fix(server): sum task LLM usage across retries and failover

The logical call now returns the usage of every physical attempt (atomic and stream), keeps the usage of failed uncommitted stream attempts without delivering their events, normalizes per-attempt echoes with MergeEcho, stops with ctx.Err() when cancelled during backoff, and treats signed/redacted thinking and tool-use signatures as committing output. The llm_usage ledger stays at one row per physical call.

.github/workflows/ci.yml

@@ -29,6 +29,10 @@ jobs:
       - uses: actions/setup-go@v5
         with:
           go-version-file: go.mod
+      - name: gofmt
+        run: |
+          files=$(gofmt -l $(git ls-files '*.go'))
+          if [ -n "$files" ]; then echo "gofmt required:"; echo "$files"; exit 1; fi
       - run: go build ./...
       - run: go vet ./...
       # Skipped tests, tracked in issue #9:

agent/chat.go

@@ -169,9 +169,12 @@ func (c *ChatAgent) Chat(ctx context.Context, agentKey, sessionID, message strin
 	if c.guard != nil {
 		opts.Hooks = c.guard.Hooks()
 	}
+	// The session id is the provider-side identity (session header, usage metering)
+	// with or without persistence; Norma only puts it on ctx when it writes a transcript.
+	opts.SessionID = sessionID
+	ctx = transcript.WithSessionID(ctx, opts.SessionID)
 	if c.tx != nil { // persist raw human↔AI conversation; one accumulating file per thread
 		opts.Transcript = c.tx
-		opts.SessionID = sessionID
 	}
 	// experimental feature: when enabled, noa takes over context compression (archive centralized under <workDir>/noa/<SessionID>, persistent).
 	enableNoa(&opts, c.noaEnabledFn, c.workDir, "chat-"+sessionID, noaWarn("chat-"+sessionID))
@@ -180,8 +183,11 @@ func (c *ChatAgent) Chat(ctx context.Context, agentKey, sessionID, message strin
 	defer s.Close()
 	// reload prior conversation so the agent has context across turns (each Chat is
 	// a fresh session). First turn: no file yet → Resume loads nothing and proceeds.
+	// A corrupt or unreadable transcript stops the turn before any model or tool call.
 	if c.tx != nil {
-		_ = s.Resume(sessionID)
+		if err := resumeSession(s, opts.SessionID); err != nil {
+			return "", err
+		}
 	}
 	// re-unlock skill-gated MCPs from prior Skill() calls in the reloaded history so
 	// revealed tools stay callable across the fresh session.

agent/mainagent.go

@@ -169,14 +169,17 @@ func (m *MainAgent) Chat(ctx context.Context, taskID int64, mainSeg int, as *db.
 	if hooks != nil { // typed-nil guard: only set when concrete (avoids harness panic)
 		opts.Hooks = hooks
 	}
+	// Segment 0 keeps the legacy "exp%d-main" name so existing transcripts still
+	// load; each new session (seg>=1) gets its own file for a clean context. The id
+	// is the provider-side identity (session header, usage metering) with or without
+	// persistence; Norma only puts it on ctx when it writes a transcript.
+	opts.SessionID = fmt.Sprintf("exp%d-main", ts.ID())
+	if mainSeg > 0 {
+		opts.SessionID = fmt.Sprintf("exp%d-main-s%d", ts.ID(), mainSeg)
+	}
+	ctx = transcript.WithSessionID(ctx, opts.SessionID)
 	if m.tx != nil { // persist raw human↔AI conversation; one accumulating file per segment
 		opts.Transcript = m.tx
-		// Segment 0 keeps the legacy "exp%d-main" name so existing transcripts still
-		// load; each new session (seg>=1) gets its own file for a clean context.
-		opts.SessionID = fmt.Sprintf("exp%d-main", ts.ID())
-		if mainSeg > 0 {
-			opts.SessionID = fmt.Sprintf("exp%d-main-s%d", ts.ID(), mainSeg)
-		}
 	}
 	// experimental feature: when enabled, noa takes over context compression (archive centralized under <workDir>/noa/<SessionID>, persistent).
 	// the session id follows the same rule as the transcript (segment-aware), keeping archive and resume aligned.
@@ -191,8 +194,11 @@ func (m *MainAgent) Chat(ctx context.Context, taskID int64, mainSeg int, as *db.
 	// reload the prior conversation from the transcript so the agent has context
 	// across turns (each Chat is a fresh session; without this it can't see earlier
 	// messages). First turn: no file yet → Resume loads nothing and proceeds.
+	// A corrupt or unreadable transcript stops the turn before any model or tool call.
 	if m.tx != nil {
-		_ = s.Resume(opts.SessionID)
+		if err := resumeSession(s, opts.SessionID); err != nil {
+			return "", err
+		}
 	}
 	// C2: this session is fresh each turn; re-unlock skill-gated MCPs from prior
 	// Skill() calls in the reloaded history so revealed tools stay callable.

agent/planner.go

@@ -567,9 +567,12 @@ func (p *Planner) Plan(ctx context.Context, taskID int64, as *db.AssetStore, ts
 	if hooks != nil { // typed-nil guard: only set when concrete (avoids harness panic)
 		opts.Hooks = hooks
 	}
+	// The session id is the provider-side identity (session header, usage metering)
+	// with or without persistence; Norma only puts it on ctx when it writes a transcript.
+	opts.SessionID = fmt.Sprintf("exp%d-planner", ts.ID())
+	ctx = transcript.WithSessionID(ctx, opts.SessionID)
 	if p.tx != nil { // persist raw LLM conversation; one accumulating file per task's planner
 		opts.Transcript = p.tx
-		opts.SessionID = fmt.Sprintf("exp%d-planner", ts.ID())
 	}
 	// Experimental: when on, noa takes over context compression (archives persist under <workDir>/noa/<SessionID>).
 	noaSession := fmt.Sprintf("exp%d-planner", ts.ID())

agent/session.go

@@ -0,0 +1,19 @@
+package agent
+
+import (
+	"fmt"
+
+	"github.com/Autumn-27/norma/agentcore"
+)
+
+// resumeSession reloads the persisted conversation of session id into s. Norma's
+// transcript.Store.Load treats a missing file as a fresh session (no records, no
+// error); any other failure, such as a corrupt line or an unreadable path, is
+// returned so the caller stops before the first model or tool call instead of
+// appending a new conversation to the damaged file.
+func resumeSession(s *agentcore.Session, id string) error {
+	if err := s.Resume(id); err != nil {
+		return fmt.Errorf("restore Norma session %q: %w", id, err)
+	}
+	return nil
+}

