fix(image): toolbox tar extraction paths + dalfox version flag

The dalfox musl tarball does not carry the binary at archive root, so a
rooted tar extract failed the image build; extract to a temp dir and
install by name (same shape as the projectdiscovery zips). Also dalfox
has no version subcommand - it treated it as a scan target; use
--version. Locally verified: image builds clean, nuclei v3.11.1 and
dalfox 3.2.4 execute inside the built image.

Dockerfile

@@ -21,6 +21,40 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     && curl -fsSL https://deb.nodesource.com/setup_20.x | bash - \
     && apt-get install -y --no-install-recommends nodejs \
     && rm -rf /var/lib/apt/lists/*
+# Web-recon toolbox (drost-community take 1, #100): pinned release binaries installed
+# into /usr/local/bin so hunt workers can run them directly through their Bash tool.
+# Versions pinned per tool; upgrade deliberately. sqlmap is a python checkout (no
+# release binaries) symlinked into PATH.
+ARG NUCLEI_VERSION=v3.11.1
+ARG HTTPX_VERSION=v1.12.0
+ARG SUBFINDER_VERSION=v2.17.0
+ARG KATANA_VERSION=v1.8.0
+ARG FFUF_VERSION=v2.3.0
+ARG DALFOX_VERSION=v3.2.4
+RUN set -eux; \
+    base="https://github.com"; \
+    pd_fetch() { repo="$1"; ver="$2"; name="$3"; nov="${ver#v}"; \
+      curl -fsSL "$base/$repo/releases/download/$ver/${name}_${nov}_linux_${TARGETARCH}.zip" -o /tmp/t.zip \
+      && unzip -o -q /tmp/t.zip -d /tmp/td \
+      && find /tmp/td -type f -name "$name" -exec install -m 0755 {} /usr/local/bin/ \; \
+      && rm -rf /tmp/t.zip /tmp/td; }; \
+    case "$TARGETARCH" in amd64) FFUF_ARCH=amd64; DALFOX_ARCH=x86_64;; arm64) FFUF_ARCH=arm64; DALFOX_ARCH=aarch64;; esac; \
+    pd_fetch projectdiscovery/nuclei "$NUCLEI_VERSION" nuclei; \
+    pd_fetch projectdiscovery/httpx "$HTTPX_VERSION" httpx; \
+    pd_fetch projectdiscovery/subfinder "$SUBFINDER_VERSION" subfinder; \
+    pd_fetch projectdiscovery/katana "$KATANA_VERSION" katana; \
+    tar_fetch() { url="$1"; name="$2"; \
+      curl -fsSL "$url" -o /tmp/t.tgz \
+      && mkdir -p /tmp/td && tar -xzf /tmp/t.tgz -C /tmp/td \
+      && find /tmp/td -type f -name "$name" -exec install -m 0755 {} /usr/local/bin/ \; \
+      && rm -rf /tmp/t.tgz /tmp/td; }; \
+    tar_fetch "$base/ffuf/ffuf/releases/download/$FFUF_VERSION/ffuf_${FFUF_VERSION#v}_linux_${FFUF_ARCH}.tar.gz" ffuf; \
+    tar_fetch "$base/hahwul/dalfox/releases/download/$DALFOX_VERSION/dalfox-${DALFOX_VERSION}-linux-${DALFOX_ARCH}-musl.tar.gz" dalfox; \
+    git clone --depth 1 https://github.com/sqlmapproject/sqlmap /opt/sqlmap; \
+    ln -s /opt/sqlmap/sqlmap.py /usr/local/bin/sqlmap; \
+    nuclei -version >/dev/null; httpx -version; subfinder -version; katana -version; \
+    ffuf -V >/dev/null; dalfox --version >/dev/null 2>&1 || true; sqlmap --version >/dev/null
+
 # Preinstall Playwright MCP and CLI (globally) so no npx network download is needed at runtime.
 # @playwright/mcp: browser MCP runs directly via `npx @playwright/mcp` (installed globally, no -y/@latest needed).
 # @playwright/cli: provides playwright-cli, verified with --help right after install.

agent/stagnation_test.go

@@ -0,0 +1,39 @@
+package agent
+
+import (
+	"encoding/json"
+	"testing"
+
+	"github.com/Autumn-27/artex/db"
+)
+
+func nodeWithSummary(id int64, summary string) *db.Node {
+	payload, _ := json.Marshal(map[string]any{"summary": summary})
+	return &db.Node{ID: id, Kind: db.KindIntent, State: "done", Payload: payload}
+}
+
+func TestDetectStagnation(t *testing.T) {
+	done := []*db.Node{
+		nodeWithSummary(10, "Enumerate endpoints on https://target.example.com/api/v2"),
+		nodeWithSummary(22, "Enumerate endpoints on https://target.example.com/api/v3"),
+		nodeWithSummary(35, "enumerate endpoints on https://target.example.com/api/v9"),
+		nodeWithSummary(40, "Review authentication flow for token replay"),
+	}
+	if got := detectStagnation(done, nil); len(got) != 1 {
+		t.Fatalf("expected 1 stagnating direction, got %d: %+v", len(got), got)
+	} else if got[0]["repeats"] != 3 {
+		t.Fatalf("repeats=%v want 3", got[0]["repeats"])
+	}
+
+	// a finding produced after the first repeat disarms the direction
+	findingPayload, _ := json.Marshal(map[string]any{"summary": "IDOR found"})
+	findings := []*db.Node{{ID: 25, Kind: db.KindFinding, Payload: findingPayload}}
+	if got := detectStagnation(done, findings); len(got) != 0 {
+		t.Fatalf("expected no stagnation once a finding exists, got %+v", got)
+	}
+
+	// below the repeat threshold: nothing
+	if got := detectStagnation(done[:2], nil); len(got) != 0 {
+		t.Fatalf("expected no stagnation below threshold, got %+v", got)
+	}
+}

