X-Git-Url: https://git.giacomolli.uk/onechat/blobdiff_plain/e1e48ba388c28105fff3b70c8ca02304f5d57f53..e73e445bd78ba98271da6b5165f132299fca04b1:/subagent_run.go?ds=inline

diff --git a/subagent_run.go b/subagent_run.go
index b2e5e62..67cf79d 100644
--- a/subagent_run.go
+++ b/subagent_run.go
@@ -2,12 +2,17 @@ package main
 
 import (
 	"context"
+	"encoding/json"
 	"errors"
 	"fmt"
+	"log"
+	"maps"
 	"os"
+	"path/filepath"
 	"slices"
 	"strings"
 	"sync"
+	"time"
 )
 
 const (
@@ -17,20 +22,39 @@ const (
 	// subagentNoteEffort follows subagentNote when the subagent has its own
 	// effort.
 	subagentNoteEffort = " at effort "
+
+	// subagentWorkFileName is the file of the names of the running
+	// subagents in the directory of an agent.
+	subagentWorkFileName = "work.json"
+)
+
+// The statuses of the end of a run, in the work event of its report.
+const (
+	subagentFinished = "finished"
+	subagentStopped  = "stopped"
+	subagentFailed   = "failed"
 )
 
 // 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 context, which only Close ends. A run is one or more
-// request chains. The first request of each chain holds the view of the agent
-// at the start of the run, then the state of the turn, then the whole log of
-// the subagent. A message of Tell that comes during a chain joins its next
+// SubagentRuns runs the subagents of one agent in the background. Each run is a
+// goroutine with its own context, which only Stop and Close end. A run is one
+// or more request chains. The first request of each chain holds the view of the
+// agent at the start of the run, then the state of the turn, then the whole log
+// of the subagent. A message of Tell that comes during a chain joins its next
 // request after the tool results. One that comes after the final answer starts
 // one more chain, which reads the log again. The report of the run goes to
-// report: the final answer of each chain, or the error of a failed call. At the
-// stop of the daemon the runs end with no report.
+// report: the final answer of each chain, or the error of a failed call. The
+// logger gets one line at the end of a run: done, failed with the cause, or
+// stopped. Stop ends one run, and its report says that it stopped. At the stop
+// of the daemon the runs end with no report.
+//
+// work.json in the directory of the agent holds the names of the running
+// subagents. Each start and each end of a run writes it whole, but Close does
+// not, so after a stop of the daemon it names the runs that the stop cut, and
+// Resume goes on with them. The logs alone cannot tell this, because a stop
+// can cut a run just after its final reply.
 //
 // Spawn and Tell split the view with the layout of the agent, which has no
 // lock. They are tools, so they run in the goroutine of the turns, as each
@@ -42,7 +66,10 @@ type SubagentRuns struct {
 	layout    *ViewLayout
 	client    ChainClient
 	state     func() string
-	report    func(text string, stopped bool)
+	report    func(name, text, status string)
+	note      func(text string) error
+	fail      func(err error)
+	logger    *log.Logger
 	notify    func(Event)
 
 	mutex   sync.Mutex
@@ -172,6 +199,61 @@ func (r *SubagentRuns) Tell(ctx context.Context, name, message string) (string,
 	return "Sent. " + name + "'s reply will reach you as a message.", nil
 }
 
+// Stop ends the run of the subagent name and tells if it ran. The chat gets the
+// note "Stopped by the user: NAME." under the mutex, so the note comes before
+// the report of the run.
+func (r *SubagentRuns) Stop(name string) (bool, error) {
+	r.mutex.Lock()
+	defer r.mutex.Unlock()
+
+	run, found := r.running[name]
+	if !found {
+		return false, nil
+	}
+
+	err := r.note("Stopped by the user: " + name + ".")
+	if err != nil {
+		return true, err
+	}
+
+	run.stopped = true
+	run.cancel()
+
+	return true, nil
+}
+
+// Resume goes on with the runs that a stop of the daemon cut, as work.json
+// names them. The chat gets the note "The daemon restarted, cutting NAMES;
+// nothing was lost: go on.", and each cut subagent gets the same text as a
+// message of Tell. A subagent that cannot go on gets a failed report.
+func (r *SubagentRuns) Resume(ctx context.Context) error {
+	names, err := r.load()
+	if err != nil || len(names) == 0 {
+		return err
+	}
+
+	text := "The daemon restarted, cutting " + strings.Join(names, ", ") + "; nothing was lost: go on."
+	err = r.note(text)
+	if err != nil {
+		return err
+	}
+
+	for _, name := range names {
+		_, err = r.Tell(ctx, name, text)
+		if err != nil {
+			r.logger.Printf("subagent %s lost at restart: %v", name, err)
+			r.report(name, subagentReport(name, []string{"failed: " + err.Error()}), subagentFailed)
+		}
+	}
+
+	// A subagent that cannot go on does not run, so work.json no longer
+	// names it, and the next start does not try it again.
+	r.mutex.Lock()
+	defer r.mutex.Unlock()
+
+	return r.save()
+}
+
 // Close ends every run and waits for them. No subagent starts after Close.
 func (r *SubagentRuns) Close() {
 	r.mutex.Lock()
@@ -236,8 +318,8 @@ func (r *SubagentRuns) effort(subagent *Subagent) (string, error) {
 	return "", fmt.Errorf("the log of %s does not start with a subagent note", subagent.name)
 }
 