agent/session_test.go

@@ -0,0 +1,305 @@
+package agent
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"iter"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+
+	"github.com/Autumn-27/artex/db"
+	"github.com/Autumn-27/norma/llm"
+	"github.com/Autumn-27/norma/transcript"
+)
+
+// sessionProbe is a provider that answers every call with one text turn and
+// records the session id it saw on ctx plus the request messages.
+type sessionProbe struct {
+	mu   sync.Mutex
+	ids  []string
+	msgs [][]llm.Message
+}
+
+func (p *sessionProbe) Stream(ctx context.Context, req llm.CompletionRequest) iter.Seq2[llm.StreamEvent, error] {
+	p.mu.Lock()
+	p.ids = append(p.ids, transcript.SessionIDFrom(ctx))
+	p.msgs = append(p.msgs, req.Messages)
+	p.mu.Unlock()
+	return func(yield func(llm.StreamEvent, error) bool) {
+		if yield(llm.StreamEvent{Type: llm.SETextDelta, Text: "done"}, nil) {
+			yield(llm.StreamEvent{Type: llm.SEMessageDelta, StopReason: "end_turn"}, nil)
+		}
+	}
+}
+
+func (p *sessionProbe) Complete(ctx context.Context, req llm.CompletionRequest) (llm.Message, string, llm.Usage, error) {
+	acc := llm.NewAccumulator()
+	for ev, err := range p.Stream(ctx, req) {
+		if err != nil {
+			return llm.Message{}, "", llm.Usage{}, err
+		}
+		acc.Add(ev)
+	}
+	return acc.Message(), acc.StopReason, acc.Usage, nil
+}
+
+func (p *sessionProbe) seen() []string {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	return append([]string(nil), p.ids...)
+}
+
+// jsonlFiles lists every .jsonl file under dir.
+func jsonlFiles(t *testing.T, dir string) []string {
+	t.Helper()
+	var out []string
+	err := filepath.WalkDir(dir, func(path string, d os.DirEntry, err error) error {
+		if err != nil {
+			return err
+		}
+		if !d.IsDir() && strings.HasSuffix(path, ".jsonl") {
+			out = append(out, path)
+		}
+		return nil
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	return out
+}
+
+// damageTranscript makes the main transcript of id corrupt (a non-JSON line) or
+// unreadable (a directory). It returns the bytes a corrupt file must keep.
+func damageTranscript(t *testing.T, store *transcript.Store, id string, unreadable bool) string {
+	t.Helper()
+	path := store.MainPath(id)
+	if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if unreadable {
+		if err := os.Mkdir(path, 0o700); err != nil {
+			t.Fatal(err)
+		}
+		return ""
+	}
+	const original = "{corrupt transcript\n"
+	if err := os.WriteFile(path, []byte(original), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	return original
+}
+
+// assertTranscriptUntouched checks that a damaged transcript was neither
+// rewritten nor appended to.
+func assertTranscriptUntouched(t *testing.T, store *transcript.Store, id, original string, unreadable bool) {
+	t.Helper()
+	path := store.MainPath(id)
+	if unreadable {
+		entries, err := os.ReadDir(path)
+		if err != nil || len(entries) != 0 {
+			t.Fatalf("unreadable transcript changed: entries=%v err=%v", entries, err)
+		}
+		return
+	}
+	b, err := os.ReadFile(path)
+	if err != nil || string(b) != original {
+		t.Fatalf("damaged transcript changed: %q err=%v", b, err)
+	}
+}
+
+func assertRestoreError(t *testing.T, err error, calls int) {
+	t.Helper()
+	if err == nil || !strings.Contains(err.Error(), "restore Norma session") || calls != 0 {
+		t.Fatalf("err=%v provider calls=%d", err, calls)
+	}
+}
+
+var damageCases = []struct {
+	name       string
+	unreadable bool
+}{{"corrupt", false}, {"unreadable-directory", true}}
+
+func TestChatSessionIdentityWithoutTranscript(t *testing.T) {
+	p := &sessionProbe{}
+	workDir := t.TempDir()
+	c := NewChatAgent(p, "test-model", workDir, nil, 100000)
+	for range 2 {
+		if _, err := c.Chat(t.Context(), "assistant", "conv-stable", "hello", 1, 0, false, nil); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if ids := p.seen(); len(ids) != 2 || ids[0] != "conv-stable" || ids[1] != "conv-stable" {
+		t.Fatalf("provider session ids: %v", ids)
+	}
+	if files := jsonlFiles(t, workDir); len(files) != 0 {
+		t.Fatalf("transcript written without persistence: %v", files)
+	}
+}
+
+func TestChatMissingTranscriptStartsFresh(t *testing.T) {
+	store := transcript.NewStore(t.TempDir())
+	p := &sessionProbe{}
+	c := NewChatAgent(p, "test-model", t.TempDir(), store, 100000)
+	if _, err := c.Chat(t.Context(), "assistant", "conv-new", "hello", 1, 0, false, nil); err != nil {
+		t.Fatal(err)
+	}
+	if ids := p.seen(); len(ids) != 1 || ids[0] != "conv-new" {
+		t.Fatalf("provider session ids: %v", ids)
+	}
+	recs, err := store.Load(store.MainPath("conv-new"))
+	if err != nil || len(recs) == 0 {
+		t.Fatalf("fresh transcript not written: %d records, err=%v", len(recs), err)
+	}
+}
+
+func TestChatRefusesDamagedTranscript(t *testing.T) {
+	for _, tc := range damageCases {
+		t.Run(tc.name, func(t *testing.T) {
+			store := transcript.NewStore(t.TempDir())
+			original := damageTranscript(t, store, "conv-broken", tc.unreadable)
+			p := &sessionProbe{}
+			c := NewChatAgent(p, "test-model", t.TempDir(), store, 100000)
+			_, err := c.Chat(t.Context(), "assistant", "conv-broken", "hello", 1, 0, false, nil)
+			assertRestoreError(t, err, len(p.seen()))
+			assertTranscriptUntouched(t, store, "conv-broken", original, tc.unreadable)
+		})
+	}
+}
+
+func TestChatResumesSignedThinkingHistory(t *testing.T) {
+	store := transcript.NewStore(t.TempDir())
+	w := store.NewWriter("conv-signed", "")
+	w.RecordMessage(llm.UserText("earlier question"), llm.Usage{})
+	w.RecordMessage(llm.Message{Role: llm.RoleAssistant, Content: []llm.ContentBlock{
+		{Type: llm.BlockThinking, Thinking: "reasoning", Signature: "signed"},
+		{Type: llm.BlockRedactedThinking, Data: "opaque"},
+		{Type: llm.BlockText, Text: "earlier answer"},
+	}}, llm.Usage{})
+	if err := w.Err(); err != nil {
+		t.Fatal(err)
+	}
+	p := &sessionProbe{}
+	c := NewChatAgent(p, "test-model", t.TempDir(), store, 100000)
+	if _, err := c.Chat(t.Context(), "assistant", "conv-signed", "next question", 1, 0, false, nil); err != nil {
+		t.Fatal(err)
+	}
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	if len(p.msgs) != 1 {
+		t.Fatalf("provider calls: %d", len(p.msgs))
+	}
+	var thinking, redacted bool
+	for _, m := range p.msgs[0] {
+		for _, b := range m.Content {
+			thinking = thinking || (b.Type == llm.BlockThinking && b.Thinking == "reasoning" && b.Signature == "signed")
+			redacted = redacted || (b.Type == llm.BlockRedactedThinking && b.Data == "opaque")
+		}
+	}
+	if !thinking || !redacted {
+		t.Fatalf("resumed history lost signed thinking=%v redacted=%v: %+v", thinking, redacted, p.msgs[0])
+	}
+}
+
+// sessionFixture is one exploration with one intent on the test database.
+type sessionFixture struct {
+	expID    int64
+	intentID int64
+	ts       *db.ExplorationStore
+	intent   *db.Node
+}
+
+func newSessionFixture(t *testing.T) sessionFixture {
+	t.Helper()
+	d := testDB(t)
+	t.Cleanup(func() { _ = d.Close() })
+	expID, err := d.CreateExploration("session identity probe", "keep the session id stable")
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { _, _ = d.Exec(`DELETE FROM explorations WHERE id=$1`, expID) })
+	ts := d.Exploration(expID)
+	payload := map[string]any{"summary": "session probe"}
+	intentID, err := ts.AddNode("intent", payload, 0, "running", "planner", nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	raw, _ := json.Marshal(payload)
+	return sessionFixture{expID: expID, intentID: intentID, ts: ts, intent: &db.Node{ID: intentID, Payload: raw}}
+}
+
+// run drives one agent kind once and returns the session id it should use.
+func (f sessionFixture) run(t *testing.T, kind string, p *sessionProbe, tx *transcript.Store) (string, error) {
+	t.Helper()
+	workDir := t.TempDir()
+	switch kind {
+	case "worker":
+		_, _, err := NewWorker(p, "test-model", workDir, tx, 0, 3).Execute(t.Context(), "work#1", f.expID, nil, f.ts, f.intent, passHooks{}, nil, nil, nil)
+		return WorkerSessionID(f.expID, f.intentID), err
+	case "planner":
+		_, _, err := NewPlanner(p, "test-model", workDir, tx, 0, 3).Plan(t.Context(), f.expID, nil, f.ts, "probe goal", nil, passHooks{}, nil)
+		return fmt.Sprintf("exp%d-planner", f.expID), err
+	case "mainagent":
+		_, err := NewMainAgent(p, "test-model", workDir, tx, 0, 3).Chat(t.Context(), f.expID, 0, nil, f.ts, "probe goal", "status?", passHooks{}, nil, nil, nil, nil, nil)
+		return fmt.Sprintf("exp%d-main", f.expID), err
+	case "mainagent-seg":
+		_, err := NewMainAgent(p, "test-model", workDir, tx, 0, 3).Chat(t.Context(), f.expID, 2, nil, f.ts, "probe goal", "status?", passHooks{}, nil, nil, nil, nil, nil)
+		return fmt.Sprintf("exp%d-main-s2", f.expID), err
+	}
+	t.Fatalf("unknown agent kind %q", kind)
+	return "", nil
+}
+
+func TestAgentSessionIdentityWithAndWithoutTranscript(t *testing.T) {
+	f := newSessionFixture(t)
+	for _, kind := range []string{"worker", "planner", "mainagent", "mainagent-seg"} {
+		t.Run(kind, func(t *testing.T) {
+			bare := &sessionProbe{}
+			want, err := f.run(t, kind, bare, nil)
+			if err != nil {
+				t.Fatal(err)
+			}
+			store := transcript.NewStore(t.TempDir())
+			persisted := &sessionProbe{}
+			if _, err := f.run(t, kind, persisted, store); err != nil {
+				t.Fatal(err)
+			}
+			for _, ids := range [][]string{bare.seen(), persisted.seen()} {
+				if len(ids) == 0 {
+					t.Fatal("provider never called")
+				}
+				for _, id := range ids {
+					if id != want {
+						t.Fatalf("provider session ids %v, want %q", ids, want)
+					}
+				}
+			}
+			if recs, err := store.Load(store.MainPath(want)); err != nil || len(recs) == 0 {
+				t.Fatalf("transcript %q not written under the session id: %d records, err=%v", want, len(recs), err)
+			}
+		})
+	}
+}
+
+func TestAgentRefusesDamagedTranscript(t *testing.T) {
+	f := newSessionFixture(t)
+	for _, kind := range []string{"worker", "mainagent"} {
+		for _, tc := range damageCases {
+			t.Run(kind+"/"+tc.name, func(t *testing.T) {
+				store := transcript.NewStore(t.TempDir())
+				id := WorkerSessionID(f.expID, f.intentID)
+				if kind == "mainagent" {
+					id = fmt.Sprintf("exp%d-main", f.expID)
+				}
+				original := damageTranscript(t, store, id, tc.unreadable)
+				p := &sessionProbe{}
+				_, err := f.run(t, kind, p, store)
+				assertRestoreError(t, err, len(p.seen()))
+				assertTranscriptUntouched(t, store, id, original, tc.unreadable)
+			})
+		}
+	}
+}

agent/tools.go

@@ -46,9 +46,9 @@ const stagnationMinRepeats = 3
 
 func detectStagnation(done, findings []*db.Node) []map[string]any {
 	type dir struct {
-		firstID  int64
-		repeats  int
-		summary  string
+		firstID int64
+		repeats int
+		summary string
 	}
 	dirs := map[string]*dir{}
 	order := []string{}
@@ -551,7 +551,7 @@ func (t *ToolSet) graphOverviewData() map[string]any {
 	// with the data so no prompt-template change is needed on deployed installs.
 	if stale := detectStagnation(recentDone, vulnNodes); len(stale) > 0 {
 		out["stagnation_warning"] = map[string]any{
-			"directions": stale,
+			"directions":  stale,
 			"instruction": "these directions were dispatched repeatedly with zero findings produced since their first attempt — do NOT re-dispatch the same approach again; open a materially different direction, or judge the goal on the existing evidence instead of grinding",
 		}
 	}

