fix(agent): task completion safety net — stop tasks sticking at Running

Implements fixes 1-3 of plans/2026-07-11-task-completion-safety-net.md.
Confirmed live that 50/50 production sessions never reached a terminal
status because the model almost never calls complete_task, even for
trivial single-tool Q&A turns SOUL.md explicitly calls out as needing it.

- Inline safety net (agent.go): a session that never called set_goal never
  framed itself as a structured task, so its first plain-text turn-end IS
  the task ending — auto-complete it there instead of leaving status stuck
  at its creation default forever.
- Idle sweep (continue.go, new completion_nudges column): goal-bearing
  sessions that stall get one nudge, then auto-close with outcome=partial
  if the nudge goes unanswered, mirroring the pattern resumeSession already
  uses for a different stuck-session failure mode.

Fix 4 (backfill of the 50 already-stuck live sessions) is deliberately
separate — deferred until this is deployed and verified live, per the
plan's implementation order.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

cmd/nomos/agent.go

@@ -205,6 +205,16 @@ func (a *agent) chatWith(ctx context.Context, sessionID, message, systemInject s
 			"[System: this task has been running long enough that only the most recent %d turns of its history are included above your context — earlier turns happened but aren't shown. If you need to know what was already tried or found, check search_knowledge/get_entity_knowledge (if you recorded it) rather than assuming it didn't happen.]",
 			historyWindowSize)))
 	}
+	// sawSetGoal / sawCompleteTask track whether this session has EVER framed
+	// itself as a structured task (set_goal) or already reached a terminal
+	// state (complete_task) — across both replayed history and this turn's
+	// own tool calls (updated again below as they happen live). Used by the
+	// end-of-turn safety net (plans/2026-07-11-task-completion-safety-net.md,
+	// fix 1): most sessions are a single trivial Q&A exchange that answers in
+	// text and never calls either tool, leaving agent_sessions.status stuck
+	// at its creation-time default forever. If a session never framed itself
+	// as a task, its first plain-text turn-end IS the task ending.
+	var sawSetGoal, sawCompleteTask bool
 	var lastAssistantCalls []persistedCall
 	for _, m := range history {
 		text := extractText(m.Content)
@@ -216,6 +226,12 @@ func (a *agent) chatWith(ctx context.Context, sessionID, message, systemInject s
 				messages = append(messages, assistantToolCallMessage(calls))
 				for _, c := range calls {
 					messages = append(messages, openai.ToolMessage(c.resultText(), c.id))
+					switch c.name {
+					case "set_goal":
+						sawSetGoal = true
+					case "complete_task":
+						sawCompleteTask = true
+					}
 				}
 				lastAssistantCalls = calls
 			}
@@ -358,6 +374,9 @@ func (a *agent) chatWith(ctx context.Context, sessionID, message, systemInject s
 
 		if len(msg.ToolCalls) == 0 {
 			emit(agentEvent{Type: "text", Data: msg.Content, SessionID: sessionID})
+			if !sawSetGoal && !sawCompleteTask {
+				a.autoCompleteTrivialTask(ctx, sessionID, msg.Content)
+			}
 			emit(agentEvent{Type: "done", Data: map[string]any{
 				"session_id":     sessionID,
 				"usage":          acc.Usage,
@@ -377,6 +396,13 @@ func (a *agent) chatWith(ctx context.Context, sessionID, message, systemInject s
 				args = map[string]any{}
 			}
 
+			switch tc.Function.Name {
+			case "set_goal":
+				sawSetGoal = true
+			case "complete_task":
+				sawCompleteTask = true
+			}
+
 			emit(agentEvent{
 				Type:      "tool_use",
 				Data:      map[string]any{"name": tc.Function.Name, "args": args, "id": tc.ID},

cmd/nomos/continue.go

@@ -32,6 +32,71 @@ func extractExecutionIDs(toolResult string) []uuid.UUID {
 	return out
 }
 