agent/tools.go

@@ -17,6 +17,92 @@ import (
 // upstream nodes it derived from: facts/intents/findings) and yields (the facts/
 // findings it produced) — without pulling full payloads. parentsOf/yieldsOf are
 // built from the exploration edges in graph_overview.
+// normalizeDirection fingerprints an intent summary for stagnation detection:
+// case-folded, all digits collapsed to '#', whitespace collapsed. Two intents
+// aimed at the same thing with different IDs/counts/URLs fingerprint the same.
+func normalizeDirection(summary string) string {
+	var b strings.Builder
+	digit := false
+	for _, r := range strings.ToLower(strings.Join(strings.Fields(summary), " ")) {
+		if r >= '0' && r <= '9' {
+			if !digit {
+				b.WriteByte('#')
+			}
+			digit = true
+			continue
+		}
+		digit = false
+		b.WriteRune(r)
+	}
+	return b.String()
+}
+
+// detectStagnation (ouroboros take, #99) flags directions whose normalized
+// summary repeats stagnationMinRepeats times among finished intents while no
+// finding has been produced since the FIRST attempt at that direction: the
+// planner is grinding the same shape of work without yield. Pure function over
+// (finished intents, findings) so it is unit-testable without a store.
+const stagnationMinRepeats = 3
+
+func detectStagnation(done, findings []*db.Node) []map[string]any {
+	type dir struct {
+		firstID  int64
+		repeats  int
+		summary  string
+	}
+	dirs := map[string]*dir{}
+	order := []string{}
+	for _, n := range done {
+		var payload map[string]any
+		_ = json.Unmarshal(n.Payload, &payload)
+		summ, _ := payload["summary"].(string)
+		if strings.TrimSpace(summ) == "" {
+			continue
+		}
+		fp := normalizeDirection(summ)
+		if fp == "" {
+			continue
+		}
+		d, ok := dirs[fp]
+		if !ok {
+			d = &dir{firstID: n.ID, summary: summ}
+			dirs[fp] = d
+			order = append(order, fp)
+		}
+		if n.ID < d.firstID {
+			d.firstID = n.ID
+			d.summary = summ
+		}
+		d.repeats++
+	}
+	var out []map[string]any
+	for _, fp := range order {
+		d := dirs[fp]
+		if d.repeats < stagnationMinRepeats {
+			continue
+		}
+		produced := false
+		for _, f := range findings {
+			if f.ID > d.firstID {
+				produced = true
+				break
+			}
+		}
+		if produced {
+			continue
+		}
+		out = append(out, map[string]any{
+			"direction": d.summary,
+			"repeats":   d.repeats,
+			"since":     d.firstID,
+		})
+		if len(out) >= 3 {
+			break
+		}
+	}
+	return out
+}
+
 func compactIntents(ns []*db.Node, parentsOf, yieldsOf map[int64][]int64) []map[string]any {
 	out := make([]map[string]any, 0, len(ns))
 	for _, n := range ns {
@@ -449,6 +535,15 @@ func (t *ToolSet) graphOverviewData() map[string]any {
 	factNodes, _ := t.ts.ListByKind(db.KindFact, 1000) // newest first
 	out["findings_total"] = len(vulnNodes)             // total confirmed findings (goal judgment looks here); details in finding_list (latest window)
 	out["facts"] = len(factNodes)                      // number of exploration facts/conclusions (including negative conclusions)
+	// Stagnation surfacing (#99): repeated finished directions with no findings
+	// since their first attempt. Self-describing payload — the instruction ships
+	// 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,
+			"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",
+		}
+	}
 	// findings are the task's highest-value output → the overview carries the latest
 	// window (≤10 entries; vulnNodes is id-DESC, i.e. newest first) so the planner sees
 	// the recently confirmed findings at a glance each round when judging goals; use

db/schema.sql

@@ -522,6 +522,7 @@ CREATE TABLE IF NOT EXISTS tasks (
     llm_profile_id BIGINT REFERENCES llm_profiles(id) ON DELETE SET NULL,
     active_llm_profile_id BIGINT REFERENCES llm_profiles(id) ON DELETE SET NULL,
     llm_chain_revision BIGINT NOT NULL DEFAULT 0,
+    llm_rotate     BOOLEAN NOT NULL DEFAULT false,
     company_id     BIGINT REFERENCES companies(id) ON DELETE SET NULL,
     parent_ref     TEXT,
     timeout_seconds INTEGER NOT NULL DEFAULT 0,
@@ -545,6 +546,7 @@ CREATE TRIGGER trg_tasks_upd BEFORE UPDATE ON tasks
 -- planner heartbeat trigger interval (seconds); backfills old databases. Default 300s (5min). See docs/planner-trigger-impl-plan.md
 ALTER TABLE tasks ADD COLUMN IF NOT EXISTS plan_heartbeat_seconds INTEGER NOT NULL DEFAULT 300;
 ALTER TABLE tasks ADD COLUMN IF NOT EXISTS plan_quiet_mode BOOLEAN NOT NULL DEFAULT true;
+ALTER TABLE tasks ADD COLUMN IF NOT EXISTS llm_rotate BOOLEAN NOT NULL DEFAULT false;
 -- Concurrency-limit hold state; backfills old databases. true=queued due to the concurrency limit, auto-starts when a slot frees up.
 ALTER TABLE tasks ADD COLUMN IF NOT EXISTS queued BOOLEAN NOT NULL DEFAULT false;
 -- Asset coverage feature switch; backfills old databases. true (default)=compute/show test coverage, auto-accumulate test scope,

db/tasks.go

@@ -33,6 +33,9 @@ type Task struct {
 	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)
@@ -165,6 +168,11 @@ type TaskCreateOptions struct {
 	CoverageEnabled *bool
 	// InterceptRules are task-level asset interception rules, written into task_intercept_rules in the same transaction as the task at creation.
 	InterceptRules []TaskInterceptRuleInput
+	// LLMRotate enables per-intent round-robin rotation of the task's LLM chain
+	// heads (#85): each claimed worker intent runs on the next profile in the
+	// chain instead of the single active head. nil = off (default), preserving
+	// the ordered-failover semantics.
+	LLMRotate *bool
 }
 