agent/worker.go

@@ -540,9 +540,12 @@ func (w *Worker) execute(ctx context.Context, name string, taskID int64, as *db.
 	if hooks != nil { // typed-nil guard: only set when concrete (avoids harness panic)
 		opts.Hooks = hooks
 	}
+	// The session id is the provider-side identity (session header, usage metering)
+	// with or without persistence; Norma only puts it on ctx when it writes a transcript.
+	opts.SessionID = WorkerSessionID(ts.ID(), intent.ID)
+	ctx = transcript.WithSessionID(ctx, opts.SessionID)
 	if w.tx != nil { // persist raw LLM conversation; one file per worked intent
 		opts.Transcript = w.tx
-		opts.SessionID = WorkerSessionID(ts.ID(), intent.ID)
 	}
 	intentID := intent.ID
 	emitWrap := func(r db.Activity) {
@@ -574,7 +577,12 @@ func (w *Worker) execute(ctx context.Context, name string, taskID int64, as *db.
 	// restarting from scratch.
 	alreadyRecorded := false
 	if w.tx != nil {
-		_ = s.Resume(opts.SessionID)
+		// A corrupt or unreadable transcript stops the run before any model or tool
+		// call. No terminal reason: the engine marks the intent blocked without the
+		// model_error reruns, which would only fail the same way again.
+		if err := resumeSession(s, opts.SessionID); err != nil {
+			return "", tsx.Writes(), err
+		}
 		alreadyRecorded = requestID != "" && hasWorkerChatMessage(s.Messages(), requestID)
 		if len(s.Messages()) > 0 && message == "" {
 			seedUnlockFromHistory(s.Messages(), def.UnlockSkill)

db/asset_intercept.go

@@ -7,12 +7,12 @@ import "time"
 // these match the *target asset*: an exact/fuzzy domain·ip·url, or a CIDR range.
 // This layer only stores rules; the matching/enforcement logic lives elsewhere.
 type AssetInterceptRule struct {
-	ID        int64     `json:"id"`
-	Enabled   bool      `json:"enabled"`
-	Kind      string    `json:"kind"` // exact_domain|exact_ip|exact_url|fuzzy_domain|fuzzy_ip|fuzzy_url|cidr
-	Pattern   string    `json:"pattern"`
-	Note      string    `json:"note"`
-	Builtin   bool      `json:"builtin"`
+	ID      int64  `json:"id"`
+	Enabled bool   `json:"enabled"`
+	Kind    string `json:"kind"` // exact_domain|exact_ip|exact_url|fuzzy_domain|fuzzy_ip|fuzzy_url|cidr
+	Pattern string `json:"pattern"`
+	Note    string `json:"note"`
+	Builtin bool   `json:"builtin"`
 	// Action is only used by task-level rules: 'block'=intercept, 'allow'=allow (whitelist).
 	// Global rules (asset_intercept_rules) do not have this column; it is always empty and treated as intercept.
 	Action    string    `json:"action,omitempty"`

db/assets_test.go

@@ -1168,11 +1168,11 @@ func TestQueryDSLInScopeMembership(t *testing.T) {
 	}
 
 	stamp := time.Now().UnixNano()
-	root := fmt.Sprintf("sc%d.invalid", stamp)           // in-scope root domain (task)
-	srcRoot := fmt.Sprintf("src%d.invalid", stamp)       // in-scope via source task
-	out := fmt.Sprintf("out%d.invalid", stamp)           // out of every scope
-	marker := fmt.Sprintf("mk%d", stamp)                 // bare-text token present in all rows
-	foreignTask := stamp + 777                           // asset produced by an unrelated task
+	root := fmt.Sprintf("sc%d.invalid", stamp)     // in-scope root domain (task)
+	srcRoot := fmt.Sprintf("src%d.invalid", stamp) // in-scope via source task
+	out := fmt.Sprintf("out%d.invalid", stamp)     // out of every scope
+	marker := fmt.Sprintf("mk%d", stamp)           // bare-text token present in all rows
+	foreignTask := stamp + 777                     // asset produced by an unrelated task
 
 	// Scope: task owns root; source task owns srcRoot; task owns an IP /24.
 	for _, sc := range []struct {
@@ -1200,7 +1200,7 @@ func TestQueryDSLInScopeMembership(t *testing.T) {
 		{"subdomain", "api." + root, root, "", "", "", marker},                        // under task root
 		{"service", "www." + root, root, "https://www." + root + "/", "", "", marker}, // service under task root
 		{"endpoint", "www." + root, root, "https://www." + root + "/a?" + marker + "=1", "GET", "", ""},
-		{"subdomain", "dev." + srcRoot, srcRoot, "", "", "", marker}, // under source-task root
+		{"subdomain", "dev." + srcRoot, srcRoot, "", "", "", marker},                                      // under source-task root
 		{"endpoint", "198.51.100.9", "198.51.100.9", "http://198.51.100.9:8080/" + marker, "GET", "", ""}, // IP-literal host, ip col empty
 		{"subdomain", "x." + out, out, "", "", "", marker},                                                // out of scope
 	}

db/exploration.go

@@ -1486,11 +1486,13 @@ type TaskListMetrics struct {
 	Findings       FindingSeverityCounts // this task's vulnerability count in the findings table (bucketed by severity)
 }
 
-// TaskListMetricsAll returns list aggregates for every live task in one query.
-// The lateral lookups use the per-exploration indexes instead of grouping the
-// complete activity history on every poll.
-func (d *DB) TaskListMetricsAll() (map[int64]TaskListMetrics, error) {
-	rows, err := d.Query(`
+// taskListMetricsQuery is the select shared by TaskListMetricsAll and
+// TaskListMetrics; each appends its own WHERE on task. The lateral lookups use
+// the per-exploration indexes instead of grouping the complete activity history
+// on every poll. latest_activity orders by the column activity.created_at, so
+// idx_act_latest (exploration_id, created_at DESC) serves the LIMIT 1; ordering
+// by the output alias would sort every activity row of the task.
+const taskListMetricsQuery = `
 		SELECT task.exploration_id,
 		       COALESCE(token_metrics.input_tokens,0),
 		       COALESCE(token_metrics.output_tokens,0),
@@ -1514,10 +1516,10 @@ func (d *DB) TaskListMetricsAll() (map[int64]TaskListMetrics, error) {
 			WHERE exploration_id=task.exploration_id AND kind='result'
 		) token_metrics ON true
 		LEFT JOIN LATERAL (
-			SELECT EXTRACT(EPOCH FROM created_at)::bigint AS created_at
+			SELECT EXTRACT(EPOCH FROM activity.created_at)::bigint AS created_at
 			FROM activity
-			WHERE exploration_id=task.exploration_id
-			ORDER BY created_at DESC
+			WHERE activity.exploration_id=task.exploration_id
+			ORDER BY activity.created_at DESC
 			LIMIT 1
 		) latest_activity ON true
 		LEFT JOIN LATERAL (
@@ -1539,30 +1541,41 @@ func (d *DB) TaskListMetricsAll() (map[int64]TaskListMetrics, error) {
 			FROM findings
 			WHERE task_id=task.id
 		) finding_metrics ON true
-		WHERE task.deleted_at IS NULL`)
+		WHERE task.deleted_at IS NULL`
+
+// scanTaskListMetrics reads one row of taskListMetricsQuery.
+func scanTaskListMetrics(row interface{ Scan(...any) error }) (int64, TaskListMetrics, error) {
+	var explorationID int64
+	var metrics TaskListMetrics
+	err := row.Scan(
+		&explorationID,
+		&metrics.Tokens.InputTokens,
+		&metrics.Tokens.OutputTokens,
+		&metrics.Tokens.CacheReadTokens,
+		&metrics.Tokens.CacheWriteTokens,
+		&metrics.LastActivity,
+		&metrics.Goals.Total,
+		&metrics.Goals.Met,
+		&metrics.RunningIntents,
+		&metrics.Findings.Critical,
+		&metrics.Findings.High,
+		&metrics.Findings.Medium,
+		&metrics.Findings.Low,
+	)
+	return explorationID, metrics, err
+}
+
+// TaskListMetricsAll returns list aggregates for every live task in one query.
+func (d *DB) TaskListMetricsAll() (map[int64]TaskListMetrics, error) {
+	rows, err := d.Query(taskListMetricsQuery)
 	if err != nil {
 		return nil, err
 	}
 	defer rows.Close()
 	out := map[int64]TaskListMetrics{}
 	for rows.Next() {
-		var explorationID int64
-		var metrics TaskListMetrics
-		if err := rows.Scan(
-			&explorationID,
-			&metrics.Tokens.InputTokens,
-			&metrics.Tokens.OutputTokens,
-			&metrics.Tokens.CacheReadTokens,
-			&metrics.Tokens.CacheWriteTokens,
-			&metrics.LastActivity,
-			&metrics.Goals.Total,
-			&metrics.Goals.Met,
-			&metrics.RunningIntents,
-			&metrics.Findings.Critical,
-			&metrics.Findings.High,
-			&metrics.Findings.Medium,
-			&metrics.Findings.Low,
-		); err != nil {
+		explorationID, metrics, err := scanTaskListMetrics(rows)
+		if err != nil {
 			return nil, err
 		}
 		out[explorationID] = metrics
@@ -1570,6 +1583,17 @@ func (d *DB) TaskListMetricsAll() (map[int64]TaskListMetrics, error) {
 	return out, rows.Err()
 }
 