+// idleTaskThreshold is how long a goal-bearing session can sit non-terminal
+// with no activity before the idle sweep nudges it, per
+// plans/2026-07-11-task-completion-safety-net.md. Arbitrary starting point,
+// not measured against real task durations — long enough that it won't fire
+// mid-turn, short enough the board doesn't lie for hours.
+const idleTaskThreshold = 15 * time.Minute
+
+// runIdleSweepWorker is the safety net for case 2 of
+// plans/2026-07-11-task-completion-safety-net.md: sessions that called
+// set_goal (so the inline safety net in agent.go correctly left them alone,
+// since they framed themselves as a real task) but then stalled without
+// ever calling complete_task. Coarser than runContinuationWorker's 4s tick
+// since "gone idle" is a much slower signal than "an execution just
+// finished." Blocks until ctx is cancelled.
+func (a *agent) runIdleSweepWorker(ctx context.Context) {
+	if a.store == nil {
+		slog.Warn("nomos: idle sweep worker disabled (no store)")
+		return
+	}
+	slog.Info("nomos: idle sweep worker started")
+	ticker := time.NewTicker(5 * time.Minute)
+	defer ticker.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+			a.processIdleSweep(ctx)
+		}
+	}
+}
+
+// processIdleSweep nudges a stalled goal-bearing session once; if it's still
+// non-terminal on the NEXT sweep (meaning the nudge itself went unanswered,
+// not just that the model is still working), auto-closes it with a
+// visible "auto-closed" outcome instead of leaving it stuck forever — same
+// reasoning resumeSession already applies below for a different failure
+// mode (a resume that produces no response at all).
+func (a *agent) processIdleSweep(ctx context.Context) {
+	stale := a.store.staleGoalSessions(ctx, idleTaskThreshold, 5)
+	for _, s := range stale {
+		s := s
+		if s.CompletionNudges == 0 {
+			safego.Go("nomos:idle-nudge:"+s.ID, func() {
+				if err := a.store.bumpCompletionNudge(ctx, s.ID); err != nil {
+					slog.Error("nomos: idle nudge bump failed", "session", s.ID, "error", err)
+					return
+				}
+				note := fmt.Sprintf("[System: this task ('%s') has been idle for %s with no complete_task call. "+
+					"If the goal is done (or can't be completed), call complete_task now with the outcome and a "+
+					"one-line summary. If you're still genuinely working through the plan, ignore this and continue.]",
+					s.Goal, idleTaskThreshold)
+				a.resumeSession(ctx, s.ID, note)
+			})
+			continue
+		}
+		safego.Go("nomos:idle-autoclose:"+s.ID, func() {
+			summary := fmt.Sprintf("Auto-closed after %s idle with no response to a completion nudge.", idleTaskThreshold)
+			if err := a.store.completeTask(ctx, s.ID, "partial", summary); err != nil {
+				slog.Error("nomos: idle auto-close failed", "session", s.ID, "error", err)
+			}
+		})
+	}
+}
+
 // runContinuationWorker is the event loop that replaces the human typing
 // "continue". It polls for gated executions that (a) were initiated by a chat
 // session and (b) have just finished, and — while that agent has an open assent

cmd/nomos/main.go

@@ -80,6 +80,11 @@ func main() {
 		// from failures) without the operator ticking it forward each step.
 		safego.Go("nomos:continuation-worker", func() { nAgent.runContinuationWorker(ctx) })
 
+		// Idle sweep for stalled goal-bearing tasks (fix 2+3 of
+		// plans/2026-07-11-task-completion-safety-net.md) — a coarser,
+		// slower-ticking counterpart to the continuation worker above.
+		safego.Go("nomos:idle-sweep-worker", func() { nAgent.runIdleSweepWorker(ctx) })
+
 		safego.Go("nomos:mcp-pool-sweeper", func() {
 			ticker := time.NewTicker(5 * time.Minute)
 			defer ticker.Stop()

cmd/nomos/store.go

@@ -508,6 +508,60 @@ func (s *store) completeTask(ctx context.Context, sessionID, outcome, summary st
 	return nil
 }
 
