--- /dev/null
+package main
+
+import (
+ "context"
+ "encoding/json"
+ "errors"
+ "fmt"
+ "strings"
+)
+
+// errChainLogWrite marks an error of a log write, which ends the chain.
+var errChainLogWrite = errors.New("log write failed")
+
+// ChainClient sends model calls.
+type ChainClient interface {
+ Create(ctx context.Context, request ModelRequest) (ModelResponse, error)
+}
+
+// Chain is one request chain of a turn or of a subagent. It sends request with
+// the input of the first request, runs the tool calls of each response, and
+// sends their results back, until a response has no tool call: the final
+// answer. The replies, the tool calls, and the results go to the log of tools
+// as they happen. The thoughts never do.
+//
+// Each later request begins with the whole input of the request before it,
+// then its output items, then the tool results. The output items hold the
+// reasoning with its encrypted content, because the calls do not store it. So
+// each request begins with all of the previous one and reads it from the prompt
+// cache. The messages of take join each later request after the tool results.
+type Chain struct {
+ client ChainClient
+ tools *Toolbox
+ request ModelRequest
+ take func() []Message
+}
+
+// Run runs the chain from input, the input of its first request, and returns
+// the text of the final answer. An error of a log write wraps
+// errChainLogWrite.
+func (c *Chain) Run(ctx context.Context, input []any) (string, error) {
+ request := c.request
+ for request.Turn = "first"; ; request.Turn = "next" {
+ request.Input = input
+ response, err := c.client.Create(ctx, request)
+ if err != nil {
+ return "", err
+ }
+
+ var text string
+ input, text, err = c.handle(ctx, input, response)
+ if err != nil {
+ return "", err
+ }
+
+ if len(input) == 0 {
+ return text, nil
+ }
+ }
+}
+
+// handle logs the replies of response and runs its tool calls. It returns the
+// input of the next request: input, the output items, the tool results, and
+// the messages of take. After the final answer it returns no input, and the
+// replies of the answer as its text.
+func (c *Chain) handle(ctx context.Context, input []any, response ModelResponse) ([]any, string, error) {
+ var outputs []any
+ var replies []string
+ for _, raw := range response.Output {
+ item, err := chainInputItem(raw)
+ if err != nil {
+ return nil, "", err
+ }
+
+ input = append(input, item)
+
+ var call ChainFunctionCall
+ err = json.Unmarshal(raw, &call)
+ if err != nil {
+ return nil, "", fmt.Errorf("decode output item: %w", err)
+ }
+
+ switch call.Type {
+ case "message":
+ text, err := modelOutputText([]json.RawMessage{raw})
+ if err != nil {
+ return nil, "", err
+ }
+
+ text = strings.TrimSpace(text)
+ if text == "" {
+ continue
+ }
+
+ _, err = c.tools.log.Append("reply", text)
+ if err != nil {
+ return nil, "", fmt.Errorf("%w: %w", errChainLogWrite, err)
+ }
+
+ replies = append(replies, text)
+ case "function_call":
+ output, err := c.tools.Call(ctx, call.Name, call.Arguments)
+ if err != nil {
+ return nil, "", fmt.Errorf("%w: %w", errChainLogWrite, err)
+ }
+
+ outputs = append(outputs, ChainFunctionCallOutput{
+ Type: "function_call_output",
+ CallID: call.CallID,
+ Output: output,
+ })
+ }
+ }
+
+ if len(outputs) == 0 {
+ return nil, strings.Join(replies, "\n\n"), nil
+ }
+
+ input = append(input, outputs...)
+ for _, message := range c.take() {
+ input = append(input, modelUserMessage(message.Text))
+ }
+
+ return input, "", nil
+}
+
+// ChainFunctionCall holds the fields of a function call output item.
+type ChainFunctionCall struct {
+ Type string `json:"type"`
+ Name string `json:"name"`
+ Arguments string `json:"arguments"`
+ CallID string `json:"call_id"`
+}
+
+// ChainFunctionCallOutput is the input item with the result of a function
+// call.
+type ChainFunctionCallOutput struct {
+ Type string `json:"type"`
+ CallID string `json:"call_id"`
+ Output string `json:"output"`
+}
+
+// chainInputItem returns an output item as an input item for the next request.
+// The calls do not store their items, so the item goes back whole, without its
+// id.