+// TaskListMetrics returns the same aggregates for the one live task of
+// exploration expID, with the same laterals as TaskListMetricsAll. A missing or
+// deleted task gives zero metrics, as a missing key of TaskListMetricsAll does.
+func (d *DB) TaskListMetrics(expID int64) (TaskListMetrics, error) {
+	_, metrics, err := scanTaskListMetrics(d.QueryRow(taskListMetricsQuery+` AND task.exploration_id=$1`, expID))
+	if errors.Is(err, sql.ErrNoRows) {
+		return TaskListMetrics{}, nil
+	}
+	return metrics, err
+}
+
 // GoalCountsAll returns goal totals (total / met) per exploration id, one query
 // for all tasks — used by listTasks so the task list shows progress for every task.
 func (d *DB) GoalCountsAll() (map[int64]GoalCounts, error) {

db/task_list_performance_test.go

@@ -117,3 +117,133 @@ VALUES ($1,'goal','{}','met'), ($1,'goal','{}','open')`, task.ExplorationID); er
 		t.Fatalf("unexpected goal metrics: %+v", metrics.Goals)
 	}
 }
+
+// TaskListMetrics(expID) is exactly the row TaskListMetricsAll returns for that
+// exploration, on every counter, and TaskListMetricsAll still lists every task.
+func TestTaskListMetricsMatchesAll(t *testing.T) {
+	d, err := Open(testDSN(t))
+	if err != nil {
+		t.Skipf("postgres unavailable (%v) - skipping", err)
+	}
+	t.Cleanup(func() { d.Close() })
+
+	newest := time.Now().Add(-time.Minute).Truncate(time.Second)
+	seed := func(name string, n int) *Task {
+		t.Helper()
+		task, err := d.CreateTask(name, "goal", nil, 0, 0)
+		if err != nil {
+			t.Fatal(err)
+		}
+		t.Cleanup(func() {
+			_, err := d.DeleteTaskCascade(task.ID, false, true)
+			cleanupErr(t, "task "+name, err)
+		})
+		exp := task.ExplorationID
+		// n result rows with tokens; the newest activity first, then an older row
+		// with a higher id, so ordering by id instead of created_at would fail.
+		for _, q := range []struct {
+			sql  string
+			args []any
+		}{
+			{`INSERT INTO activity(exploration_id,kind,input_tokens,output_tokens,cache_read_tokens,cache_write_tokens,created_at)
+			 SELECT $1,'result',10,5,3,1,$3::timestamptz - interval '1 hour' FROM generate_series(1,$2::int)`, []any{exp, n, newest}},
+			{`INSERT INTO activity(exploration_id,kind,created_at) VALUES ($1,'text',$2::timestamptz)`, []any{exp, newest}},
+			{`INSERT INTO activity(exploration_id,kind,created_at) SELECT $1,'text',$3::timestamptz - interval '2 hours' FROM generate_series(1,$2::int)`, []any{exp, n, newest}},
+			{`INSERT INTO exploration_nodes(exploration_id,kind,payload,state)
+			 SELECT $1,'goal','{}',CASE WHEN g<=$2::int THEN 'met' ELSE 'open' END FROM generate_series(1,$2::int+1) g`, []any{exp, n}},
+			{`INSERT INTO exploration_nodes(exploration_id,kind,payload,state)
+			 SELECT $1,'intent','{}',CASE WHEN g<=$2::int THEN 'running' ELSE 'done' END FROM generate_series(1,$2::int+2) g`, []any{exp, n}},
+		} {
+			if _, err := d.Exec(q.sql, q.args...); err != nil {
+				t.Fatal(err)
+			}
+		}
+		if _, err := d.Exec(`INSERT INTO findings(task_id,severity)
+SELECT $1, s FROM unnest(ARRAY['critical','high','high','medium','low','info']) s, generate_series(1,$2)`, task.ID, n); err != nil {
+			t.Fatal(err)
+		}
+		return task
+	}
+	a := seed("task-list-metrics-a", 2)
+	b := seed("task-list-metrics-b", 3)
+
+	all, err := d.TaskListMetricsAll()
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, c := range []struct {
+		task *Task
+		n    int
+	}{{a, 2}, {b, 3}} {
+		row, ok := all[c.task.ExplorationID]
+		if !ok {
+			t.Fatalf("TaskListMetricsAll misses task %d", c.task.ID)
+		}
+		one, err := d.TaskListMetrics(c.task.ExplorationID)
+		if err != nil {
+			t.Fatal(err)
+		}
+		if one != row {
+			t.Fatalf("task %d: TaskListMetrics=%+v, TaskListMetricsAll=%+v", c.task.ID, one, row)
+		}
+		n := c.n
+		want := TaskListMetrics{
+			Tokens:         TokenUsage{InputTokens: 10 * n, OutputTokens: 5 * n, CacheReadTokens: 3 * n, CacheWriteTokens: n},
+			LastActivity:   newest.Unix(),
+			Goals:          GoalCounts{Total: n + 1, Met: n},
+			RunningIntents: n,
+			Findings:       FindingSeverityCounts{Critical: n, High: 2 * n, Medium: n, Low: n},
+		}
+		if one != want {
+			t.Fatalf("task %d metrics=%+v, want %+v", c.task.ID, one, want)
+		}
+	}
+	if none, err := d.TaskListMetrics(-1); err != nil || none != (TaskListMetrics{}) {
+		t.Fatalf("missing exploration: %+v, %v; want zero metrics", none, err)
+	}
+}
+
+// BenchmarkTaskListMetrics compares the per-exploration metrics query with the
+// all-tasks one on three tasks of 40,000 activity rows each. No time assertion.
+func BenchmarkTaskListMetrics(b *testing.B) {
+	dsn, _, err := DSN()
+	if err != nil {
+		b.Skipf("no database config (%v) - skipping", err)
+	}
+	d, err := Open(dsn)
+	if err != nil {
+		b.Skipf("postgres unavailable (%v) - skipping", err)
+	}
+	b.Cleanup(func() { d.Close() })
+	var exp int64
+	for i := range 3 {
+		task, err := d.CreateTask(fmt.Sprintf("bench-task-list-metrics-%d", i), "goal", nil, 0, 0)
+		if err != nil {
+			b.Fatal(err)
+		}
+		b.Cleanup(func() { _, _ = d.DeleteTaskCascade(task.ID, false, true) })
+		if _, err := d.Exec(`INSERT INTO activity(exploration_id,kind,input_tokens,created_at)
+SELECT $1, CASE WHEN g%20=0 THEN 'result' ELSE 'tool_use' END, 100, now() - (g||' seconds')::interval
+FROM generate_series(1,40000) g`, task.ExplorationID); err != nil {
+			b.Fatal(err)
+		}
+		exp = task.ExplorationID
+	}
+	if _, err := d.Exec(`ANALYZE activity`); err != nil {
+		b.Fatal(err)
+	}
+	b.Run("one", func(b *testing.B) {
+		for range b.N {
+			if _, err := d.TaskListMetrics(exp); err != nil {
+				b.Fatal(err)
+			}
+		}
+	})
+	b.Run("all", func(b *testing.B) {
+		for range b.N {
+			if _, err := d.TaskListMetricsAll(); err != nil {
+				b.Fatal(err)
+			}
+		}
+	})
+}

db/task_scope_test.go

@@ -6,14 +6,14 @@ import "testing"
 // ip/cidr scope path so a target like "10.0.188.136:3000" no longer 404s.
 func TestStripHostPort(t *testing.T) {
 	cases := []struct{ in, want string }{
-		{"10.0.188.136:3000", "10.0.188.136"},          // the reported IP case
-		{"10.0.188.136", "10.0.188.136"},               // bare IPv4 unchanged
-		{"[2001:db8::1]:8080", "2001:db8::1"},          // bracketed IPv6 + port
-		{"2001:db8::1", "2001:db8::1"},                 // bare IPv6 unchanged (has colons)
-		{" 1.2.3.4:80 ", "1.2.3.4"},                    // trims surrounding space
-		{"example.com:443", "example.com"},             // domain + port → bare host
-		{"api.example.com:8080", "api.example.com"},    // subdomain + port
-		{"example.com", "example.com"},                 // bare domain unchanged
+		{"10.0.188.136:3000", "10.0.188.136"},       // the reported IP case
+		{"10.0.188.136", "10.0.188.136"},            // bare IPv4 unchanged
+		{"[2001:db8::1]:8080", "2001:db8::1"},       // bracketed IPv6 + port
+		{"2001:db8::1", "2001:db8::1"},              // bare IPv6 unchanged (has colons)
+		{" 1.2.3.4:80 ", "1.2.3.4"},                 // trims surrounding space
+		{"example.com:443", "example.com"},          // domain + port → bare host
+		{"api.example.com:8080", "api.example.com"}, // subdomain + port
+		{"example.com", "example.com"},              // bare domain unchanged
 	}
 	for _, c := range cases {
 		if got := stripHostPort(c.in); got != c.want {

db/tasks.go

@@ -28,19 +28,19 @@ type Task struct {
 	LLMProfileID  *int64     `json:"llm_profile_id,omitempty"`
 	// Task-level ordered LLM chain. LLMProfileID remains the compatibility alias
 	// for ActiveLLMProfileID while older API clients still send one profile id.
-	LLMProfileIDs      []int64    `json:"llm_profile_ids,omitempty"`
-	ActiveLLMProfileID *int64     `json:"active_llm_profile_id,omitempty"`
-	LLMChainRevision   int64      `json:"-"`
-	LLMFailoverState   string     `json:"llm_failover_state,omitempty"`
-	LLMFailoverReason  string     `json:"llm_failover_reason,omitempty"`
+	LLMProfileIDs      []int64 `json:"llm_profile_ids,omitempty"`
+	ActiveLLMProfileID *int64  `json:"active_llm_profile_id,omitempty"`
+	LLMChainRevision   int64   `json:"-"`
+	LLMFailoverState   string  `json:"llm_failover_state,omitempty"`
+	LLMFailoverReason  string  `json:"llm_failover_reason,omitempty"`
 	// LLMRotate: per-intent chain-head rotation switch (#85); false = classic
 	// ordered failover from the active head.
-	LLMRotate bool `json:"llm_rotate"`
-	SourceTaskIDs      []int64    `json:"source_task_ids,omitempty"`
-	CompanyIDs         []int64    `json:"company_ids,omitempty"`
-	ParentRef          string     `json:"parent_ref,omitempty"` // parent task id (recorded by orchestration spawn; empty=top level)
-	CreatedAt          time.Time  `json:"created_at"`
-	CompletedAt        *time.Time `json:"completed_at,omitempty"` // when it entered a terminal state (done/failed/timeout); nil for non-terminal
+	LLMRotate     bool       `json:"llm_rotate"`
+	SourceTaskIDs []int64    `json:"source_task_ids,omitempty"`
+	CompanyIDs    []int64    `json:"company_ids,omitempty"`
+	ParentRef     string     `json:"parent_ref,omitempty"` // parent task id (recorded by orchestration spawn; empty=top level)
+	CreatedAt     time.Time  `json:"created_at"`
+	CompletedAt   *time.Time `json:"completed_at,omitempty"` // when it entered a terminal state (done/failed/timeout); nil for non-terminal
 	// Task-level timeout (see the task-level timeout and wrap-up design doc).
 	TimeoutSeconds int        `json:"timeout_seconds"`        // 0=no time limit
 	FirstRunAt     *time.Time `json:"first_run_at,omitempty"` // when it first actually started running (not created_at); nil=not yet run

server/assembly_test.go

@@ -6,9 +6,9 @@ import (
 	"path/filepath"
 	"testing"
 
