the agent with given input - GET /sessions — list all sessions - GET /sessions/{id} — get session result - GET /sessions/{id}/checkpoint — get the call log and snapshot manifest metadata - GET /sessions/{id}/snapshot — inspect snapshot manifest metadata without raw VM bytes - POST /sessions/{id}/resume — resume a paused input() or approval session - POST /sessions/{id}/replay — replay from a session's checkpoint - POST /sessions/{id}/cancel — cancel a running or stored session - POST /sessions/stream — run a session with SSE call and prompt progress events

🐍 Python SDK

The Python SDK is a pure-stdlib HTTP client that talks to a running chidori serve instance. No pip install, no native bindings.

```python import sys sys.path.insert(0, "sdk/python")

from chidori import AgentClient, Checkpoint

client = AgentClient("http://localhost:8080")