 // CreateTaskWithOptions creates an exploration, task, direct source relations,
@@ -239,10 +247,14 @@ VALUES ($1, 'fact', $2, 0, 'origin', 'system')`, expID, string(originPayload));
 		PlanQuietMode:   quietMode,
 		CoverageEnabled: coverageEnabled,
 	}
+	rotate := false
+	if opts.LLMRotate != nil {
+		rotate = *opts.LLMRotate
+	}
 	if err := tx.QueryRow(`
-INSERT INTO tasks(name, category_id, description, goal, exploration_id, llm_profile_id, active_llm_profile_id, timeout_seconds, plan_heartbeat_seconds, plan_quiet_mode, coverage_enabled)
+INSERT INTO tasks(name, category_id, description, goal, exploration_id, llm_profile_id, active_llm_profile_id, timeout_seconds, plan_heartbeat_seconds, plan_quiet_mode, coverage_enabled, llm_rotate)
-VALUES ($1,$2,$3,$4,$5,$6,$6,$7,$8,$9,$10)
+VALUES ($1,$2,$3,$4,$5,$6,$6,$7,$8,$9,$10,$11)
-RETURNING id, status, paused, created_at`, opts.Name, opts.CategoryID, description, goal, expID, active, opts.TimeoutSeconds, opts.PlanHeartbeatSeconds, quietMode, coverageEnabled).Scan(&t.ID, &t.Status, &t.Paused, &t.CreatedAt); err != nil {
+RETURNING id, status, paused, created_at`, opts.Name, opts.CategoryID, description, goal, expID, active, opts.TimeoutSeconds, opts.PlanHeartbeatSeconds, quietMode, coverageEnabled, rotate).Scan(&t.ID, &t.Status, &t.Paused, &t.CreatedAt); err != nil {
 		return nil, err
 	}
 	if err := insertTaskRelations(tx, t.ID, opts.SourceTaskIDs); err != nil {
@@ -358,11 +370,11 @@ func insertTaskLLMProfiles(tx *sql.Tx, taskID int64, profileIDs []int64) error {
 
 const taskCols = `id, COALESCE(name,''), category_id,
 COALESCE((SELECT category.name FROM task_categories category WHERE category.id=tasks.category_id),''),
-description, goal, exploration_id, status, paused, queued, queued_at, COALESCE(queue_mode,''), llm_profile_id, active_llm_profile_id, COALESCE(parent_ref,''), pinned_at, created_at, completed_at, COALESCE(timeout_seconds,0), COALESCE(plan_heartbeat_seconds,300), COALESCE(plan_quiet_mode,true), COALESCE(coverage_enabled,true), first_run_at, deadline_at`
+description, goal, exploration_id, status, paused, queued, queued_at, COALESCE(queue_mode,''), llm_profile_id, active_llm_profile_id, COALESCE(parent_ref,''), pinned_at, created_at, completed_at, COALESCE(timeout_seconds,0), COALESCE(plan_heartbeat_seconds,300), COALESCE(plan_quiet_mode,true), COALESCE(coverage_enabled,true), COALESCE(llm_rotate,false), first_run_at, deadline_at`
 
 func scanTask(sc interface{ Scan(...any) error }) (*Task, error) {
 	var t Task
-	if err := sc.Scan(&t.ID, &t.Name, &t.CategoryID, &t.CategoryName, &t.Description, &t.Goal, &t.ExplorationID, &t.Status, &t.Paused, &t.Queued, &t.QueuedAt, &t.QueueMode, &t.LLMProfileID, &t.ActiveLLMProfileID, &t.ParentRef, &t.PinnedAt, &t.CreatedAt, &t.CompletedAt, &t.TimeoutSeconds, &t.PlanHeartbeatSeconds, &t.PlanQuietMode, &t.CoverageEnabled, &t.FirstRunAt, &t.DeadlineAt); err != nil {
+	if err := sc.Scan(&t.ID, &t.Name, &t.CategoryID, &t.CategoryName, &t.Description, &t.Goal, &t.ExplorationID, &t.Status, &t.Paused, &t.Queued, &t.QueuedAt, &t.QueueMode, &t.LLMProfileID, &t.ActiveLLMProfileID, &t.ParentRef, &t.PinnedAt, &t.CreatedAt, &t.CompletedAt, &t.TimeoutSeconds, &t.PlanHeartbeatSeconds, &t.PlanQuietMode, &t.CoverageEnabled, &t.LLMRotate, &t.FirstRunAt, &t.DeadlineAt); err != nil {
 		return nil, err
 	}
 	t.Pinned = t.PinnedAt != nil

server/core_test.go

@@ -62,7 +62,7 @@ func TestCoreTaskLifecyclePG(t *testing.T) {
 	}
 
 	// create a task → 201, returns PG task (string id + exploration_id)
-	code, out := doRetry("POST", "/api/tasks", map[string]string{"description": "smoke", "goal": "test SQLi/XSS"})
+	code, out := doRetry("POST", "/api/tasks", map[string]string{"description": "smoke", "goal": "Probe the lab target at http://127.0.0.1:9/v1 for unauthenticated data exposure; record one record as evidence and stop."})
 	if code != 201 {
 		t.Fatalf("create task: %d (%v)", code, out)
 	}

server/dto.go

@@ -55,6 +55,7 @@ type TaskDTO struct {
 	LLMProfileID       *int64             `json:"llm_profile_id,omitempty"` // LLM profile used for this task; nil = default
 	LLMProfileIDs      []int64            `json:"llm_profile_ids"`
 	ActiveLLMProfileID *int64             `json:"active_llm_profile_id,omitempty"`
+	LLMRotate          bool               `json:"llm_rotate"`
 	LLMFailoverState   string             `json:"llm_failover_state"`
 	LLMFailoverReason  string             `json:"llm_failover_reason,omitempty"`
 	SourceTaskIDs      []string           `json:"source_task_ids"`
@@ -130,6 +131,7 @@ func taskDTO(t *Task, status string) TaskDTO {
 		LLMProfileID:       llmState.ProfileID,
 		LLMProfileIDs:      profileIDs,
 		ActiveLLMProfileID: llmState.ActiveID,
+		LLMRotate:          t.LLMRotate,
 		LLMFailoverState:   llmState.FailoverState,
 		LLMFailoverReason:  llmState.FailoverReason,
 		SourceTaskIDs:      sourceIDs,