-	actool "github.com/Autumn-27/norma/tool"
 	"github.com/Autumn-27/artex/agent"
 	"github.com/Autumn-27/artex/db"
+	actool "github.com/Autumn-27/norma/tool"
 )
 
 // hasSkillTool reports whether the packed tool set contains the Skill meta-tool.

server/goal_gate_test.go

@@ -7,19 +7,19 @@ import (
 
 func TestGoalGateCheck(t *testing.T) {
 	good := map[string]string{
-		"url target":            "Audit " + strings.Repeat("x ", 60) + " at https://example.com/app every module for injection issues and record one finding per issue with evidence.",
-		"bare domain":           strings.Repeat("probe ", 25) + "the api at target.example.org for unauthenticated data exposure, one record proof, then stop.",
-		"ipv4":                  strings.Repeat("recon ", 25) + "10.0.0.5 and map every listening service before any test.",
-		"localhost port":        strings.Repeat("verify ", 25) + "localhost:3000 health endpoints respond identically before and after config reload.",
-		"path":                  strings.Repeat("review ", 25) + "the module at /app/data/core for hardcoded secrets and report each hit.",
-		"targetless opt-out":    "[targetless] " + strings.Repeat("summarize ", 20) + "the attached material into a prioritized hypothesis list.",
+		"url target":         "Audit " + strings.Repeat("x ", 60) + " at https://example.com/app every module for injection issues and record one finding per issue with evidence.",
+		"bare domain":        strings.Repeat("probe ", 25) + "the api at target.example.org for unauthenticated data exposure, one record proof, then stop.",
+		"ipv4":               strings.Repeat("recon ", 25) + "10.0.0.5 and map every listening service before any test.",
+		"localhost port":     strings.Repeat("verify ", 25) + "localhost:3000 health endpoints respond identically before and after config reload.",
+		"path":               strings.Repeat("review ", 25) + "the module at /app/data/core for hardcoded secrets and report each hit.",
+		"targetless opt-out": "[targetless] " + strings.Repeat("summarize ", 20) + "the attached material into a prioritized hypothesis list.",
 	}
 	good["targetless short but marked"] = "[targetless] analyze nothing"
 
 	bad := map[string]string{
-		"no target long":  strings.Repeat("do a thorough security review of everything and report anything interesting that comes up during the exploration phase of this audit, thanks. ", 2),
-		"too short":       "audit the site",
-		"short and bare":  "find bugs",
+		"no target long": strings.Repeat("do a thorough security review of everything and report anything interesting that comes up during the exploration phase of this audit, thanks. ", 2),
+		"too short":      "audit the site",
+		"short and bare": "find bugs",
 	}
 	for name, g := range good {
 		if name == "targetless short+:url" {

server/manager.go

@@ -55,13 +55,13 @@ type Task struct {
 	LLMChainRevision   int64   `json:"-"`
 	LLMFailoverState   string  `json:"llm_failover_state,omitempty"`
 	// LLMRotate (#85): per-intent round-robin over the task's LLM chain heads.
-	LLMRotate  bool `json:"llm_rotate,omitempty"`
-	rotateMu   sync.Mutex
-	rotateNext int
-	LLMFailoverReason  string  `json:"llm_failover_reason,omitempty"`
-	SourceTaskIDs      []int64 `json:"source_task_ids,omitempty"`
-	CompanyIDs         []int64 `json:"company_ids,omitempty"`
-	Status             string  `json:"status"` // persisted lifecycle status (done/failed/timeout are terminal; empty/other is derived from runtime state)
+	LLMRotate         bool `json:"llm_rotate,omitempty"`
+	rotateMu          sync.Mutex
+	rotateNext        int
+	LLMFailoverReason string  `json:"llm_failover_reason,omitempty"`
+	SourceTaskIDs     []int64 `json:"source_task_ids,omitempty"`
+	CompanyIDs        []int64 `json:"company_ids,omitempty"`
+	Status            string  `json:"status"` // persisted lifecycle status (done/failed/timeout are terminal; empty/other is derived from runtime state)
 	// Task-level timeout (see docs/task-timeout-and-settle-design.md). DeadlineAt/FirstRunAt are unix seconds, 0 = unset/not yet run.
 	TimeoutSeconds       int                    `json:"timeout_seconds"`
 	PlanHeartbeatSeconds int                    `json:"plan_heartbeat_seconds"` // planner heartbeat trigger interval (seconds)

server/server.go

@@ -1609,12 +1609,12 @@ type createTaskReq struct {
 	CategoryID           *int64   `json:"category_id,omitempty"`
 	Description          string   `json:"description"`
 	Goal                 string   `json:"goal"`
-	LLMProfileID         *int64   `json:"llm_profile_id,omitempty"`    // the LLM profile to run this task on; omitted/null = use the active profile
-	LLMProfileIDs        []int64  `json:"llm_profile_ids,omitempty"`   // ordered task-level profile chain; the first entry takes effect initially
-	SourceTaskIDs        []string `json:"source_task_ids,omitempty"`   // direct, read-only inherited source tasks
-	CompanyIDs           []int64  `json:"company_ids,omitempty"`       // links the company scope and snapshots its current assets; does not copy assets or force intent generation
+	LLMProfileID         *int64   `json:"llm_profile_id,omitempty"`  // the LLM profile to run this task on; omitted/null = use the active profile
+	LLMProfileIDs        []int64  `json:"llm_profile_ids,omitempty"` // ordered task-level profile chain; the first entry takes effect initially
+	SourceTaskIDs        []string `json:"source_task_ids,omitempty"` // direct, read-only inherited source tasks
+	CompanyIDs           []int64  `json:"company_ids,omitempty"`     // links the company scope and snapshots its current assets; does not copy assets or force intent generation
 	TimeoutSeconds       int      `json:"timeout_seconds"`
-	LLMRotate            *bool    `json:"llm_rotate"`             // task-level timeout (seconds); 0/omitted = no limit
+	LLMRotate            *bool    `json:"llm_rotate"`                  // task-level timeout (seconds); 0/omitted = no limit
 	PlanHeartbeatSeconds int      `json:"plan_heartbeat_seconds"`      // planner heartbeat interval (seconds); 0/omitted = default 600 (10min); floor = default = 600, lower values are raised to 600
 	PlanQuietMode        *bool    `json:"plan_quiet_mode,omitempty"`   // planner quiet-mode switch (#86); omitted = default on; false = edge rounds always plan
 	TargetGate           *bool    `json:"target_gate,omitempty"`       // approval gate for asset targets, opt-in; omitted/null = off
@@ -1927,11 +1927,11 @@ func (s *Server) getTask(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 	dto := taskDTO(t, s.resolvedTaskStatus(t))
-	metrics, err := s.m.PG().TaskListMetricsAll()
+	metrics, err := s.m.PG().TaskListMetrics(t.ExpID)
 	if err != nil {
 		log.Printf("[tasks] detail metrics query failed for %s — metrics zeroed: %v", t.ID, err)
 	}
-	s.applyTaskMetrics(&dto, t, metrics[t.ExpID])
+	s.applyTaskMetrics(&dto, t, metrics)
 	archiveBlockers, _ := s.m.PG().TaskArchiveBlockers()
 	applyTaskArchiveBlocker(&dto, archiveBlockers)
 	writeJSON(w, 200, dto)

server/task_detail_metrics_test.go

@@ -38,6 +38,21 @@ VALUES ($1,'goal','{}','met'), ($1,'goal','{}','open')`, task.ExpID); err != nil
 		t.Fatal(err)
 	}
 
+	// A second task, so the detail must pick its own row and the list must keep
+	// serving every task.
+	other, err := m.CreateTask("detail metrics other", "goal", nil, 0, 0)
+	if err != nil {
+		t.Fatal(err)
+	}
+	otherID, _ := strconv.ParseInt(other.ID, 10, 64)
+	defer func() { _, _ = m.pg.DeleteTaskCascade(otherID, false, true) }()
+	defer func() { _, _ = m.pg.Exec(`DELETE FROM findings WHERE task_id=$1`, taskID) }()
+	if _, err := m.pg.Exec(`
+INSERT INTO findings(task_id, severity) VALUES ($1,'critical'), ($1,'high'), ($1,'high'), ($1,'low'), ($2,'medium')`,
+		taskID, otherID); err != nil {
+		t.Fatal(err)
+	}
+
 	ctx, cancel := context.WithCancel(context.Background())
 	s := New(ctx, m, t.TempDir(), t.TempDir(), t.TempDir())
 	defer stopTestServer(t, s, cancel)
@@ -45,7 +60,19 @@ VALUES ($1,'goal','{}','met'), ($1,'goal','{}','open')`, task.ExpID); err != nil
 	if err != nil {
 		t.Fatal(err)
 	}
+	// After New: restoring the task resets running intents to open.
+	if _, err := m.pg.Exec(`
+INSERT INTO exploration_nodes(exploration_id, kind, payload, state)
+VALUES ($1,'intent','{}','running'), ($1,'intent','{}','done')`, task.ExpID); err != nil {
+		t.Fatal(err)
+	}
 
