"encoding/json"
"errors"
"fmt"
+ "log/slog"
"maps"
"os"
"path/filepath"
"slices"
"strings"
"sync"
+ "time"
)
const (
// 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
report func(name, text, status string)
note func(text string) error
fail func(err error)
+ logger *slog.Logger
notify func(Event)
mutex sync.Mutex
}
r.mutex.Lock()
+
closed := r.closed
r.mutex.Unlock()
// 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()
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)
}
}
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()
+
r.closed = true
for _, run := range r.running {
run.cancel()
}
+
r.mutex.Unlock()
r.workers.Wait()
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()
+
_, found := r.running[run.subagent.name]
if r.closed || found {
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,
}
var file SubagentWorkFile
+
err = json.Unmarshal(data, &file)
if err != nil {
return nil, fmt.Errorf("decode running subagents: %w", err)
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
+// 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 (
+ replies []string
+ failure error
+ )
+
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:
case err != nil:
reply = "failed: " + err.Error()
status = subagentFailed
+ failure = err
}
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 {
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)
}
// 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
cancel context.CancelFunc
pending []Message
stopped bool
+ started time.Time
}
// SubagentWorkFile is the content of work.json.
projects / onechat / blobdiff