X-Git-Url: http://git.giacomolli.uk/onechat/blobdiff_plain/eeb821ff17ad79c87ca9b64cd9be26f8612f4a03..f2e67d5579593e2d4f1cb2d9165b851c52a06715:/subagent_run.go

diff --git a/subagent_run.go b/subagent_run.go
index 4a88800..16cf2bd 100644
--- a/subagent_run.go
+++ b/subagent_run.go
@@ -2,9 +2,12 @@ package main
 
 import (
 	"context"
+	"encoding/json"
 	"errors"
 	"fmt"
+	"maps"
 	"os"
+	"path/filepath"
 	"slices"
 	"strings"
 	"sync"
@@ -17,6 +20,10 @@ 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.
@@ -40,6 +47,12 @@ var errSubagentClosed = errors.New("the daemon stops, so no subagent starts")
 // 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
 // split of a turn does.
@@ -52,6 +65,7 @@ type SubagentRuns struct {
 	state     func() string
 	report    func(name, text, status string)
 	note      func(text string) error
+	fail      func(err error)
 	notify    func(Event)
 
 	mutex   sync.Mutex
@@ -204,6 +218,37 @@ func (r *SubagentRuns) Stop(name string) (bool, error) {
 	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.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()
@@ -287,6 +332,15 @@ 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()
 
@@ -300,6 +354,50 @@ func (r *SubagentRuns) start(ctx context.Context, run *SubagentRun) error {
 	return nil
 }
 
+// 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 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. A chain that Stop ends gives the reply "stopped".
@@ -325,11 +423,19 @@ func (r *SubagentRuns) run(ctx context.Context, run *SubagentRun) {
 
 		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)
 	}
 
@@ -410,6 +516,11 @@ type SubagentRun struct {
 	stopped  bool
 }
 
+// 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
 // name in brackets, a new line, the replies, and the zoom that gives its whole
 // chat.