Back up the daemon and restore its data

[onechat] / subagent_run.go

diff --git a/subagent_run.go b/subagent_run.go
index 4a88800c3bed065aace3c242510d302a01ea9325..8084c390a7012d5213019f1f784ec3214ce08bbe 100644 (file)
--- a/subagent_run.go
+++ b/subagent_run.go
@@ -2,12 +2,17 @@ package main
 
 import (
        "context"
 
 import (
        "context"

+       "encoding/json"

        "errors"
        "fmt"
        "errors"
        "fmt"

+       "log/slog"
+       "maps"

        "os"
        "os"

+       "path/filepath"

        "slices"
        "strings"
        "sync"
        "slices"
        "strings"
        "sync"

+       "time"

 )
 
 const (
 )
 
 const (


@@ -17,6 +22,10 @@ const (
        // subagentNoteEffort follows subagentNote when the subagent has its own
        // effort.
        subagentNoteEffort = " at effort "
        // 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.
 )
 
 // The statuses of the end of a run, in the work event of its report.


@@ -36,9 +45,16 @@ var errSubagentClosed = errors.New("the daemon stops, so no subagent starts")
 // 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
 // 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. Stop
-// ends one run, and its report says that it stopped. 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
 //
 // 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


@@ -52,6 +68,8 @@ type SubagentRuns struct {
        state     func() string
        report    func(name, text, status string)
        note      func(text string) error
        state     func() string
        report    func(name, text, status string)
        note      func(text string) error

+       fail      func(err error)
+       logger    *slog.Logger

        notify    func(Event)
 
        mutex   sync.Mutex
        notify    func(Event)
 
        mutex   sync.Mutex


@@ -69,6 +87,7 @@ func (r *SubagentRuns) Spawn(ctx context.Context, tasks []string, effort *string
        }
 
        r.mutex.Lock()
        }
 
        r.mutex.Lock()

+

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


@@ -119,6 +138,7 @@ func (r *SubagentRuns) Spawn(ctx context.Context, tasks []string, effort *string
 // its log with a new run, with the view of now and the effort of its note.
 func (r *SubagentRuns) Tell(ctx context.Context, name, message string) (string, error) {
        r.mutex.Lock()
 // its log with a new run, with the view of now and the effort of its note.
 func (r *SubagentRuns) Tell(ctx context.Context, name, message string) (string, error) {
        r.mutex.Lock()

+

        run, found := r.running[name]
        if found {
                defer r.mutex.Unlock()
        run, found := r.running[name]
        if found {
                defer r.mutex.Unlock()


@@ -204,13 +224,59 @@ func (r *SubagentRuns) Stop(name string) (bool, error) {
        return true, nil
 }
 
        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.Error("subagent lost at restart", "name", name, "err", 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()
+}
+
+// Snapshot locks the run state, memory, and subagent logs while snapshot reads
+// the agent files. The lock order agrees with Stop, which logs under the run
+// lock.
+func (r *SubagentRuns) Snapshot(ctx context.Context, snapshot func() error) error {
+       r.mutex.Lock()
+       defer r.mutex.Unlock()
+
+       return r.memory.Snapshot(ctx, func() error {
+               return r.subagents.Snapshot(ctx, snapshot)
+       })
+}
+

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

+

        r.mutex.Unlock()
 
        r.workers.Wait()
        r.mutex.Unlock()
 
        r.workers.Wait()


@@ -268,12 +334,13 @@ func (r *SubagentRuns) effort(subagent *Subagent) (string, error) {
        return "", fmt.Errorf("the log of %s does not start with a subagent note", subagent.name)
 }
 
        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 Stop and Close end 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))
 
        r.mutex.Lock()
 func (r *SubagentRuns) start(ctx context.Context, run *SubagentRun) error {
        ctx, run.cancel = context.WithCancel(context.WithoutCancel(ctx))
 
        r.mutex.Lock()

+

        _, found := r.running[run.subagent.name]
        if r.closed || found {
                r.mutex.Unlock()
        _, found := r.running[run.subagent.name]
        if r.closed || found {
                r.mutex.Unlock()


@@ -287,9 +354,20 @@ func (r *SubagentRuns) start(ctx context.Context, run *SubagentRun) error {
        }
 
        r.running[run.subagent.name] = run
        }
 
        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()
 
        r.workers.Add(1)
        r.mutex.Unlock()
 

+       run.started = r.subagents.clock()
+       r.logger.Info("subagent started", "name", run.subagent.name)

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


@@ -300,20 +378,70 @@ func (r *SubagentRuns) start(ctx context.Context, run *SubagentRun) error {
        return nil
 }
 
        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
+// 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()
 
 // 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 (
+               replies []string
+               failure error
+       )
+

        stopped := false
        status := subagentFinished
        for more := true; more; {
                reply, err := r.chain(ctx, run)
 
                r.mutex.Lock()
        stopped := false
        status := subagentFinished
        for more := true; more; {
                reply, err := r.chain(ctx, run)
 
                r.mutex.Lock()

+

                stopped = run.stopped
                switch {
                case err != nil && stopped:
                stopped = run.stopped
                switch {
                case err != nil && stopped:


@@ -321,15 +449,28 @@ func (r *SubagentRuns) run(ctx context.Context, run *SubagentRun) {
                case err != nil:
                        reply = "failed: " + err.Error()
                        status = subagentFailed
                case err != nil:
                        reply = "failed: " + err.Error()
                        status = subagentFailed

+                       failure = err

                }
 
                more = err == nil && ctx.Err() == nil && len(run.pending) > 0
                run.pending = nil
                }
 
                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 !more {
                        delete(r.running, run.subagent.name)

+
+                       if !r.closed {
+                               saveErr = r.save()
+                       }

                }
                }

+

                r.mutex.Unlock()
 
                r.mutex.Unlock()
 

+               if saveErr != nil {
+                       r.fail(saveErr)
+               }
+

                replies = append(replies, reply)
        }
 
                replies = append(replies, reply)
        }
 


@@ -341,6 +482,16 @@ func (r *SubagentRuns) run(ctx context.Context, run *SubagentRun) {
                status = subagentStopped
        }
 
                status = subagentStopped
        }
 

+       length := r.subagents.clock().Sub(run.started)
+       switch status {
+       case subagentStopped:
+               r.logger.Info("subagent stopped", "name", run.subagent.name, "after", length)
+       case subagentFailed:
+               r.logger.Error("subagent failed", "name", run.subagent.name, "after", length, "err", failure)
+       default:
+               r.logger.Info("subagent done", "name", run.subagent.name, "after", length)
+       }
+

        r.report(run.subagent.name, subagentReport(run.subagent.name, replies), status)
 }
 
        r.report(run.subagent.name, subagentReport(run.subagent.name, replies), status)
 }
 


@@ -399,7 +550,7 @@ 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, the messages of Tell that wait for a request,
 
 // SubagentRun is one run of a subagent: the view that its requests get, its
 // effort, the end of its context, the messages of Tell that wait for a request,

-// and whether Stop ended it. The mutex of SubagentRuns protects the messages
+// whether Stop ended it, and the time it started. The mutex of SubagentRuns protects the messages

 // and the stop.
 type SubagentRun struct {
        subagent *Subagent
 // and the stop.
 type SubagentRun struct {
        subagent *Subagent


@@ -408,6 +559,12 @@ type SubagentRun struct {
        cancel   context.CancelFunc
        pending  []Message
        stopped  bool
        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
 }
 
 // subagentReport returns the report of the subagent name with its replies: the