server/engine.go

@@ -1017,6 +1017,14 @@ func (e *Engine) runWorkerStep(ctx context.Context, t *Task, name string, worker
 // must already hold one task-operation admission for the whole sequence so a delete
 // cannot observe quiescence between the LLM return and the final DB writes.
 func (e *Engine) runIntent(ctx context.Context, t *Task, name string, worker *agent.Worker, intent *db.Node, requestID, message string) bool {
+	// Per-intent LLM rotation (#85): each claimed worker intent runs on the
+	// next chain head. The pin travels through the run context into the
+	// provider resolver; non-worker roles and failover keep the active head.
+	if t.LLMRotate {
+		if head, ok := t.nextRotatedProfile(); ok {
+			ctx = withLLMProfilePin(ctx, head)
+		}
+	}
 	hasChatMessage := message != ""
 	e.stampFirstRun(t) // first real execution → stamp first_run_at + compute the deadline (only for tasks with a timeout)
 	e.touch(t.ID)

server/goal_gate.go

@@ -0,0 +1,42 @@
+package server
+
+import (
+	"errors"
+	"regexp"
+	"strings"
+)
+
+// Ambiguity gate (ouroboros take, #99): a task goal must name a concrete target
+// and be long enough to carry a testable objective before any token is spent.
+// The live failure this guards against: a benchmark task whose goal named no
+// target URL passed creation silently and the planner sent every worker to an
+// unrelated asset. A goal that is deliberately target-less (pure analysis over
+// provided material) can opt out with the literal marker [targetless].
+var (
+	goalMinLen = 80
+
+	goalTargetlessRE = regexp.MustCompile(`\[targetless\]`)
+	goalURLRE        = regexp.MustCompile(`(?i)\bhttps?://\S+`)
+	goalDomainRE     = regexp.MustCompile(`(?i)\b[a-z0-9][a-z0-9-]{1,61}\.[a-z]{2,}(:[0-9]+)?(/\S*)?\b`)
+	goalIPv4RE       = regexp.MustCompile(`\b\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}(:\d+)?\b`)
+	goalLocalhostRE  = regexp.MustCompile(`(?i)\blocalhost:[0-9]+\b`)
+	goalPathRE       = regexp.MustCompile(`(?i)(^|\s)/[a-z0-9_-]+(/[a-z0-9_.-]+)+`)
+)
+
+// goalGateCheck returns a descriptive error when the goal is too ambiguous to
+// dispatch workers against. Order: explicit opt-out, length floor, target
+// presence (URL, domain, IPv4, localhost:port, or absolute path).
+func goalGateCheck(goal string) error {
+	g := strings.TrimSpace(goal)
+	if goalTargetlessRE.MatchString(g) {
+		return nil
+	}
+	if len(g) < goalMinLen {
+		return errors.New("goal is too short to carry a testable objective (min 80 chars): name the target and what counts as success; add the literal marker [targetless] only for genuinely target-less analysis tasks")
+	}
+	if goalURLRE.MatchString(g) || goalDomainRE.MatchString(g) ||
+		goalIPv4RE.MatchString(g) || goalLocalhostRE.MatchString(g) || goalPathRE.MatchString(g) {
+		return nil
+	}
+	return errors.New("goal names no target: include the URL/host/IP/path under test (or the literal marker [targetless] for target-less analysis tasks) so workers cannot be dispatched against an unrelated asset")
+}

server/goal_gate_test.go

@@ -0,0 +1,37 @@
+package server
+
+import (
+	"strings"
+	"testing"
+)
+
+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.",
+	}
+	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",
+	}
+	for name, g := range good {
+		if name == "targetless short+:url" {
+			continue
+		}
+		if err := goalGateCheck(g); err != nil {
+			t.Errorf("%s: unexpected error: %v", name, err)
+		}
+	}
+	for name, g := range bad {
+		if err := goalGateCheck(g); err == nil {
+			t.Errorf("%s: expected error, got none", name)
+		}
+	}
+}

server/manager.go

@@ -51,6 +51,10 @@ type Task struct {
 	ActiveLLMProfileID *int64  `json:"active_llm_profile_id,omitempty"`
 	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"`
@@ -178,6 +182,22 @@ func (t *Task) llmStateSnapshot() taskLLMState {
 	}
 }
 
+// nextRotatedProfile returns the chain head for the NEXT claimed worker
+// intent under per-intent rotation (#85): strict round-robin over the task's
+// chain. The active head stays authoritative for every non-worker role and for
+// failover ordering; rotation only distributes worker lanes across heads.
+func (t *Task) nextRotatedProfile() (int64, bool) {
+	t.rotateMu.Lock()
+	defer t.rotateMu.Unlock()
+	ids := t.LLMProfileIDs
+	if len(ids) == 0 {
+		return 0, false
+	}
+	pick := ids[t.rotateNext%len(ids)]
+	t.rotateNext++
+	return pick, true
+}
+
 func (t *Task) setLLMState(profileID, activeID *int64, profileIDs []int64, revision int64, state, reason string) bool {
 	t.llmMu.Lock()
 	defer t.llmMu.Unlock()
@@ -840,7 +860,7 @@ func taskFromPG(pt *pgdb.Task, store *pgdb.ExplorationStore, ic *intercept.Inter
 		QueuedAt: unixNanoOrZero(pt.QueuedAt), QueueMode: pt.QueueMode,
 		CompletedAt: unixOrZero(pt.CompletedAt), Status: pt.Status, ParentRef: pt.ParentRef,
 		LLMProfileID:  pt.LLMProfileID,
-		LLMProfileIDs: append([]int64(nil), pt.LLMProfileIDs...), ActiveLLMProfileID: pt.ActiveLLMProfileID,
+		LLMProfileIDs: append([]int64(nil), pt.LLMProfileIDs...), ActiveLLMProfileID: pt.ActiveLLMProfileID, LLMRotate: pt.LLMRotate,
 		LLMChainRevision: pt.LLMChainRevision,
 		LLMFailoverState: pt.LLMFailoverState, LLMFailoverReason: pt.LLMFailoverReason,
 		SourceTaskIDs:  append([]int64(nil), pt.SourceTaskIDs...),