+	type findingCounts struct {
+		Critical int `json:"critical"`
+		High     int `json:"high"`
+		Medium   int `json:"medium"`
+		Low      int `json:"low"`
+	}
 	var probe struct {
 		Tokens struct {
 			InputTokens      int `json:"input_tokens"`
@@ -53,8 +80,10 @@ VALUES ($1,'goal','{}','met'), ($1,'goal','{}','open')`, task.ExpID); err != nil
 			CacheReadTokens  int `json:"cache_read_tokens"`
 			CacheWriteTokens int `json:"cache_write_tokens"`
 		} `json:"tokens"`
-		GoalsMet   int `json:"goals_met"`
-		GoalsTotal int `json:"goals_total"`
+		GoalsMet   int           `json:"goals_met"`
+		GoalsTotal int           `json:"goals_total"`
+		InFlight   int           `json:"in_flight"`
+		Findings   findingCounts `json:"findings"`
 	}
 	req := httptest.NewRequest(http.MethodGet, "/api/tasks/"+task.ID, nil)
 	req.Header.Set("Authorization", "Bearer "+token)
@@ -73,6 +102,9 @@ VALUES ($1,'goal','{}','met'), ($1,'goal','{}','open')`, task.ExpID); err != nil
 	if probe.GoalsMet != 1 || probe.GoalsTotal != 2 {
 		t.Fatalf("detail goal progress not filled from metrics: %d/%d", probe.GoalsMet, probe.GoalsTotal)
 	}
+	if probe.InFlight != 1 || probe.Findings != (findingCounts{Critical: 1, High: 2, Low: 1}) {
+		t.Fatalf("detail in-flight/findings not filled from this task's metrics: %d %+v", probe.InFlight, probe.Findings)
+	}
 
 	// The list endpoint must report the same numbers for the same task.
 	listReq := httptest.NewRequest(http.MethodGet, "/api/tasks", nil)
@@ -89,27 +121,32 @@ VALUES ($1,'goal','{}','met'), ($1,'goal','{}','open')`, task.ExpID); err != nil
 				InputTokens  int `json:"input_tokens"`
 				OutputTokens int `json:"output_tokens"`
 			} `json:"tokens"`
-			GoalsMet   int `json:"goals_met"`
-			GoalsTotal int `json:"goals_total"`
+			GoalsMet   int           `json:"goals_met"`
+			GoalsTotal int           `json:"goals_total"`
+			InFlight   int           `json:"in_flight"`
+			Findings   findingCounts `json:"findings"`
 		} `json:"tasks"`
 	}
 	if err := json.Unmarshal(listRec.Body.Bytes(), &listed); err != nil {
 		t.Fatal(err)
 	}
-	found := false
+	found, foundOther := false, false
 	for _, lt := range listed.Tasks {
+		if lt.ID == other.ID {
+			foundOther = lt.Findings == (findingCounts{Medium: 1})
+		}
 		if lt.ID != task.ID {
 			continue
 		}
 		found = true
 		if lt.Tokens.InputTokens != probe.Tokens.InputTokens ||
 			lt.Tokens.OutputTokens != probe.Tokens.OutputTokens ||
-			lt.GoalsMet != probe.GoalsMet || lt.GoalsTotal != probe.GoalsTotal {
-			t.Fatalf("list and detail metrics disagree: list=%+v/%d detail=%+v/%d",
-				lt.Tokens, lt.GoalsMet, probe.Tokens, probe.GoalsTotal)
+			lt.GoalsMet != probe.GoalsMet || lt.GoalsTotal != probe.GoalsTotal ||
+			lt.InFlight != probe.InFlight || lt.Findings != probe.Findings {
+			t.Fatalf("list and detail metrics disagree: list=%+v detail=%+v", lt, probe)
 		}
 	}
-	if !found {
-		t.Fatalf("task %s missing from list response", task.ID)
+	if !found || !foundOther {
+		t.Fatalf("list misses a task or its metrics: task %s found=%v, task %s with its findings=%v", task.ID, found, other.ID, foundOther)
 	}
 }

server/task_llm.go

@@ -317,10 +317,14 @@ func (r *taskLLMRuntime) Complete(ctx context.Context, req llm.CompletionRequest
 }
 
 func completeTaskLLM(ctx context.Context, taskID string, req llm.CompletionRequest, hooks taskLLMStreamHooks) (llm.Message, string, llm.Usage, error) {
+	// total is the usage of the logical call: the sum of every physical attempt
+	// (same-provider retries and profile failover). Each attempt is already billed
+	// once by the recorder below; this sum only feeds the caller's accounting.
+	var total llm.Usage
 	for {
 		selection, err := hooks.current(ctx)
 		if err != nil {
-			return llm.Message{}, "", llm.Usage{}, err
+			return llm.Message{}, "", total, err
 		}
 		var (
 			msg     llm.Message
@@ -333,13 +337,14 @@ func completeTaskLLM(ctx context.Context, taskID string, req llm.CompletionReque
 		retries, backoffOf := sameProviderRetryPolicy(selection.retry)
 		for attempt := 0; ; attempt++ {
 			msg, sr, usage, callErr = selection.provider.Complete(ctx, req)
+			total.Add(usage)
 			if callErr != nil && ctx.Err() == nil &&
 				attempt < retries && isRetryableStreamError(callErr) {
 				backoff := backoffOf(attempt)
 				log.Printf("[task-llm] task %s non-streaming call failed, retrying on the same provider after %v (%d/%d): %v",
 					taskID, backoff, attempt+1, retries, callErr)
 				if sleepCtx(ctx, backoff) {
-					break // ctx cancelled during backoff → stop retrying
+					return llm.Message{}, "", total, ctx.Err() // cancelled during backoff → stop
 				}
 				continue
 			}
@@ -347,7 +352,7 @@ func completeTaskLLM(ctx context.Context, taskID string, req llm.CompletionReque
 		}
 		if callErr == nil {
 			resetLLMFailStreak(taskID, selection.profileID)
-			return msg, sr, usage, nil
+			return msg, sr, total, nil
 		}
 		// Non-quota failures on a chain profile escalate after a consecutive
 		// streak (#99): same advance path as quota exhaust, truthful reason.
@@ -356,31 +361,31 @@ func completeTaskLLM(ctx context.Context, taskID string, req llm.CompletionReque
 				cause := fmt.Errorf("profile #%d failed %d consecutive calls (last: %w)", selection.profileID, llmFailStreakThreshold, callErr)
 				transition, markErr := hooks.exhaust(selection, cause)
 				if markErr != nil {
-					return llm.Message{}, "", llm.Usage{}, fmt.Errorf("mark profile exhausted after %v: %w", callErr, markErr)
+					return llm.Message{}, "", total, fmt.Errorf("mark profile exhausted after %v: %w", callErr, markErr)
 				}
 				if transition.Advanced && !transition.Stale && hooks.transition != nil {
 					hooks.transition(selection, transition, callErr)
 				}
 				if !transition.Stale && transition.NextProfileID == nil {
-					return llm.Message{}, "", llm.Usage{}, &taskLLMError{taskID: taskID, chainExhausted: transition.ChainExhausted, cause: callErr}
+					return llm.Message{}, "", total, &taskLLMError{taskID: taskID, chainExhausted: transition.ChainExhausted, cause: callErr}
 				}
 				continue // replay the same logical request on the next profile
 			}
-			return llm.Message{}, "", llm.Usage{}, callErr
+			return llm.Message{}, "", total, callErr
 		}
 		// profileID=0: the explicit chain was cleared; non-quota error: pass through. Neither changes the task failover state.
 		if selection.profileID == 0 || !isQuotaExhaustedError(callErr) {
-			return llm.Message{}, "", llm.Usage{}, callErr
+			return llm.Message{}, "", total, callErr
 		}
 		transition, markErr := hooks.exhaust(selection, callErr)
 		if markErr != nil {
-			return llm.Message{}, "", llm.Usage{}, fmt.Errorf("mark profile quota exhausted after %v: %w", callErr, markErr)
+			return llm.Message{}, "", total, fmt.Errorf("mark profile quota exhausted after %v: %w", callErr, markErr)
 		}
 		if transition.Advanced && !transition.Stale && hooks.transition != nil {
 			hooks.transition(selection, transition, callErr)
 		}
 		if !transition.Stale && transition.NextProfileID == nil {
-			return llm.Message{}, "", llm.Usage{}, &taskLLMError{taskID: taskID, chainExhausted: transition.ChainExhausted, cause: callErr}
+			return llm.Message{}, "", total, &taskLLMError{taskID: taskID, chainExhausted: transition.ChainExhausted, cause: callErr}
 		}
 		// No output has been delivered to the caller, so replaying the same logical request on the next profile is safe.
 	}
@@ -397,6 +402,35 @@ func streamTaskLLM(ctx context.Context, taskID string, req llm.CompletionRequest
 			committed := false
 			var pending []llm.StreamEvent
 			var streamErr error
+			// The caller's llm.Accumulator sums usage across the whole logical call
+			// (every attempt and profile): message_start is added, message_delta is
+			// folded with MergeEcho. Folding one attempt's echo against the sum of
+			// earlier attempts would lose input/cache tokens, so each attempt's echoes
+			// are normalized here against that attempt alone and only the increment
+			// is passed on, as a message_start. The recorder below already billed the
+			// physical attempt; this only shapes what the caller adds up.
+			var attemptUsage llm.Usage
+			forward := func(event llm.StreamEvent) bool {
+				switch event.Type {
+				case llm.SEMessageStart:
+					attemptUsage.Add(event.Usage)
+				case llm.SEMessageDelta:
+					next := attemptUsage
+					next.MergeEcho(event.Usage)
+					inc := llm.Usage{
+						InputTokens:      next.InputTokens - attemptUsage.InputTokens,
+						OutputTokens:     next.OutputTokens - attemptUsage.OutputTokens,
+						CacheReadTokens:  next.CacheReadTokens - attemptUsage.CacheReadTokens,
+						CacheWriteTokens: next.CacheWriteTokens - attemptUsage.CacheWriteTokens,
+					}
+					attemptUsage = next
+					if inc != (llm.Usage{}) && !yield(llm.StreamEvent{Type: llm.SEMessageStart, Usage: inc}, nil) {
+						return false
+					}
+					event.Usage = llm.Usage{}
+				}
+				return yield(event, nil)
+			}
 			retries, backoffOf := sameProviderRetryPolicy(selection.retry)
 			// Same-provider safe-window retry: a transient failure before committed
 			// (nothing delivered to the caller yet) can be replayed verbatim without
@@ -405,6 +439,7 @@ func streamTaskLLM(ctx context.Context, taskID string, req llm.CompletionRequest
 				committed = false
 				pending = nil
 				streamErr = nil
+				attemptUsage = llm.Usage{}
 				for event, err := range selection.provider.Stream(ctx, req) {
 					if err != nil {
 						streamErr = err
@@ -416,24 +451,34 @@ func streamTaskLLM(ctx context.Context, taskID string, req llm.CompletionRequest
 					}
 					if !committed {
 						for _, buffered := range pending {
-							if !yield(buffered, nil) {
+							if !forward(buffered) {
 								return
 							}
 						}
 						pending = nil
 						committed = true
 					}
-					if !yield(event, nil) {
+					if !forward(event) {
 						return
 					}
 				}
+				if streamErr != nil && !committed {
+					// A failed uncommitted attempt delivers none of its events (they are
+					// never replayed), but the usage it reported survives as one
+					// message_start so the logical call still counts it.
+					if u := pendingUsage(pending); u != (llm.Usage{}) && !yield(llm.StreamEvent{Type: llm.SEMessageStart, Usage: u}, nil) {
+						return
+					}
+					pending = nil
+				}
 				if streamErr != nil && !committed && ctx.Err() == nil &&
 					attempt < retries && isRetryableStreamError(streamErr) {
 					backoff := backoffOf(attempt)
 					log.Printf("[task-llm] task %s pre-commit stream failure, retrying on the same provider after %v (%d/%d): %v",
 						taskID, backoff, attempt+1, retries, streamErr)
 					if sleepCtx(ctx, backoff) {
-						break // ctx cancelled during backoff → stop retrying
+						yield(llm.StreamEvent{}, ctx.Err()) // cancelled during backoff → stop
+						return
 					}
 					continue
 				}
@@ -442,7 +487,7 @@ func streamTaskLLM(ctx context.Context, taskID string, req llm.CompletionRequest
 			if streamErr == nil {
 				resetLLMFailStreak(taskID, selection.profileID)
 				for _, buffered := range pending {
-					if !yield(buffered, nil) {
+					if !forward(buffered) {
 						return
 					}
 				}
@@ -474,11 +519,6 @@ func streamTaskLLM(ctx context.Context, taskID string, req llm.CompletionRequest
 			// bundle was still in use. Agent/global fallback errors follow the legacy
 			// behavior and never mutate task failover state.
 			if selection.profileID == 0 || !isQuotaExhaustedError(streamErr) {
-				for _, buffered := range pending {
-					if !yield(buffered, nil) {
-						return
-					}
-				}
 				yield(llm.StreamEvent{}, streamErr)
 				return
 			}
@@ -508,9 +548,25 @@ func streamTaskLLM(ctx context.Context, taskID string, req llm.CompletionRequest
 	}
 }
 
