README
¶
Agent examples
Runnable examples for the stateful agent.Agent. They call a live model API and
may incur provider charges. Set the DeepSeek API key and run them from the
repository root:
export DEEPSEEK_API_KEY=your-deepseek-api-key
go run ./example/agent/basic # smallest: one tool, one prompt
go run ./example/agent/tool # interactive terminal session
go run ./example/agent/hooks # interception points (needs MINIMAX_CN_API_KEY too)
basic
The smallest stateful agent: one tool and one prompt. It subscribes to the event stream to print the answer as it streams and to show when the tool runs.
tool
An interactive terminal with colored reasoning, answers, tool progress, and a
compact run summary. The same Agent retains the conversation across prompts.
All lifecycle events are handled; enter /debug to display low-level events.
The /model <provider> <model-id> command switches models while preserving the
entire conversation transcript.
Ask as many questions as you like:
You › What is the weather in Shanghai?
You › What about Beijing?
You › /quit
Enter /quit, quit, or exit to end the session. Enter /thinking to show
or hide provider-supplied reasoning. Set NO_COLOR=1 to disable ANSI colors.
Weather data is simulated, so the example needs no second API or service.
For example, switch from DeepSeek V4 Flash to V4 Pro after an answer:
You › /model deepseek deepseek-v4-pro
↻ MODEL Switched to deepseek/deepseek-v4-pro · 2 messages preserved
hooks
A non-interactive program showing the agent's interception points: BeforeToolCall
gates each call and blocks one tool, AfterToolCall annotates results,
PrepareNextTurn switches to a different model — across wire protocols — after the
tool turn, and ShouldStopAfterTurn guards against a runaway loop. It runs tools
on DeepSeek (OpenAI-compatible) and answers on MiniMax (Anthropic-compatible), so
it needs both DEEPSEEK_API_KEY and MINIMAX_CN_API_KEY.
Directories
¶
| Path | Synopsis |
|---|---|
|
Command basic demonstrates the smallest stateful agent: one tool, one prompt.
|
Command basic demonstrates the smallest stateful agent: one tool, one prompt. |
|
Command hooks demonstrates the agent's interception points: a policy gate that blocks a tool (BeforeToolCall), an annotation applied to results (AfterToolCall), a model switch between turns across wire protocols (PrepareNextTurn), and a turn guard (ShouldStopAfterTurn).
|
Command hooks demonstrates the agent's interception points: a policy gate that blocks a tool (BeforeToolCall), an annotation applied to results (AfterToolCall), a model switch between turns across wire protocols (PrepareNextTurn), and a turn guard (ShouldStopAfterTurn). |
|
Command tool demonstrates an interactive stateful agent with tool use.
|
Command tool demonstrates an interactive stateful agent with tool use. |