fix: persist and retry run alerts and honor notification modes

This commit is contained in:
2026-10-20 15:42:21 +01:00
parent e4f48ecbb1
commit 016fcb552a
20 changed files with 697 additions and 41 deletions
+80 -14
View File
@@ -3,6 +3,7 @@ package server
import (
"context"
"fmt"
"git.a-wan.net/alan/pipelines/internal/pipeline"
"git.a-wan.net/alan/pipelines/internal/protocol"
"git.a-wan.net/alan/pipelines/internal/stage"
"io"
@@ -34,10 +35,6 @@ func (r *runner) notifyStep(id int64, name, workspace string, policy *protocol.N
if !enabled {
return
}
// The stage's own policy already covers this attention event, so the
// run-level default must not repeat it. Recorded before delivery so a failed
// stage notification is reported by its own receipt, not silently duplicated.
r.markAttentionSent(id, step.Sequence, stageAttentionEvent(event))
ctx, cancel := context.WithTimeout(r.ctx, 30*time.Second)
defer cancel()
summary := fmt.Sprintf("%s · run #%d · stage %s · %s (%s)", name, id, step.Stage, event, step.Outcome)
@@ -128,6 +125,18 @@ func (r *runner) notifyStep(id int64, name, workspace string, policy *protocol.N
if !sentArtifact {
send("", nil)
}
// Only a successful publish covers the run alert. A failed stage publisher
// must not swallow the durable run-level fallback and its retries.
for _, receipt := range receipts {
if receipt.Status == "sent" {
covered := stageAttentionEvent(event)
if event == "success" && step.Target == protocol.SuccessEnd {
covered = "succeeded"
}
r.markAttentionSent(id, step.Sequence, covered)
break
}
}
}
// runAttentionEvent maps a run status to the operator attention notification it
@@ -167,8 +176,16 @@ func (r *runner) markAttentionSent(id int64, sequence int, event string) {
if event == "" {
return
}
if r.store != nil {
if err := r.store.CoverRunNotification(r.ctx, id, sequence, event); err != nil {
log.Printf("record stage notification coverage run=%d: %v", id, err)
}
}
r.attentionMu.Lock()
defer r.attentionMu.Unlock()
if r.attentionSent == nil {
r.attentionSent = map[int64]map[string]bool{}
}
if r.attentionSent[id] == nil {
r.attentionSent[id] = map[string]bool{}
}
@@ -206,6 +223,9 @@ func (r *runner) notifyMode(id int64) string {
if r.store != nil {
if stored, err := r.store.RunNotifyMode(r.ctx, id); err == nil && stored != "" {
r.notifyMu.Lock()
if r.notifyModes == nil {
r.notifyModes = map[int64]string{}
}
r.notifyModes[id] = stored
r.notifyMu.Unlock()
return stored
@@ -240,29 +260,32 @@ func runAttentionClick(id int64) string {
// notifyStep and reuses the same PublishNtfy machinery, destination and token
// handling, so no second publisher, topic or service is introduced. The run id
// and pipeline name identify the run and the click action opens it in the app.
// A run that needs nobody produces nothing, and a delivery failure is logged and
// never fails the run: this is a best-effort alert, not persisted run evidence.
// Attempts are persisted before publication. Publisher acknowledgement is not
// proof of device reception; a crash in that acknowledgement window can repeat.
func (r *runner) notifyRun(id int64, name, status string, sequence int, currentStage, diagnostic string) {
if r.notifyMode(id) == "off" {
return
}
event := runAttentionEvent(status)
if status == "succeeded" && r.notifyMode(id) == "on" {
event = "succeeded"
}
if event == "" {
return
}
if r.attentionAlready(id, sequence, event) {
return
}
// Record the delivery attempt before publishing so a repeated transition for
// the same run and event (a re-request, a restart-free re-entry) never
// republishes; delivery itself is never retried.
r.markAttentionSent(id, sequence, event)
title := "Run needs your decision"
summary := fmt.Sprintf("Run #%d · %s · waiting for a human decision", id, name)
if event == "failed" {
title = "Run failed"
summary = fmt.Sprintf("Run #%d · %s · run failed", id, name)
}
if event == "succeeded" {
title = "Run succeeded"
summary = fmt.Sprintf("Run #%d · %s · run succeeded", id, name)
}
if currentStage != "" {
summary += " at stage " + currentStage
}
@@ -270,10 +293,53 @@ func (r *runner) notifyRun(id int64, name, status string, sequence int, currentS
summary += ": " + diagnostic
}
summary = truncateUTF8(summary, 2000)
ctx, cancel := context.WithTimeout(r.ctx, 30*time.Second)
defer cancel()
if _, err := stage.PublishNtfy(ctx, "system", title, summary, "", nil, runAttentionClick(id)); err != nil {
log.Printf("run attention notification run=%d: %v", id, err)
if err := r.store.EnqueueRunNotification(r.ctx, pipeline.RunNotification{RunID: id, Sequence: sequence, Event: event, Channel: "system", Title: title, Body: summary, Click: runAttentionClick(id)}); err != nil {
log.Printf("enqueue run notification run=%d: %v", id, err)
return
}
r.deliverRunNotificationBatch(time.Now(), 1)
}
// resumeRunNotifications also recovers pending/expired leases after restart.
func (r *runner) resumeRunNotifications() {
ticker := time.NewTicker(time.Second)
defer ticker.Stop()
for {
select {
case <-r.ctx.Done():
return
case <-ticker.C:
r.deliverRunNotifications(time.Now())
}
}
}
func (r *runner) deliverRunNotifications(now time.Time) {
r.deliverRunNotificationBatch(now, 32)
}
func (r *runner) deliverRunNotificationBatch(now time.Time, limit int) {
// Bound one drain so a queue cannot monopolize the runner.
started := time.Now()
for i := 0; i < limit; i++ {
n, err := r.store.ClaimRunNotification(r.ctx, now.Add(time.Since(started)).Unix())
if err != nil {
log.Printf("claim run notification: %v", err)
return
}
if n == nil {
return
}
ctx, cancel := context.WithTimeout(r.ctx, 30*time.Second)
result, publishErr := stage.PublishNtfy(ctx, n.Channel, n.Title, n.Body, "", nil, n.Click)
cancel()
messageID, _ := result["message_id"].(string)
save, done := context.WithTimeout(context.Background(), 5*time.Second)
err = r.store.FinishRunNotification(save, *n, now.Add(time.Since(started)).Unix(), messageID, publishErr)
done()
if err != nil {
log.Printf("save run notification run=%d: %v", n.RunID, err)
}
}
}