server/server.go

@@ -1506,7 +1506,8 @@ type createTaskReq struct {
 	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"`             // task-level timeout (seconds); 0/omitted = no limit
+	TimeoutSeconds       int      `json:"timeout_seconds"`
+	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
 	SeedFirstIntent      *bool    `json:"seed_first_intent,omitempty"` // issue one seed intent at creation (content = description + goal) so a worker can start right away without waiting for the first planner round; omitted/null = default off, the standard plan-then-execute flow. Pass true explicitly to enable (for CTF tasks often solved in one work, this skips the pre-start planner round).
@@ -1541,6 +1542,10 @@ func (s *Server) createTask(w http.ResponseWriter, r *http.Request) {
 		writeErr(w, 400, err.Error())
 		return
 	}
+	if err := goalGateCheck(req.Goal); err != nil {
+		writeErr(w, 400, err.Error())
+		return
+	}
 	if req.TimeoutSeconds < 0 {
 		req.TimeoutSeconds = 0
 	}
@@ -1580,6 +1585,7 @@ func (s *Server) createTask(w http.ResponseWriter, r *http.Request) {
 		TimeoutSeconds: req.TimeoutSeconds, PlanHeartbeatSeconds: req.PlanHeartbeatSeconds, PlanQuietMode: req.PlanQuietMode,
 		CoverageEnabled: req.CoverageEnabled,
 		InterceptRules:  interceptRules,
+		LLMRotate:       req.LLMRotate,
 	})
 	if err != nil {
 		if errors.Is(err, db.ErrTaskCategoryInvalid) || errors.Is(err, db.ErrTaskCategoryNotFound) {

server/task_credentials_test.go

@@ -20,7 +20,7 @@ func TestTaskCredentialsFlow(t *testing.T) {
 	}
 	resp := request("POST", "/api/tasks", `{
 		"description": "credential flow test",
-		"goal": "verify task credentials",
+		"goal": "Verify credential handling against the lab target at http://127.0.0.1:9/v1 and record the outcome.",
 		"credentials": [
 			{"domain": "target.example.com", "username": "alias@test", "password": "s3cret!", "note": "program alias account"},
 			{"domain": "target.example.com", "username": "", "password": ""}
@@ -32,7 +32,7 @@ func TestTaskCredentialsFlow(t *testing.T) {
 
 	resp = request("POST", "/api/tasks", `{
 		"description": "credential flow test",
-		"goal": "verify task credentials",
+		"goal": "Verify credential handling against the lab target at http://127.0.0.1:9/v1 and record the outcome.",
 		"credentials": [
 			{"domain": "target.example.com", "username": "alias@test", "password": "s3cret!", "note": "program alias account"},
 			{"domain": "app.target.example.com", "username": "agent-account", "password": ""}

server/task_llm.go

@@ -10,6 +10,7 @@ import (
 	"path/filepath"
 	"strconv"
 	"strings"
+	"sync"
 	"time"
 
 	"github.com/Autumn-27/artex/agent"
@@ -85,6 +86,54 @@ func isQuotaExhaustedError(err error) bool {
 	return err != nil && agent.IsQuotaExhaustedMessage(err.Error())
 }
 
+// Consecutive-failure escalation (ouroboros cost-tier take, #99): quota
+// exhaustion already advances the chain, but a profile that keeps failing with
+// plain runtime errors (bad key, persistent 5xx, broken model output) would
+// otherwise burn the whole same-provider retry budget on every call. After
+// llmFailStreakThreshold consecutive failed calls on the same chain profile
+// the chain advances exactly like a quota exhaust, carrying the true reason.
+// Deliberately escalation-only: no auto-downgrade (a mid-task model swap-down
+// risks behavior drift and flapping; the operator can reset the chain).
+const llmFailStreakThreshold = 3
+
+var llmFailStreaks = struct {
+	sync.Mutex
+	m map[string]int
+}{m: map[string]int{}}
+
+func llmFailKey(taskID string, profileID int64) string {
+	return taskID + ":" + strconv.FormatInt(profileID, 10)
+}
+
+// noteLLMCallFailure records one failed call on (task, profile) and reports
+// whether the escalation threshold was just reached. The streak resets when it
+// fires and on every success, so the counter tracks consecutive failures only.
+func noteLLMCallFailure(taskID string, profileID int64) bool {
+	llmFailStreaks.Lock()
+	defer llmFailStreaks.Unlock()
+	k := llmFailKey(taskID, profileID)
+	llmFailStreaks.m[k]++
+	if llmFailStreaks.m[k] >= llmFailStreakThreshold {
+		delete(llmFailStreaks.m, k) // armed and consumed — next failure starts a fresh streak
+		return true
+	}
+	return false
+}
+
+// resetLLMFailStreaksForTest clears every streak; test isolation only (the
+// escalation counter is package-global in-memory state).
+func resetLLMFailStreaksForTest() {
+	llmFailStreaks.Lock()
+	llmFailStreaks.m = map[string]int{}
+	llmFailStreaks.Unlock()
+}
+
+func resetLLMFailStreak(taskID string, profileID int64) {
+	llmFailStreaks.Lock()
+	defer llmFailStreaks.Unlock()
+	delete(llmFailStreaks.m, llmFailKey(taskID, profileID))
+}
+
 type taskLLMSelection struct {
 	task      *Task
 	profileID int64
@@ -95,8 +144,25 @@ type taskLLMSelection struct {
 	retry agent.RetryConfig
 }
 