+// staleGoalSession is a goal-bearing task that's gone idle without reaching
+// a terminal state — the idle-sweep worker's work list (fix 2+3 of
+// plans/2026-07-11-task-completion-safety-net.md).
+type staleGoalSession struct {
+	ID               string
+	Goal             string
+	CompletionNudges int
+}
+
+// staleGoalSessions finds sessions that framed themselves as a real task
+// (goal != '', so the inline safety net in agent.go intentionally left them
+// alone) but have sat non-terminal past idleThreshold. completion_nudges
+// tells the caller whether to nudge (0) or give up and auto-close (>=1) —
+// see processIdleSweep in continue.go.
+func (s *store) staleGoalSessions(ctx context.Context, idleThreshold time.Duration, limit int) []staleGoalSession {
+	if s == nil {
+		return nil
+	}
+	rows, err := s.pool.Query(ctx, `
+		SELECT id, goal, completion_nudges
+		FROM agent_sessions
+		WHERE goal <> ''
+		  AND status IN ('active', 'planning', 'executing')
+		  AND last_active_at < now() - ($1 * interval '1 second')
+		ORDER BY last_active_at
+		LIMIT $2`, idleThreshold.Seconds(), limit)
+	if err != nil {
+		return nil
+	}
+	defer rows.Close()
+	var out []staleGoalSession
+	for rows.Next() {
+		var s staleGoalSession
+		if err := rows.Scan(&s.ID, &s.Goal, &s.CompletionNudges); err == nil {
+			out = append(out, s)
+		}
+	}
+	return out
+}
+
+// bumpCompletionNudge records that the idle sweep nudged a stalled session,
+// stamping last_active_at so it isn't picked up again until it's genuinely
+// idle again (a fresh nudge shouldn't fire every tick while the model is
+// mid-response to the previous one).
+func (s *store) bumpCompletionNudge(ctx context.Context, sessionID string) error {
+	if s == nil {
+		return nil
+	}
+	_, err := s.pool.Exec(ctx, `
+		UPDATE agent_sessions SET completion_nudges = completion_nudges + 1, last_active_at = now()
+		WHERE id = $1`, sessionID)
+	return err
+}
+
 // planStep is a persisted plan step, as returned to the frontend for hydration
 // (the panel otherwise only sees steps live via plan.proposed/plan.step.*).
 type planStep struct {

cmd/nomos/tasks.go

@@ -266,3 +266,28 @@ func (a *agent) handleTaskTool(ctx context.Context, sessionID, name string, args
 		return nil, false
 	}
 }
+
+// autoCompleteTrivialTask is the case-1 fix from
+// plans/2026-07-11-task-completion-safety-net.md: a session that never
+// called set_goal never framed itself as a structured task, so a turn that
+// ends with a plain-text answer and no further tool calls IS the task
+// ending — but the model consistently skips complete_task for exactly this
+// case (confirmed live: 43/50 production sessions were a single trivial
+// Q&A exchange, none of which ever reached a terminal status). Rather than
+// leave agent_sessions.status stuck at its creation-time default forever,
+// close it out mechanically here: no judgment call needed, since SOUL.md
+// already treats a one-shot answered question as done by definition.
+func (a *agent) autoCompleteTrivialTask(ctx context.Context, sessionID, responseText string) {
+	summary := strings.TrimSpace(responseText)
+	summary = strings.SplitN(summary, "\n", 2)[0] // first line only — the board shows one line
+	const maxLen = 120
+	if len(summary) > maxLen {
+		summary = summary[:maxLen] + "…"
+	}
+	if summary == "" {
+		summary = "Answered without further action needed."
+	}
+	if err := a.store.completeTask(ctx, sessionID, "success", summary); err != nil {
+		slog.Error("nomos: auto-complete trivial task failed", "session", sessionID, "error", err)
+	}
+}