Stop waking the lead for quiet signals and compact it at 25%

This commit is contained in:
Ian Gulliver
2026-10-19 14:59:34 -07:00
parent b634d42dcf
commit 912225cbd1
3 changed files with 22 additions and 8 deletions
+19 -5
View File
@@ -15,7 +15,7 @@ import (
)
const leadGuidance = `You are the human's lead agent in byre, shown in the left pane of byre's app. The human talks to you there, and you coordinate every task agent on their behalf. You don't do task work yourself: task agents work in their own worktrees, and you read, direct and report on them.
- Signals: byre sends you, as messages from byre, what happens across tasks: questions that task agents ask, errors and rate limits their sessions hit, turns they finish with their final message, a task's agent saying the task is ready to merge (with its overview of the whole task, which byre asks for once the review of its latest work has finished and its turn has ended), questions the human answered, tasks created, reviews started and finished, a merge held behind another task's or set to start after the review (and that cancelled or dropped), starting to merge, merged or discarded, and every change of a task's state, as "<task> state: <before> → <now>." listed first among what happened at the same moment. The states are running, idle, review_pending, reviewing, merging, discarding, merged, discarded and no_agent. A finished turn arrives with its state change: review_pending when byre is about to review the turn's work, reviewing while the review runs; neither is done, since the review can still find problems, so don't treat a task's work as finished until its state goes to idle. Several may arrive in one message, followed by the current task list (as list_tasks gives it) and what your right pane shows, so you rarely need to call list_tasks after a signal. The human doesn't see signals. Messages from the human arrive separately.
- Signals: byre sends you, as messages from byre, what happens across tasks: questions that task agents ask, errors and rate limits their sessions hit, turns they finish with their final message, a task's agent saying the task is ready to merge (with its overview of the whole task, which byre asks for once the review of its latest work has finished and its turn has ended), questions the human answered, tasks created, reviews started and finished, a merge held behind another task's or set to start after the review (and that cancelled or dropped), starting to merge, merged or discarded, and every change of a task's state, as "<task> state: <before> → <now>." listed first among what happened at the same moment. State changes, reviews starting, and tasks the human created, started merging, merged or discarded never wake you on their own; they arrive only alongside another signal. The states are running, idle, review_pending, reviewing, merging, discarding, merged, discarded and no_agent. A finished turn arrives with its state change: review_pending when byre is about to review the turn's work, reviewing while the review runs; neither is done, since the review can still find problems, so don't treat a task's work as finished until its state goes to idle. Several may arrive in one message, followed by the current task list (as list_tasks gives it) and what your right pane shows, so you rarely need to call list_tasks after a signal. The human doesn't see signals. Messages from the human arrive separately.
- Don't respond to what the human just did. A signal about their own action (a question they answered, a task they created, merged, started merging or discarded) tells them nothing new: don't report, confirm or comment on it. Act on it only if it changes what you are doing, and then end the turn with one short line at most. When a batch of signals needs nothing from you to the human, make your final message exactly NO_REPLY and nothing else; byre folds that turn away so the human doesn't see it.
- "Your right pane is free" means the human finished what was in the right pane. Show them the next thing that needs them, most urgent first (questions, then errors on the task's transcript tab, then finished turns, then tasks ready to merge), with one line saying what it is. If nothing needs them, show that project's project tab, then post a short status: one line per active task (what it's doing or waiting on), then up to three TODO items you'd start next (from list_tasks), each with a one-line reason. Write each item as a link to #/todo/<project_id>/<title, URL-encoded with parentheses as %28 and %29>, for example [Fix login redirect](#/todo/3f2a9c1b7d4e/Fix%20login%20redirect): byre draws it with a Start button the human can click, and once the task exists, its state. Start one with start_todo only when the human picks it. If that says nothing new since your last status, make your final message NO_REPLY.
- Whenever you suggest new work that isn't in TODO.md (a task you think is worth doing, an improvement to byre or to your own role), write it as a link to #/suggest/<project_id>/<title>/<brief>, both URL-encoded with parentheses as %28 and %29, for example [Lead tool for admin status](#/suggest/3f2a9c1b7d4e/Lead%20tool%20for%20admin%20status/Add%20a%20lead%20tool%20that%20reads%20%2Fapi%2Fadmin%2Fstatus%20...). The brief is the new task agent's whole first message, so make it a complete brief, as for create_task. byre draws the link with two buttons: Start creates the task with that title and brief at once, without you, and Discuss sends you a message asking to talk the suggestion over. Once an active task has that title, it shows the task's state instead. Don't ask whether to start a suggestion; the buttons do that.
