chore: import upstream snapshot with attribution
govulncheck / govulncheck (push) Has been cancelled
Lint / golangci-lint (push) Has been cancelled
Run Tests / Unit Tests (push) Has been cancelled
Run Tests / Etcd Integration Tests (push) Has been cancelled
Harness (E2E) / Harnesses (mock LLM) (push) Has been cancelled
Harness (E2E) / Provider harnesses (live LLM conformance) (push) Has been cancelled
govulncheck / govulncheck (push) Has been cancelled
Lint / golangci-lint (push) Has been cancelled
Run Tests / Unit Tests (push) Has been cancelled
Run Tests / Etcd Integration Tests (push) Has been cancelled
Harness (E2E) / Harnesses (mock LLM) (push) Has been cancelled
Harness (E2E) / Provider harnesses (live LLM conformance) (push) Has been cancelled
This commit is contained in:
@@ -0,0 +1,63 @@
|
||||
# Agent Tool Wrappers
|
||||
|
||||
Middleware around an agent's tool execution, the same way
|
||||
`client.CallWrapper` and `server.HandlerWrapper` wrap RPCs.
|
||||
|
||||
Every tool call an agent makes runs through `ai.ToolHandler`:
|
||||
|
||||
```go
|
||||
type ToolHandler func(ctx context.Context, call ai.ToolCall) ai.ToolResult
|
||||
type ToolWrapper func(ai.ToolHandler) ai.ToolHandler
|
||||
```
|
||||
|
||||
`WrapTool` (exposed as `micro.AgentWrapTool`) registers a wrapper: it
|
||||
takes the next handler and returns a new one. Code before `next(...)`
|
||||
runs before the tool, code after runs after. That single seam covers
|
||||
the whole lifecycle — before/after hooks, timing, metrics, retries,
|
||||
inspecting results.
|
||||
|
||||
## What this example does
|
||||
|
||||
One flaky `weather` service and one agent with two wrappers:
|
||||
|
||||
- **observe** — times every call and records a per-tool count, logging
|
||||
the correlation ID (`call.ID`) carried through from the provider. It
|
||||
observes; it changes nothing.
|
||||
- **retry** — re-runs a call whose result is an error, up to three
|
||||
attempts. The weather service fails the first time it's hit and
|
||||
succeeds after, so retry turns a transient failure into a success the
|
||||
model never sees.
|
||||
|
||||
Wrappers compose **outermost-first**: `observe` is registered first, so
|
||||
it wraps `retry` and sees one logical call even when retry runs the tool
|
||||
twice.
|
||||
|
||||
```go
|
||||
micro.NewAgent("forecaster",
|
||||
micro.AgentServices("weather"),
|
||||
micro.AgentProvider(provider),
|
||||
micro.AgentAPIKey(apiKey),
|
||||
micro.AgentWrapTool(m.observe, retry(3)),
|
||||
)
|
||||
```
|
||||
|
||||
## Wrappers vs. guardrails
|
||||
|
||||
Developer wrappers run **outside** the built-in guardrails (`MaxSteps`,
|
||||
`LoopLimit`, `ApproveTool`), so they see every call and its result —
|
||||
including a guardrail's refusal. The flip side: a retry wrapper's
|
||||
`next` is the full guardrail stack, so each retry is also counted by
|
||||
loop detection. Keep `LoopLimit` at or above your retry count, or set
|
||||
`AgentLoopLimit(0)` when a wrapper owns the repetition.
|
||||
|
||||
See the [Agent Guardrails guide](../../internal/website/docs/guides/agent-guardrails.md)
|
||||
for the full picture.
|
||||
|
||||
## Run
|
||||
|
||||
Needs an LLM provider key:
|
||||
|
||||
```bash
|
||||
export ANTHROPIC_API_KEY=sk-ant-... # or OPENAI_API_KEY, GEMINI_API_KEY, ...
|
||||
go run main.go
|
||||
```
|
||||
Reference in New Issue
Block a user