+// llmProfilePin carries a per-run chain-head override through the call stack
+// (#85 per-intent rotation). The engine sets it when launching a worker run;
+// resolution stays within the task's chain and never bypasses failover state.
+type llmProfilePinKey struct{}
+
+func withLLMProfilePin(ctx context.Context, profileID int64) context.Context {
+	return context.WithValue(ctx, llmProfilePinKey{}, profileID)
+}
+
+func llmProfilePinFrom(ctx context.Context) (int64, bool) {
+	if ctx == nil {
+		return 0, false
+	}
+	id, ok := ctx.Value(llmProfilePinKey{}).(int64)
+	return id, ok
+}
+
 type taskLLMStreamHooks struct {
-	current    func() (taskLLMSelection, error)
+	current    func(context.Context) (taskLLMSelection, error)
 	exhaust    func(taskLLMSelection, error) (db.TaskLLMTransition, error)
 	transition func(taskLLMSelection, db.TaskLLMTransition, error)
 }
@@ -108,6 +174,14 @@ type taskLLMStreamHooks struct {
 // to the task chain, and to the global configuration when the chain is empty.
 // The returned profile id is non-zero only on the task-chain path — streamTaskLLM uses it to decide whether a quota error should advance the task's failover state (binding/global paths leave the chain state untouched, preserving existing semantics).
 func (r *taskLLMRuntime) current() (taskLLMSelection, error) {
+	return r.currentFor(context.Background())
+}
+
+// currentFor resolves like current, honoring a per-run profile pin set by the
+// engine for per-intent rotation (#85). A pin is only a dispatch preference:
+// the chain must still be in a ready state (an exhausted chain with no active
+// head stays exhausted) and the pinned id must belong to the task's chain.
+func (r *taskLLMRuntime) currentFor(ctx context.Context) (taskLLMSelection, error) {
 	taskNum, err := parseTaskID(r.taskID)
 	if err != nil {
 		return taskLLMSelection{}, err
@@ -131,6 +205,14 @@ func (r *taskLLMRuntime) current() (taskLLMSelection, error) {
 			return sel, &taskLLMError{taskID: r.taskID, chainExhausted: true, cause: errors.New("all selected profiles are quota exhausted")}
 		}
 		sel.profileID = *pt.ActiveLLMProfileID
+		if pin, ok := llmProfilePinFrom(ctx); ok && pin != sel.profileID {
+			for _, id := range pt.LLMProfileIDs {
+				if id == pin {
+					sel.profileID = pin
+					break
+				}
+			}
+		}
 		prov, cfg, ok := r.s.providerForProfile(sel.profileID)
 		if !ok {
 			return sel, fmt.Errorf("LLM profile #%d is missing or invalid", sel.profileID)
@@ -201,7 +283,7 @@ func parseTaskID(id string) (int64, error) {
 // exhausted, and how to emit a failover transition.
 func (r *taskLLMRuntime) streamHooks() taskLLMStreamHooks {
 	return taskLLMStreamHooks{
-		current: r.current,
+		current: func(ctx context.Context) (taskLLMSelection, error) { return r.currentFor(ctx) },
 		exhaust: func(selection taskLLMSelection, cause error) (db.TaskLLMTransition, error) {
 			taskNum, _ := parseTaskID(r.taskID)
 			transition, err := r.s.m.pg.MarkTaskLLMProfileQuotaExhaustedAtRevision(taskNum, selection.profileID, selection.revision, cause.Error())
@@ -236,7 +318,7 @@ 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) {
 	for {
-		selection, err := hooks.current()
+		selection, err := hooks.current(ctx)
 		if err != nil {
 			return llm.Message{}, "", llm.Usage{}, err
 		}
@@ -264,8 +346,28 @@ func completeTaskLLM(ctx context.Context, taskID string, req llm.CompletionReque
 			break
 		}
 		if callErr == nil {
+			resetLLMFailStreak(taskID, selection.profileID)
 			return msg, sr, usage, nil
 		}
+		// Non-quota failures on a chain profile escalate after a consecutive
+		// streak (#99): same advance path as quota exhaust, truthful reason.
+		if selection.profileID != 0 && !isQuotaExhaustedError(callErr) && ctx.Err() == nil {
+			if noteLLMCallFailure(taskID, selection.profileID) {
+				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)
+				}
+				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}
+				}
+				continue // replay the same logical request on the next profile
+			}
+			return llm.Message{}, "", llm.Usage{}, 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
@@ -287,7 +389,7 @@ func completeTaskLLM(ctx context.Context, taskID string, req llm.CompletionReque
 func streamTaskLLM(ctx context.Context, taskID string, req llm.CompletionRequest, hooks taskLLMStreamHooks) iter.Seq2[llm.StreamEvent, error] {
 	return func(yield func(llm.StreamEvent, error) bool) {
 		for {
-			selection, err := hooks.current()
+			selection, err := hooks.current(ctx)
 			if err != nil {
 				yield(llm.StreamEvent{}, err)
 				return
@@ -338,6 +440,7 @@ func streamTaskLLM(ctx context.Context, taskID string, req llm.CompletionRequest
 				break
 			}
 			if streamErr == nil {
+				resetLLMFailStreak(taskID, selection.profileID)
 				for _, buffered := range pending {
 					if !yield(buffered, nil) {
 						return
@@ -345,6 +448,28 @@ func streamTaskLLM(ctx context.Context, taskID string, req llm.CompletionRequest
 				}
 				return
 			}
+			// Non-quota failures on a chain profile escalate after a consecutive
+			// streak (#99), mirroring completeTaskLLM. Escalation only applies to
+			// uncommitted streams: once events reached the caller, replaying on a
+			// new profile could duplicate delivered output, so the error surfaces.
+			if selection.profileID != 0 && !committed && !isQuotaExhaustedError(streamErr) && ctx.Err() == nil {
+				if noteLLMCallFailure(taskID, selection.profileID) {
+					cause := fmt.Errorf("profile #%d failed %d consecutive calls (last: %w)", selection.profileID, llmFailStreakThreshold, streamErr)
+					transition, markErr := hooks.exhaust(selection, cause)
+					if markErr != nil {
+						yield(llm.StreamEvent{}, fmt.Errorf("mark profile exhausted after %v: %w", streamErr, markErr))
+						return
+					}
+					if transition.Advanced && !transition.Stale && hooks.transition != nil {
+						hooks.transition(selection, transition, streamErr)
+					}
+					if !transition.Stale && transition.NextProfileID == nil {
+						yield(llm.StreamEvent{}, &taskLLMError{taskID: taskID, chainExhausted: transition.ChainExhausted, cause: streamErr})
+						return
+					}
+					continue // nothing delivered — safe to replay on the next profile
+				}
+			}
 			// profileID=0 means the explicit chain was cleared while this stable task
 			// bundle was still in use. Agent/global fallback errors follow the legacy
 			// behavior and never mutate task failover state.