+// pendingUsage is the usage the buffered (uncommitted) events of one attempt
+// reported. Only message_start can be buffered: message_delta commits.
+func pendingUsage(pending []llm.StreamEvent) llm.Usage {
+	var u llm.Usage
+	for _, ev := range pending {
+		if ev.Type == llm.SEMessageStart {
+			u.Add(ev.Usage)
+		}
+	}
+	return u
+}
+
+// streamEventCommitsOutput reports whether delivering event makes the attempt
+// unreplayable. Signed and redacted thinking and tool-use signatures are opaque
+// output the model bound to this exact response, so they commit like text.
 func streamEventCommitsOutput(event llm.StreamEvent) bool {
 	switch event.Type {
-	case llm.SETextDelta, llm.SEThinkingDelta, llm.SEToolInputJSON:
+	case llm.SETextDelta, llm.SEThinkingDelta, llm.SEToolInputJSON,
+		llm.SEThinkingSignature, llm.SERedactedThinking, llm.SEToolUseSignature:
 		return event.Text != ""
 	case llm.SEToolUseStart, llm.SEMessageDelta, llm.SEMessageStop:
 		return true

server/task_llm_escalation_test.go

@@ -77,4 +77,3 @@ func TestEscalationAdvancesAfterConsecutiveFailures(t *testing.T) {
 		t.Fatalf("exhaust calls=%d want 2", exhaustCalls)
 	}
 }
-