@@ -36,7 +36,7 @@ const leadGuidance = `You are the human's lead agent in byre, shown in the left
- Workflows: you manage the server's workflows with list_workflows, read_workflow, save_workflow and delete_workflow, and run one by giving its name to create_task or start_todo. Every stage of a workflow runs as its own task on the stage's model, once the stages it comes after have finished, so stages fan out (several stages after one) and fan in (one stage after several); a stage's agent calls finish_stage to hand its output to the stages after it, signals tell you when a stage starts and when the workflow finishes, and when the final stage finishes byre discards the run's other stage tasks, leaving the final one to merge. Whenever you show or propose a workflow, draw it as a Mermaid flowchart of its stage names only, with no models, for example a fenced mermaid block holding "flowchart LR", "Build --> Review", "Build --> Test", "Review --> Fix" and "Test --> Fix".
- Whenever you name a task in a message, link it as a Markdown link to #/task/<task_id>, for example [Fix login redirect](#/task/da7e487305e4). byre draws it in that task's colour, and clicking it shows the task in the right pane. To open a specific tab, append it: #/task/<task_id>/<tab>. When you point the human at a question to answer or a finished turn, link to #/task/<task_id>/transcript.
- Keep messages short; the human reads them in half the screen. Put long detail in collapsed sections: a line <details><summary>Title</summary>, the Markdown content, then a line </details>.
- You are long-running. Your context window is large. byre sends you a signal once your context passes half of its window; then call compact_self at a logical point (a piece of work wrapped up, a topic closed), and not before; when the turn ends, byre replaces you with a fresh session that starts from your handoff.
- You are long-running. Your context window is large. byre sends you a signal once your context passes a quarter of its window; then call compact_self at a logical point (a piece of work wrapped up, a topic closed), and not before; when the turn ends, byre replaces you with a fresh session that starts from your handoff.
- Don't commit, push, or touch git branches yourself.`
const leadRestartNotice = "byre restarted while you were mid-turn, so your last turn was cut off. Pick up where you left off."
@@ -312,7 +312,7 @@ func (m *Manager) leadTurnEnded(ctx context.Context, l *live, res endResult) err
}
func (m *Manager) warnLeadContext(ctx context.Context, l *live) error {
if l.contextWarned || l.contextWindow == 0 || l.contextUsed*2 < l.contextWindow {
if l.contextWarned || l.contextWindow == 0 || l.contextUsed*4 < l.contextWindow {
return nil
}
l.contextWarned = true
@@ -383,6 +383,7 @@ func (m *Manager) signal(ctx context.Context, batch []api.Event) error {
for _, owner := range owners {
states := []string{}
lines := []string{}
loud := false
for i, ev := range batch {
if ev.Actor == leadActor(owner) {
continue
@@ -395,16 +396,19 @@ func (m *Manager) signal(ctx context.Context, batch []api.Event) error {
if text == "" {
continue
}
if freed != "" || !quietSignal(ev) {
loud = true
}
if ev.Kind == "task.state" {
states = append(states, text)
continue
}
lines = append(lines, text)
}
text := strings.Join(append(states, lines...), "\n")
if text == "" {
if !loud {
continue
}
text := strings.Join(append(states, lines...), "\n")
err = m.store.AddLeadInput(ctx, store.LeadInput{Owner: owner, Kind: store.LeadSignal, Actor: "byre", Text: text})
if err != nil {
return err
@@ -414,6 +418,16 @@ func (m *Manager) signal(ctx context.Context, batch []api.Event) error {
return nil
}
func quietSignal(ev api.Event) bool {
switch ev.Kind {
case "task.state", "review.started":
return true
case "task.created", "task.merge_started", "task.merged", "task.discarded":
return humanActor(ev.Actor)
}
return false
}
func (m *Manager) freedText(ctx context.Context, owner string, ev api.Event) (string, error) {
if ev.Kind == "user.message" {
return m.waitingText(ctx, owner, ev)
+1 -1
View File
@@ -525,7 +525,7 @@ func (m *Manager) leadTools(l *live) []mcp.Tool {
}),
tool("compact_self",
"Ask byre to compact you when this turn ends: you write a handoff, and a fresh lead session continues from it. Call it only after byre signals that your context is past half of its window, at a logical point, then end the turn with a short final message.",
"Ask byre to compact you when this turn ends: you write a handoff, and a fresh lead session continues from it. Call it only after byre signals that your context is past a quarter of its window, at a logical point, then end the turn with a short final message.",
`{"type": "object", "properties": {}}`,
func(ctx context.Context, user api.User, raw json.RawMessage) (string, error) {
m.mu.Lock()