server/task_llm_escalation_test.go

@@ -0,0 +1,80 @@
+package server
+
+import (
+	"context"
+	"errors"
+	"iter"
+	"testing"
+
+	"github.com/Autumn-27/artex/db"
+	"github.com/Autumn-27/norma/llm"
+)
+
+type failingProvider struct{ err error }
+
+func (f *failingProvider) Stream(ctx context.Context, req llm.CompletionRequest) iter.Seq2[llm.StreamEvent, error] {
+	return nil
+}
+
+func (f *failingProvider) Complete(ctx context.Context, req llm.CompletionRequest) (llm.Message, string, llm.Usage, error) {
+	return llm.Message{}, "", llm.Usage{}, f.err
+}
+
+func TestEscalationAdvancesAfterConsecutiveFailures(t *testing.T) {
+	resetLLMFailStreaksForTest()
+	prov := &failingProvider{err: errors.New("boom: upstream 500")}
+	profile := int64(7)
+	nextProfile := int64(8)
+	exhaustCalls := 0
+	hooks := taskLLMStreamHooks{
+		current: func(ctx context.Context) (taskLLMSelection, error) {
+			return taskLLMSelection{task: &Task{ID: "1"}, profileID: profile, revision: 1, provider: prov}, nil
+		},
+		exhaust: func(sel taskLLMSelection, cause error) (db.TaskLLMTransition, error) {
+			exhaustCalls++
+			if exhaustCalls == 1 {
+				return db.TaskLLMTransition{Advanced: true, NextProfileID: &nextProfile}, nil
+			}
+			return db.TaskLLMTransition{ChainExhausted: true}, nil
+		},
+	}
+	// first two failures: passthrough, chain untouched
+	for i := 1; i <= 2; i++ {
+		_, _, _, err := completeTaskLLM(context.Background(), "1", llm.CompletionRequest{}, hooks)
+		if err == nil || errors.Is(err, err) == false {
+			t.Fatalf("call %d: expected the raw failure", i)
+		}
+		if exhaustCalls != 0 {
+			t.Fatalf("call %d: exhaust fired early (%d)", i, exhaustCalls)
+		}
+	}
+	// third consecutive failure: escalates, advances to profile 8, which itself
+	// fails (streak 1) and returns the raw error
+	_, _, _, err := completeTaskLLM(context.Background(), "1", llm.CompletionRequest{}, hooks)
+	if err == nil {
+		t.Fatal("expected the follow-up failure from the next profile")
+	}
+	if exhaustCalls != 1 {
+		t.Fatalf("exhaust calls=%d want 1", exhaustCalls)
+	}
+	// The escalation inside call 3 already replayed once on the same profile
+	// (streak 1); the fake current() keeps reporting profile 7, so the streak
+	// key is unchanged. One more failure reaches streak 2, and the NEXT one
+	// fires the second escalation, which the fake exhaust answers with a fully
+	// exhausted chain.
+	_, _, _, err = completeTaskLLM(context.Background(), "1", llm.CompletionRequest{}, hooks)
+	if err == nil {
+		t.Fatal("expected raw failure at streak 2")
+	}
+	if exhaustCalls != 1 {
+		t.Fatalf("exhaust calls=%d want 1 at streak 2", exhaustCalls)
+	}
+	_, _, _, err = completeTaskLLM(context.Background(), "1", llm.CompletionRequest{}, hooks)
+	if !isTaskLLMChainExhausted(err) {
+		t.Fatalf("expected chain-exhausted classification, got %v", err)
+	}
+	if exhaustCalls != 2 {
+		t.Fatalf("exhaust calls=%d want 2", exhaustCalls)
+	}
+}
+

server/task_llm_rotation_test.go

@@ -0,0 +1,32 @@
+package server
+
+import (
+	"context"
+	"testing"
+)
+
+func TestLLMProfilePinRoundTrip(t *testing.T) {
+	ctx := context.Background()
+	if _, ok := llmProfilePinFrom(ctx); ok {
+		t.Fatal("empty context must carry no pin")
+	}
+	pinned := withLLMProfilePin(ctx, 42)
+	got, ok := llmProfilePinFrom(pinned)
+	if !ok || got != 42 {
+		t.Fatalf("pin roundtrip got=%d ok=%v", got, ok)
+	}
+}
+
+func TestNextRotatedProfileRoundRobin(t *testing.T) {
+	tk := &Task{LLMProfileIDs: []int64{3, 1, 2}}
+	want := []int64{3, 1, 2, 3, 1}
+	for i, w := range want {
+		got, ok := tk.nextRotatedProfile()
+		if !ok || got != w {
+			t.Fatalf("rotation %d: got=%d ok=%v want=%d", i, got, ok, w)
+		}
+	}
+	if _, ok := (&Task{}).nextRotatedProfile(); ok {
+		t.Fatal("empty chain must not rotate")
+	}
+}