server/task_llm_usage_test.go

@@ -0,0 +1,286 @@
+package server
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"iter"
+	"testing"
+	"time"
+
+	"github.com/Autumn-27/artex/agent"
+	"github.com/Autumn-27/artex/db"
+	"github.com/Autumn-27/artex/llmrec"
+	"github.com/Autumn-27/norma/llm"
+)
+
+// Usage reported by one failed attempt, and by a successful one. A stream
+// reports the input side in message_start and echoes it, with the output, in
+// message_delta; the atomic path returns the attempt's normalized usage.
+var (
+	failedAttemptUsage = llm.Usage{InputTokens: 11, OutputTokens: 2, CacheReadTokens: 3, CacheWriteTokens: 2}
+	okStartUsage       = llm.Usage{InputTokens: 13, CacheReadTokens: 3, CacheWriteTokens: 2}
+	okDeltaUsage       = llm.Usage{InputTokens: 17, OutputTokens: 5, CacheReadTokens: 4, CacheWriteTokens: 2}
+	okAttemptUsage     = llm.Usage{InputTokens: 17, OutputTokens: 5, CacheReadTokens: 4, CacheWriteTokens: 2}
+)
+
+// usageAttemptProvider fails its first fail attempts (always, when fail < 0)
+// after reporting failedAttemptUsage, then answers "done".
+type usageAttemptProvider struct {
+	fail    int
+	failErr error
+	onFail  func()
+	// extra events a failed stream attempt emits before its error.
+	failEvents []llm.StreamEvent
+	calls      int
+}
+
+func (p *usageAttemptProvider) failing() bool {
+	p.calls++
+	failed := p.fail < 0 || p.calls <= p.fail
+	if failed && p.onFail != nil {
+		p.onFail()
+	}
+	return failed
+}
+
+func (p *usageAttemptProvider) Complete(context.Context, llm.CompletionRequest) (llm.Message, string, llm.Usage, error) {
+	if p.failing() {
+		return llm.Message{}, "", failedAttemptUsage, p.failErr
+	}
+	return llm.Message{Role: llm.RoleAssistant, Content: []llm.ContentBlock{llm.TextBlock("done")}}, "end_turn", okAttemptUsage, nil
+}
+
+func (p *usageAttemptProvider) Stream(context.Context, llm.CompletionRequest) iter.Seq2[llm.StreamEvent, error] {
+	return func(yield func(llm.StreamEvent, error) bool) {
+		if p.failing() {
+			events := append([]llm.StreamEvent{{Type: llm.SEMessageStart, Usage: failedAttemptUsage}}, p.failEvents...)
+			for _, ev := range events {
+				if !yield(ev, nil) {
+					return
+				}
+			}
+			yield(llm.StreamEvent{}, p.failErr)
+			return
+		}
+		for _, ev := range []llm.StreamEvent{
+			{Type: llm.SEMessageStart, Usage: okStartUsage},
+			{Type: llm.SETextDelta, Text: "done"},
+			{Type: llm.SEMessageDelta, StopReason: "end_turn", Usage: okDeltaUsage},
+			{Type: llm.SEMessageStop},
+		} {
+			if !yield(ev, nil) {
+				return
+			}
+		}
+	}
+}
+
+// logicalResult is what the SDK sees for one logical call.
+type logicalResult struct {
+	usage  llm.Usage
+	text   string
+	events []llm.StreamEvent
+	err    error
+}
+
+func runLogicalCall(ctx context.Context, stream bool, taskID string, hooks taskLLMStreamHooks) logicalResult {
+	if !stream {
+		msg, _, usage, err := completeTaskLLM(ctx, taskID, llm.CompletionRequest{}, hooks)
+		return logicalResult{usage: usage, text: msg.Text(), err: err}
+	}
+	var r logicalResult
+	acc := llm.NewAccumulator()
+	for ev, err := range streamTaskLLM(ctx, taskID, llm.CompletionRequest{}, hooks) {
+		if err != nil {
+			r.err = err
+			break
+		}
+		r.events = append(r.events, ev)
+		acc.Add(ev)
+	}
+	r.usage, r.text = acc.Usage, acc.Message().Text()
+	return r
+}
+
+func singleProfileHooks(p llm.Provider, retry agent.RetryConfig) taskLLMStreamHooks {
+	return taskLLMStreamHooks{
+		current: func(context.Context) (taskLLMSelection, error) {
+			return taskLLMSelection{provider: p, retry: retry}, nil
+		},
+		exhaust: func(taskLLMSelection, error) (db.TaskLLMTransition, error) {
+			return db.TaskLLMTransition{}, nil
+		},
+	}
+}
+
+func sumUsage(us ...llm.Usage) llm.Usage {
+	var t llm.Usage
+	for _, u := range us {
+		t.Add(u)
+	}
+	return t
+}
+
+var usageModes = []struct {
+	name   string
+	stream bool
+}{{"atomic", false}, {"stream", true}}
+
+func TestTaskLLMRetrySumsAttemptUsage(t *testing.T) {
+	for _, m := range usageModes {
+		t.Run(m.name, func(t *testing.T) {
+			p := &usageAttemptProvider{fail: 1, failErr: errors.New("connection reset by peer")}
+			r := runLogicalCall(t.Context(), m.stream, "p12-retry", singleProfileHooks(p, agent.RetryConfig{StreamAttempts: 1, StreamInterval: time.Millisecond}))
+			if r.err != nil {
+				t.Fatal(r.err)
+			}
+			want := sumUsage(failedAttemptUsage, okAttemptUsage) // {28 7 7 4}
+			if r.usage != want || p.calls != 2 || r.text != "done" {
+				t.Fatalf("usage=%+v want %+v, calls=%d, text=%q", r.usage, want, p.calls, r.text)
+			}
+		})
+	}
+}
+
+func TestTaskLLMExhaustedRetriesReturnSummedUsage(t *testing.T) {
+	for _, m := range usageModes {
+		t.Run(m.name, func(t *testing.T) {
+			p := &usageAttemptProvider{fail: -1, failErr: errors.New("connection reset by peer")}
+			r := runLogicalCall(t.Context(), m.stream, "p12-exhausted", singleProfileHooks(p, agent.RetryConfig{StreamAttempts: 1, StreamInterval: time.Millisecond}))
+			if r.err == nil || r.err.Error() != "connection reset by peer" {
+				t.Fatalf("err=%v", r.err)
+			}
+			want := sumUsage(failedAttemptUsage, failedAttemptUsage)
+			if r.usage != want || p.calls != 2 {
+				t.Fatalf("usage=%+v want %+v, calls=%d", r.usage, want, p.calls)
+			}
+		})
+	}
+}
+
+func TestTaskLLMProfileFailoverSumsUsage(t *testing.T) {
+	for _, m := range usageModes {
+		t.Run(m.name, func(t *testing.T) {
+			first := &usageAttemptProvider{fail: -1, failErr: errors.New("openai: status 402: insufficient_quota")}
+			second := &usageAttemptProvider{}
+			active := int64(11)
+			hooks := taskLLMStreamHooks{
+				current: func(context.Context) (taskLLMSelection, error) {
+					if active == 11 {
+						return taskLLMSelection{profileID: 11, provider: first}, nil
+					}
+					return taskLLMSelection{profileID: 22, provider: second}, nil
+				},
+				exhaust: func(taskLLMSelection, error) (db.TaskLLMTransition, error) {
+					active = 22
+					next := active
+					return db.TaskLLMTransition{PreviousProfileID: 11, NextProfileID: &next, Advanced: true}, nil
+				},
+			}
+			r := runLogicalCall(t.Context(), m.stream, "p12-failover-"+m.name, hooks)
+			if r.err != nil {
+				t.Fatal(r.err)
+			}
+			want := sumUsage(failedAttemptUsage, okAttemptUsage)
+			if r.usage != want || first.calls != 1 || second.calls != 1 || r.text != "done" {
+				t.Fatalf("usage=%+v want %+v, calls first=%d second=%d, text=%q", r.usage, want, first.calls, second.calls, r.text)
+			}
+		})
+	}
+}
+
+func TestTaskLLMCancelDuringBackoffStops(t *testing.T) {
+	for _, m := range usageModes {
+		t.Run(m.name, func(t *testing.T) {
+			ctx, cancel := context.WithCancel(t.Context())
+			defer cancel()
+			p := &usageAttemptProvider{fail: -1, failErr: errors.New("connection reset by peer")}
+			// Cancel once the retry is waiting out its (long) backoff.
+			p.onFail = func() { time.AfterFunc(20*time.Millisecond, cancel) }
+			r := runLogicalCall(ctx, m.stream, "p12-cancel", singleProfileHooks(p, agent.RetryConfig{StreamAttempts: 2, StreamInterval: time.Minute}))
+			if !errors.Is(r.err, context.Canceled) {
+				t.Fatalf("err=%v, want context.Canceled", r.err)
+			}
+			if p.calls != 1 || r.usage != failedAttemptUsage {
+				t.Fatalf("calls=%d usage=%+v", p.calls, r.usage)
+			}
+		})
+	}
+}
+
+func TestTaskLLMUncommittedStreamAttemptKeepsUsageWithoutReplay(t *testing.T) {
+	p := &usageAttemptProvider{
+		fail:    1,
+		failErr: errors.New("anthropic: overloaded_error"),
+		// Buffered, non-committing events of the failed attempt.
+		failEvents: []llm.StreamEvent{{Type: llm.SEThinkingDelta}, {Type: llm.SEToolInputJSON}},
+	}
+	r := runLogicalCall(t.Context(), true, "p12-uncommitted", singleProfileHooks(p, agent.RetryConfig{StreamAttempts: 1, StreamInterval: time.Millisecond}))
+	if r.err != nil {
+		t.Fatal(r.err)
+	}
+	if want := sumUsage(failedAttemptUsage, okAttemptUsage); r.usage != want {
+		t.Fatalf("usage=%+v want %+v", r.usage, want)
+	}
+	if r.text != "done" {
+		t.Fatalf("text=%q, want the second attempt's only", r.text)
+	}
+	// The failed attempt contributes one usage-only message_start, nothing else.
+	if len(r.events) == 0 || r.events[0].Type != llm.SEMessageStart || r.events[0].Usage != failedAttemptUsage {
+		t.Fatalf("first event %+v", r.events)
+	}
+	for _, ev := range r.events {
+		if ev.Type == llm.SEThinkingDelta || ev.Type == llm.SEToolInputJSON {
+			t.Fatalf("buffered event of the failed attempt was delivered: %+v", r.events)
+		}
+	}
+}
+
+func TestTaskLLMOpaqueThinkingCommitsStream(t *testing.T) {
+	defer withZeroRetryBackoff()()
+	for _, kind := range []llm.StreamEventType{llm.SEThinkingSignature, llm.SERedactedThinking, llm.SEToolUseSignature} {
+		t.Run(string(kind), func(t *testing.T) {
+			if !streamEventCommitsOutput(llm.StreamEvent{Type: kind, Text: "opaque"}) {
+				t.Fatalf("%s does not commit", kind)
+			}
+			// A retryable failure after the opaque block is surfaced, not replayed.
+			p := &scriptedLLMProvider{events: []llm.StreamEvent{{Type: kind, Text: "opaque"}}, err: errors.New("connection reset by peer")}
+			events, err := collectTaskLLMStream(streamTaskLLM(t.Context(), "p12-opaque", llm.CompletionRequest{}, singleProfileHooks(p, agent.RetryConfig{})))
+			if err == nil || p.calls != 1 || len(events) != 1 || events[0].Type != kind {
+				t.Fatalf("calls=%d events=%+v err=%v", p.calls, events, err)
+			}
+		})
+	}
+}
+
+// The recorder below the task router writes one llm_usage row per physical
+// attempt; summing the attempts for the logical call adds no row.
+func TestTaskLLMRetryKeepsOneLedgerRowPerPhysicalCall(t *testing.T) {
+	m, err := NewManager(t.TempDir(), "")
+	if err != nil {
+		t.Skipf("postgres unavailable (%v) — skipping", err)
+	}
+	defer m.Close()
+	pg := m.PG()
+	for _, mode := range usageModes {
+		t.Run(mode.name, func(t *testing.T) {
+			taskID := fmt.Sprintf("p12-ledger-%s-%d", mode.name, time.Now().UnixNano())
+			t.Cleanup(func() { _, _ = pg.Exec(`DELETE FROM llm_usage WHERE task_id=$1`, taskID) })
+			p := &usageAttemptProvider{fail: 1, failErr: errors.New("connection reset by peer")}
+			rec := llmrec.Wrap(p, pg, "test-model", "test-profile", "", "", func() bool { return false })
+			ctx := llmrec.WithTaskID(t.Context(), taskID)
+			r := runLogicalCall(ctx, mode.stream, taskID, singleProfileHooks(rec, agent.RetryConfig{StreamAttempts: 1, StreamInterval: time.Millisecond}))
+			if r.err != nil {
+				t.Fatal(r.err)
+			}
+			var rows, failed int
+			if err := pg.QueryRow(`SELECT count(*), count(*) FILTER (WHERE status='error') FROM llm_usage WHERE task_id=$1`, taskID).Scan(&rows, &failed); err != nil {
+				t.Fatal(err)
+			}
+			if rows != 2 || failed != 1 || p.calls != 2 {
+				t.Fatalf("llm_usage rows=%d (error %d), physical calls=%d; want 2 rows, 1 error", rows, failed, p.calls)
+			}
+		})
+	}
+}

server/tools_wire_test.go

@@ -5,9 +5,9 @@ import (
 	"encoding/json"
 	"testing"
 
-	actool "github.com/Autumn-27/norma/tool"
 	"github.com/Autumn-27/artex/agent"
 	"github.com/Autumn-27/artex/db"
+	actool "github.com/Autumn-27/norma/tool"
 )
 
 func names(tools []actool.CoreTool) map[string]actool.CoreTool {

web/package.json

@@ -21,7 +21,6 @@
   },
   "dependencies": {
     "@antv/g6": "^5.1.1",
-    "@antv/graphin": "^3.0.5",
     "@base-ui/react": "^1.6.0",
     "@dnd-kit/core": "^6.3.1",
     "@dnd-kit/modifiers": "^9.0.0",
@@ -40,7 +39,7 @@
     "geist": "^1.7.2",
     "input-otp": "^1.4.2",
     "lucide-react": "^1.22.0",
-    "next": "^16.2.9",
+    "next": "^16.4.6",
     "next-themes": "^0.4.6",
     "radix-ui": "^1.6.0",
     "react": "^19.2.7",
@@ -52,7 +51,6 @@
     "react-resizable-panels": "^4.12.0",
     "recharts": "^3.8.0",
     "remark-gfm": "^4.0.1",
-    "shadcn": "^4.12.0",
     "simple-icons": "^16.24.1",
     "sonner": "^2.0.7",
     "tailwind-merge": "^3.6.0",
@@ -74,6 +72,7 @@
     "husky": "^9.1.7",
     "lint-staged": "^16.4.0",
     "postcss": "^8.5.16",
+    "shadcn": "^4.12.0",
     "tailwindcss": "^4.1.5",
     "ts-node": "^10.9.2",
     "tw-animate-css": "^1.4.0",