Run subagents in the background with spawn

[onechat] / subagent_run.go

package main

import (
        "context"
        "errors"
        "fmt"
        "os"
        "slices"
        "strings"
        "sync"
)

const (
        // subagentNote is the first message of the log of a subagent.
        subagentNote = "subagent"

        // subagentNoteEffort follows subagentNote when the subagent has its own
        // effort.
        subagentNoteEffort = " at effort "
)

// errSubagentClosed tells that the runs of the subagents ended with the loop.
var errSubagentClosed = errors.New("the daemon stops, so no subagent starts")

// SubagentRuns runs the subagents of one agent in the background. Each run is
// a goroutine with its own Chain and its own context, which only Close ends.
// The first request of each chain holds the view of the agent at the spawn,
// then the state of the turn, then the whole log of the subagent. The report of
// the run goes to report: its final answer, or the error of a failed call. At
// the stop of the daemon the runs end with no report.
//
// Spawn splits the view with the layout of the agent, which has no lock. Spawn
// is a tool, so it runs in the goroutine of the turns, as each split of a turn
// does.
type SubagentRuns struct {
        agent     *Agent
        memory    *Memory
        subagents *Subagents
        layout    *ViewLayout
        client    ChainClient
        state     func() string
        report    func(text string, stopped bool)
        notify    func(Event)

        mutex   sync.Mutex
        running map[string]*SubagentRun
        closed  bool
        workers sync.WaitGroup
}

// Spawn starts one subagent for each task and returns their names at once. It
// waits until the view has no pending summary, and gives the view to each
// subagent. A nil effort is the effort of the agent.
func (r *SubagentRuns) Spawn(ctx context.Context, tasks []string, effort *string) (string, error) {
        if len(tasks) == 0 {
                return "", errors.New("no tasks")
        }

        r.mutex.Lock()
        closed := r.closed
        r.mutex.Unlock()

        if closed {
                return "", errSubagentClosed
        }

        note := subagentNote
        runEffort := r.agent.config.Effort
        if effort != nil && !slices.Contains(promptEfforts, *effort) {
                return "", fmt.Errorf("no effort %q: one of %s", *effort, strings.Join(promptEfforts, ", "))
        }

        if effort != nil && *effort != runEffort {
                note += subagentNoteEffort + *effort
                runEffort = *effort
        }

        err := r.memory.Settle(ctx)
        if err != nil {
                return "", fmt.Errorf("wait for the summaries: %w", err)
        }

        view, err := r.layout.Split(r.memory.ViewLines(nil))
        if err != nil {
                return "", err
        }

        names := make([]string, 0, len(tasks))
        for _, task := range tasks {
                name, err := r.create(ctx, note, task, view, runEffort)
                if err != nil && len(names) > 0 {
                        return "", fmt.Errorf("started %s, then: %w", strings.Join(names, ", "), err)
                }

                if err != nil {
                        return "", err
                }

                names = append(names, name)
        }

        return "Started " + strings.Join(names, ", ") + ". Each one's report will reach you as a message when it finishes.", nil
}

// create makes a new subagent with its note, its task, and its work directory,
// and starts its run.
func (r *SubagentRuns) create(ctx context.Context, note, task string, view []string, effort string) (string, error) {
        subagent, err := r.subagents.Create(ctx)
        if err != nil {
                return "", err
        }

        for _, message := range [][2]string{{"note", note}, {"user", task}} {
                _, err = subagent.Append(message[0], message[1])
                if err != nil {
                        return "", err
                }
        }

        err = subagent.Flush(ctx)
        if err != nil {
                return "", err
        }

        err = os.MkdirAll(subagent.WorkDir(), 0o700)
        if err != nil {
                return "", fmt.Errorf("create work directory: %w", err)
        }

        return subagent.name, r.start(ctx, &SubagentRun{
                subagent: subagent,
                view:     view,
                effort:   effort,
        })
}

// start runs run in the background. Its context keeps the values of ctx, but
// only Close ends it.
func (r *SubagentRuns) start(ctx context.Context, run *SubagentRun) error {
        ctx, run.cancel = context.WithCancel(context.WithoutCancel(ctx))

        r.mutex.Lock()
        if r.closed {
                r.mutex.Unlock()
                run.cancel()

                return errSubagentClosed
        }

        r.running[run.subagent.name] = run
        r.workers.Add(1)
        r.mutex.Unlock()

        r.notify(Event{
                Kind: "subagent",
                Text: run.subagent.name,
        })

        go r.run(ctx, run)

        return nil
}

// run runs the chain of run and reports its end, unless Close ended it.
func (r *SubagentRuns) run(ctx context.Context, run *SubagentRun) {
        defer r.workers.Done()
        defer run.cancel()

        reply := r.chain(ctx, run)

        r.mutex.Lock()
        delete(r.running, run.subagent.name)
        r.mutex.Unlock()

        if ctx.Err() != nil {
                return
        }

        r.report(subagentReport(run.subagent.name, []string{reply}), false)
}

// chain runs one request chain of run and returns its final answer, or the
// error of the chain after "failed: ".
func (r *SubagentRuns) chain(ctx context.Context, run *SubagentRun) string {
        instructions, err := r.agent.SubagentPrompt()
        if err != nil {
                return "failed: " + err.Error()
        }

        log, err := run.subagent.Show()
        if err != nil {
                return "failed: " + err.Error()
        }

        input := make([]any, 0, len(run.view)+2)
        for _, text := range run.view {
                input = append(input, modelUserMessage(text))
        }

        input = append(input, modelUserMessage(r.state()), modelUserMessage(log))
        chain := &Chain{
                client: r.client,
                tools: &Toolbox{
                        log:       run.subagent,
                        chat:      r.memory,
                        subagents: r.subagents,
                        workdir:   run.subagent.WorkDir(),
                },
                request: ModelRequest{
                        Model:        r.agent.config.Model,
                        Effort:       run.effort,
                        Instructions: instructions,
                        Tools:        promptSubagentTools(),
                        Agent:        run.subagent.name,
                },
                take: func() []Message {
                        return nil
                },
        }

        reply, err := chain.Run(ctx, input)
        if err != nil {
                return "failed: " + err.Error()
        }

        return reply
}

// Close ends every run and waits for them. No subagent starts after Close.
func (r *SubagentRuns) Close() {
        r.mutex.Lock()
        r.closed = true
        for _, run := range r.running {
                run.cancel()
        }
        r.mutex.Unlock()

        r.workers.Wait()
}

// SubagentRun is one run of a subagent: the view that its requests get, its
// effort, and the end of its context.
type SubagentRun struct {
        subagent *Subagent
        view     []string
        effort   string
        cancel   context.CancelFunc
}

// subagentReport returns the report of the subagent name with its replies: the
// name in brackets, a new line, the replies, and the zoom that gives its whole
// chat.
func subagentReport(name string, replies []string) string {
        return "[" + name + "]\n" + strings.Join(replies, "\n\n") + "\n\nFull chat: zoom(\"" + name + "\")"
}