+func chainInputItem(raw json.RawMessage) (map[string]any, error) {
+ var item map[string]any
+ err := json.Unmarshal(raw, &item)
+ if err != nil {
+ return nil, fmt.Errorf("decode output item: %w", err)
+ }
+
+ delete(item, "id")
+
+ return item, nil
+}
--- /dev/null
+package main
+
+import (
+ "net/http"
+ "path/filepath"
+ "testing"
+
+ "github.com/stretchr/testify/require"
+)
+
+func TestChainReturnsTheRepliesOfTheFinalAnswer(t *testing.T) {
+ server := newModelFake(t)
+ server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
+ if call.number == 0 {
+ modelWriteEvents(w, modelCompletedEvent([]any{
+ modelMessageItem("Looking."),
+ modelFunctionCallItem("fc_1", "call_1", "shell", `{"command":"echo hi"}`),
+ }, 100, 0, 5))
+
+ return
+ }
+
+ modelWriteEvents(w, modelCompletedEvent([]any{
+ modelMessageItem("It said hi."),
+ modelMessageItem("Done."),
+ }, 120, 100, 5))
+ }
+
+ memory, err := openMemory(t.Context(), t.TempDir(), newTestClock().Now, newFakeSummarizer())
+ require.NoError(t, err)
+
+ t.Cleanup(func() {
+ require.NoError(t, memory.Close())
+ })
+
+ chain := &Chain{
+ client: newModelClient(server.URL, ModelAPIKey(modelTestKey), "agent-key", filepath.Join(t.TempDir(), "usage.jsonl")),
+ tools: &Toolbox{
+ log: memory,
+ chat: memory,
+ workdir: t.TempDir(),
+ },
+ request: ModelRequest{
+ Model: "gpt-test",
+ Effort: "low",
+ },
+ take: func() []Message {
+ return nil
+ },
+ }
+
+ text, err := chain.Run(t.Context(), []any{modelUserMessage("Run echo hi.")})
+ require.NoError(t, err)
+ require.Equal(t, "It said hi.\n\nDone.", text)
+ require.Equal(t, 2, server.Calls())
+
+ var kinds []string
+ for id := range memory.Len() {
+ message, err := memory.Message(id)
+ require.NoError(t, err)
+
+ kinds = append(kinds, message.Kind)
+ }
+
+ require.Equal(t, []string{"reply", "tool", "echo", "reply", "reply"}, kinds)
+}
import (
"context"
- "encoding/json"
"errors"
- "fmt"
"path/filepath"
- "strings"
"sync"
"time"
)
-// errTurnLogWrite marks an error of a log write, which ends the loop.
-var errTurnLogWrite = errors.New("log write failed")
-
// newTurnLoop returns the loop of agent with its memory. Its calls go through
// client, and notify gets each event. The clock gives the date of a turn.
-func newTurnLoop(agent *Agent, memory *Memory, client TurnClient, clock func() time.Time, notify func(Event)) (*TurnLoop, error) {
+func newTurnLoop(agent *Agent, memory *Memory, client ChainClient, clock func() time.Time, notify func(Event)) (*TurnLoop, error) {
layout, err := loadViewLayout(filepath.Join(agent.dir, viewLayoutFileName))
if err != nil {
return nil, err
return l, nil
}
-// TurnClient sends model calls.
-type TurnClient interface {
- Create(ctx context.Context, request ModelRequest) (ModelResponse, error)
-}
-
// Event tells the clients what the agent does as it happens. Kind is the kind
// of a logged message (user, reply, tool, echo), or one of ready, busy, idle,
// and error. ID is the ID of a logged message. Source names the client that
// TurnLoop runs the turns of one agent. A user message goes to the log when it
// comes, and it wakes the loop. A turn waits until every message before it is
-// summarized. Then its first request holds the prompt, the tools, and the input
-// of turnInput. The loop runs the tool calls of each response and sends their
-// results back, until a response has no tool call: the final answer.
+// summarized. Then the turn runs a Chain whose first request holds the prompt,
+// the tools, and the input of turnInput.
//
// The view goes out as the messages of a ViewLayout, so the first request of a
-// turn reads the view of the turn before from the prompt cache.
-//
-// Each later request of a turn begins with the whole input of the request
-// before it, then its output items, then the tool results. The output items
-// hold the reasoning with its encrypted content, because the calls do not
-// store it. So each request begins with all of the previous one and reads it
-// from the prompt cache. User messages that come during a turn join the next
-// request after the tool results. The replies, the tool calls, and the results
-// go to the log as they happen. The thoughts never do.