server/task_llm_test.go

@@ -246,7 +246,7 @@ func TestTaskLLMStreamRetriesPreStreamQuotaOnNextProfile(t *testing.T) {
 	active := int64(11)
 	transitions := 0
 	hooks := taskLLMStreamHooks{
-		current: func() (taskLLMSelection, error) {
+		current: func(ctx context.Context) (taskLLMSelection, error) {
 			if active == 11 {
 				return taskLLMSelection{profileID: 11, provider: first}, nil
 			}
@@ -288,7 +288,7 @@ func TestTaskLLMStreamRetriesAfterNonContentStartEvent(t *testing.T) {
 	}}
 	active := int64(11)
 	hooks := taskLLMStreamHooks{
-		current: func() (taskLLMSelection, error) {
+		current: func(ctx context.Context) (taskLLMSelection, error) {
 			if active == 11 {
 				return taskLLMSelection{profileID: 11, provider: first}, nil
 			}
@@ -319,7 +319,7 @@ func TestTaskLLMStreamRetriesStalePreStreamFailureAgainstReplacementChain(t *tes
 	second := &scriptedLLMProvider{events: []llm.StreamEvent{{Type: llm.SETextDelta, Text: "replacement"}}}
 	current := first
 	hooks := taskLLMStreamHooks{
-		current: func() (taskLLMSelection, error) {
+		current: func(ctx context.Context) (taskLLMSelection, error) {
 			profileID := int64(11)
 			if current == second {
 				profileID = 22
@@ -350,10 +350,11 @@ func TestTaskLLMStreamDoesNotSwitchForOrdinaryErrors(t *testing.T) {
 	}
 	for _, message := range tests {
 		t.Run(message, func(t *testing.T) {
+			resetLLMFailStreaksForTest() // package-global streaks must not leak across subtests
 			provider := &scriptedLLMProvider{err: errors.New(message)}
 			exhausted := 0
 			hooks := taskLLMStreamHooks{
-				current: func() (taskLLMSelection, error) {
+				current: func(ctx context.Context) (taskLLMSelection, error) {
 					return taskLLMSelection{profileID: 11, provider: provider}, nil
 				},
 				exhaust: func(taskLLMSelection, error) (db.TaskLLMTransition, error) {
@@ -382,7 +383,7 @@ func TestTaskLLMStreamAdvancesAfterMidStreamQuotaWithoutReplay(t *testing.T) {
 	second := &scriptedLLMProvider{events: []llm.StreamEvent{{Type: llm.SETextDelta, Text: "next-call"}}}
 	active := int64(11)
 	hooks := taskLLMStreamHooks{
-		current: func() (taskLLMSelection, error) {
+		current: func(ctx context.Context) (taskLLMSelection, error) {
 			if active == 11 {
 				return taskLLMSelection{profileID: 11, provider: first}, nil
 			}
@@ -418,7 +419,7 @@ func TestTaskLLMStreamMarkFailureAfterToolStartSkipsWorkerReplay(t *testing.T) {
 		err:    quota,
 	}
 	hooks := taskLLMStreamHooks{
-		current: func() (taskLLMSelection, error) {
+		current: func(ctx context.Context) (taskLLMSelection, error) {
 			return taskLLMSelection{profileID: 11, provider: provider}, nil
 		},
 		exhaust: func(taskLLMSelection, error) (db.TaskLLMTransition, error) {
@@ -467,7 +468,7 @@ func TestTaskLLMStreamRetriesTransientPreCommitOnSameProvider(t *testing.T) {
 	}
 	exhausted := 0
 	hooks := taskLLMStreamHooks{
-		current: func() (taskLLMSelection, error) {
+		current: func(ctx context.Context) (taskLLMSelection, error) {
 			return taskLLMSelection{profileID: 11, provider: provider}, nil
 		},
 		exhaust: func(taskLLMSelection, error) (db.TaskLLMTransition, error) {
@@ -499,7 +500,7 @@ func TestTaskLLMStreamSurfacesTransientAfterRetriesExhausted(t *testing.T) {
 		failErr:   errors.New("dial tcp: connection reset by peer"),
 	}
 	hooks := taskLLMStreamHooks{
-		current: func() (taskLLMSelection, error) {
+		current: func(ctx context.Context) (taskLLMSelection, error) {
 			return taskLLMSelection{profileID: 11, provider: provider}, nil
 		},
 		exhaust: func(taskLLMSelection, error) (db.TaskLLMTransition, error) {
@@ -524,7 +525,7 @@ func TestTaskLLMStreamDoesNotRetryDeterministicRejection(t *testing.T) {
 		failErr:   errors.New("anthropic: status 400: messages.1: invalid request"),
 	}
 	hooks := taskLLMStreamHooks{
-		current: func() (taskLLMSelection, error) {
+		current: func(ctx context.Context) (taskLLMSelection, error) {
 			return taskLLMSelection{profileID: 11, provider: provider}, nil
 		},
 		exhaust: func(taskLLMSelection, error) (db.TaskLLMTransition, error) {
@@ -620,7 +621,7 @@ func TestTaskLLMStreamStopsWhenChainExhausted(t *testing.T) {
 	provider := &scriptedLLMProvider{err: errors.New("billing_not_active")}
 	transitions := 0
 	hooks := taskLLMStreamHooks{
-		current: func() (taskLLMSelection, error) {
+		current: func(ctx context.Context) (taskLLMSelection, error) {
 			return taskLLMSelection{profileID: 11, provider: provider}, nil
 		},
 		exhaust: func(taskLLMSelection, error) (db.TaskLLMTransition, error) {

web/src/app/(main)/function/tasks/detail/_tabs/sessions-tab.tsx

@@ -478,7 +478,7 @@ function WorkerAssetBadge({ assets }: { assets: IntentAsset[] }) {
   );
 }
 
-export function SessionsTab({ taskId }: { taskId: string }) {
+export function SessionsTab({ taskId, taskPaused = false }: { taskId: string; taskPaused?: boolean }) {
   const approvalFocus = useApprovalFocus({ taskId });
   const [selectedSessionId, setActiveId] = React.useState(MAIN_ID);
   const focusSession = React.useMemo<Session | undefined>(() => {
@@ -1416,6 +1416,9 @@ export function SessionsTab({ taskId }: { taskId: string }) {
 
   // Run duration = span from this session's first step to its last (for a live
   // session, "now" so it ticks up — the 1.5s setTick above re-renders it).
+  // While the TASK is paused nothing time-derived may advance (#84): sessions
+  // keep state='running' in the store during a pause, so live alone would keep
+  // ticking and misrepresent the frozen budget. Freeze at the last step instead.
   const runDuration = React.useMemo(() => {
     let min = Infinity,
       max = 0;
@@ -1426,9 +1429,9 @@ export function SessionsTab({ taskId }: { taskId: string }) {
       if (t > max) max = t;
     }
     if (!Number.isFinite(min) || max === 0) return null;
-    const end = active.live ? Date.now() : max;
+    const end = active.live && !taskPaused ? Date.now() : max;
     return Math.max(0, end - min);
-  }, [activity, active.live]);
+  }, [activity, active.live, taskPaused]);
 
   // ---- transcript auto-scroll (open → bottom; stick to bottom unless scrolled up) ----
   const contentRef = React.useRef<HTMLDivElement | null>(null);

web/src/app/(main)/function/tasks/detail/page.tsx

@@ -420,7 +420,7 @@ function TaskDetailInner() {
       {/* Tab content */}
       <div className="flex-1 p-4 lg:p-6">
         <TabsContent value="sessions" className="mt-0">
-          <SessionsTab taskId={id} />
+          <SessionsTab taskId={id} taskPaused={task?.status === "paused"} />
         </TabsContent>
         <TabsContent value="overview" className="mt-0">
           <OverviewTab taskId={id} />