-// start runs run in the background. Its context keeps the values of ctx, but
-// only Close ends it.
+// start runs run in the background and logs its start. Its context keeps the
+// values of ctx, but only Stop and Close end it.
 func (r *SubagentRuns) start(ctx context.Context, run *SubagentRun) error {
 	ctx, run.cancel = context.WithCancel(context.WithoutCancel(ctx))
 
@@ -255,9 +337,20 @@ func (r *SubagentRuns) start(ctx context.Context, run *SubagentRun) error {
 	}
 
 	r.running[run.subagent.name] = run
+	err := r.save()
+	if err != nil {
+		delete(r.running, run.subagent.name)
+		r.mutex.Unlock()
+		run.cancel()
+
+		return err
+	}
+
 	r.workers.Add(1)
 	r.mutex.Unlock()
 
+	run.started = r.subagents.clock()
+	r.logger.Printf("subagent %s started", run.subagent.name)
 	r.notify(Event{
 		Kind: "subagent",
 		Text: run.subagent.name,
@@ -268,36 +361,112 @@ func (r *SubagentRuns) start(ctx context.Context, run *SubagentRun) error {
 	return nil
 }
 
-// run runs the chains of run and reports its end, unless Close ended it. A
-// chain follows while messages wait after the final answer of the chain before
-// it, and no chain failed.
+// load reads the names of work.json. A missing file names none.
+func (r *SubagentRuns) load() ([]string, error) {
+	data, err := os.ReadFile(r.workPath())
+	if errors.Is(err, os.ErrNotExist) {
+		return nil, nil
+	}
+
+	if err != nil {
+		return nil, fmt.Errorf("read running subagents: %w", err)
+	}
+
+	var file SubagentWorkFile
+	err = json.Unmarshal(data, &file)
+	if err != nil {
+		return nil, fmt.Errorf("decode running subagents: %w", err)
+	}
+
+	return file.Running, nil
+}
+
+// save writes the names of the running subagents to work.json. The caller
+// holds the mutex.
+func (r *SubagentRuns) save() error {
+	file := SubagentWorkFile{
+		Running: append([]string{}, slices.Sorted(maps.Keys(r.running))...),
+	}
+
+	data, err := json.Marshal(file)
+	if err != nil {
+		return fmt.Errorf("encode running subagents: %w", err)
+	}
+
+	err = writeFileAtomic(r.workPath(), append(data, '\n'))
+	if err != nil {
+		return fmt.Errorf("save running subagents: %w", err)
+	}
+
+	return nil
+}
+
+func (r *SubagentRuns) workPath() string {
+	return filepath.Join(r.agent.dir, subagentWorkFileName)
+}
+
+// run runs the chains of run and reports and logs its end with the length of
+// the run, unless Close ended it. A chain follows while messages wait after the final answer of the chain before
+// it, and no chain failed. A chain that Stop ends gives the reply "stopped".
 func (r *SubagentRuns) run(ctx context.Context, run *SubagentRun) {
 	defer r.workers.Done()
 	defer run.cancel()
 
 	var replies []string
+	var failure error
+	stopped := false
+	status := subagentFinished
 	for more := true; more; {
 		reply, err := r.chain(ctx, run)
-		if err != nil {
+
+		r.mutex.Lock()
+		stopped = run.stopped
+		switch {
+		case err != nil && stopped:
+			reply = "stopped"
+		case err != nil:
 			reply = "failed: " + err.Error()
+			status = subagentFailed
+			failure = err
 		}
 
-		replies = append(replies, reply)
-
-		r.mutex.Lock()
-		more = err == nil && len(run.pending) > 0
+		more = err == nil && ctx.Err() == nil && len(run.pending) > 0
 		run.pending = nil
+		var saveErr error
 		if !more {
 			delete(r.running, run.subagent.name)
+			if !r.closed {
+				saveErr = r.save()
+			}
 		}
 		r.mutex.Unlock()
+
+		if saveErr != nil {
+			r.fail(saveErr)
+		}
+
+		replies = append(replies, reply)
 	}
 
-	if ctx.Err() != nil {
+	if ctx.Err() != nil && !stopped {
 		return
 	}
 
-	r.report(subagentReport(run.subagent.name, replies), false)
+	if stopped {
+		status = subagentStopped
+	}
+
+	length := durationText(r.subagents.clock().Sub(run.started))
+	switch status {
+	case subagentStopped:
+		r.logger.Printf("subagent %s stopped after %s", run.subagent.name, length)
+	case subagentFailed:
+		r.logger.Printf("subagent %s failed after %s: %v", run.subagent.name, length, failure)
+	default:
+		r.logger.Printf("subagent %s done after %s", run.subagent.name, length)
+	}
+
+	r.report(run.subagent.name, subagentReport(run.subagent.name, replies), status)
 }
 
 // chain runs one request chain of run and returns its final answer.
@@ -354,14 +523,22 @@ func (r *SubagentRuns) take(run *SubagentRun) []Message {
 }
 
 // SubagentRun is one run of a subagent: the view that its requests get, its
-// effort, the end of its context, and the messages of Tell that wait for a
-// request. The mutex of SubagentRuns protects the messages.
+// effort, the end of its context, the messages of Tell that wait for a request,
+// whether Stop ended it, and the time it started. The mutex of SubagentRuns protects the messages
+// and the stop.
 type SubagentRun struct {
 	subagent *Subagent
 	view     []string
 	effort   string
 	cancel   context.CancelFunc
 	pending  []Message
+	stopped  bool
+	started  time.Time
+}
+
+// SubagentWorkFile is the content of work.json.
+type SubagentWorkFile struct {
+	Running []string `json:"running"`
 }
 
 // subagentReport returns the report of the subagent name with its replies: the