+// turn reads the view of the turn before from the prompt cache. User messages
+// that come during a turn join the next request of its chain after the tool
+// results.
type TurnLoop struct {
agent *Agent
memory *Memory
- client TurnClient
+ client ChainClient
tools *Toolbox
layout *ViewLayout
clock func() time.Time
})
err := l.chain(ctx)
- if err != nil && !errors.Is(err, errTurnLogWrite) {
+ if err != nil && !errors.Is(err, errChainLogWrite) {
l.notify(Event{
Kind: "error",
Text: turnError(ctx, err),
return err
}
- input := turnInput(view, l.state(), messages)
- for turn := "first"; ; turn = "next" {
- response, err := l.client.Create(ctx, ModelRequest{
+ chain := &Chain{
+ client: l.client,
+ tools: l.tools,
+ request: ModelRequest{
Model: l.agent.config.Model,
Effort: l.agent.config.Effort,
Instructions: instructions,
Tools: promptTools(),
- Input: input,
- Turn: turn,
- })
- if err != nil {
- return err
- }
+ },
+ take: l.take,
+ }
- input, err = l.handle(ctx, input, response)
- if err != nil {
- return err
- }
+ _, err = chain.Run(ctx, turnInput(view, l.state(), messages))
- if len(input) == 0 {
- return nil
- }
- }
+ return err
}
// take returns the user messages that wait for a request, and removes them.
return "</chat>\n\nNow: " + l.clock().Format("Monday 2006-01-02 15:04 -07:00")
}
-// handle logs the replies of response and runs its tool calls. It returns the
-// input of the next request: input, the output items, the tool results, and
-// the user messages that came in the meantime. After the final answer it
-// returns no input.
-func (l *TurnLoop) handle(ctx context.Context, input []any, response ModelResponse) ([]any, error) {
- var outputs []any
- for _, raw := range response.Output {
- item, err := turnInputItem(raw)
- if err != nil {
- return nil, err
- }
-
- input = append(input, item)
-
- var call TurnFunctionCall
- err = json.Unmarshal(raw, &call)
- if err != nil {
- return nil, fmt.Errorf("decode output item: %w", err)
- }
-
- switch call.Type {
- case "message":
- text, err := modelOutputText([]json.RawMessage{raw})
- if err != nil {
- return nil, err
- }
-
- text = strings.TrimSpace(text)
- if text == "" {
- continue
- }
-
- _, err = l.Append("reply", text)
- if err != nil {
- return nil, fmt.Errorf("%w: %w", errTurnLogWrite, err)
- }
- case "function_call":
- output, err := l.tools.Call(ctx, call.Name, call.Arguments)
- if err != nil {
- return nil, fmt.Errorf("%w: %w", errTurnLogWrite, err)
- }
-
- outputs = append(outputs, TurnFunctionCallOutput{
- Type: "function_call_output",
- CallID: call.CallID,
- Output: output,
- })
- }
- }
-
- if len(outputs) == 0 {
- return nil, nil
- }
-
- input = append(input, outputs...)
- for _, message := range l.take() {
- input = append(input, modelUserMessage(message.Text))
- }
-
- return input, nil
-}
-
-// TurnFunctionCall holds the fields of a function call output item.
-type TurnFunctionCall struct {
- Type string `json:"type"`
- Name string `json:"name"`
- Arguments string `json:"arguments"`
- CallID string `json:"call_id"`
-}
-
-// TurnFunctionCallOutput is the input item with the result of a function call.
-type TurnFunctionCallOutput struct {
- Type string `json:"type"`
- CallID string `json:"call_id"`
- Output string `json:"output"`
-}
-
// turnError returns the text of the error that ended a turn.
func turnError(ctx context.Context, err error) string {
if ctx.Err() != nil {
return input
}
-
-// turnInputItem returns an output item as an input item for the next request.
-// The calls do not store their items, so the item goes back whole, without its
-// id.
-func turnInputItem(raw json.RawMessage) (map[string]any, error) {
- var item map[string]any
- err := json.Unmarshal(raw, &item)
- if err != nil {
- return nil, fmt.Errorf("decode output item: %w", err)
- }
-
- delete(item, "id")
-
- return item, nil
-}
projects / onechat / commitdiff