feat(advisor): P1 outlets — findings push to SSE + MCP scheduler_advisor_findings

The advisor's findings previously lived only in its JSONL decision log on
its own host. P1 adds two observe-only outlets so diagnoses reach humans
and other agents live:

apiserver side:
- backend/broker.go: EventAdvisorFinding ("advisor.finding") event type +
  AdvisorFinding struct + Event.AdvisorFindings field.
- controllers/events.go: ReportFindings — POST /xce/v1/advisor/findings
  (generated_at/scenario/findings), publishes one broker event. Broadcast
  only: the decision log remains the store of record.
- router/routers.go: route behind RequirePermission("advisor:report").
- middleware/whitelist.go: admit /xce/v1/advisor/findings in the POST URL
  whitelist (the URL middleware 403s before the router's RBAC otherwise —
  same shape as bug-1489).
- config/rbac.yaml: advisor_agent role (advisor:report + the six *:view
  grants the snapshot needs) and advisor:report in the permission
  whitelist. mcp_readonly stays strictly *:view. New
  TestRBACYAML_AdvisorAgentRole_PinsReportSurface pins the role surface.

advisor side:
- internal/push: Pusher — best-effort POST of each round's findings, one
  retry, nil/unconfigured/empty findings are no-ops; failures log and
  never affect the observe loop.
- internal/agent: SetPusher wired after the decision-log append (the log
  is authoritative; push is a fan-out).
- internal/config + cmd: push default-on to the same XCE_API_URL with the
  same XCE_API_KEY (the key's role needs advisor:report); ADVISOR_PUSH=
  off|false|0 disables.

mcp-server side:
- internal/advisor.go: lenient JSONL reader for the advisor's decision
  log (ADVISOR_LOG env), tail-bounded, corrupt lines skipped.
- scheduler_advisor_findings tool: recent records with parsed findings;
  friendly errors when the log is missing/unconfigured. Passive reader —
  never talks to the advisor process.

Cluster verification (t1/t2/t3): redeployed apiserver closure, deployed
rbac.yaml, issued advisor_agent key. Forced-PEND job → pend_stall fired
on the real LLM (glm-5-3-flash, 19.4s, degraded=false) → 2 findings, all
evidence verbatim in the snapshot, unverified=false → SSE broadcast
captured on /xce/v1/events → MCP tool returned both records. Second PEND
job joining the stalled set fired a new decision within one round
(change-triggered semantics). Verification jobs killed; sanity job
submitted→ran→DONE; all daemons and 3 hosts healthy.

Design doc: architecture diagram gains the push branch; roadmap P1
marked done (CLI outlet not built).

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

src/docs/advisor-design.md

@@ -21,7 +21,7 @@
  apiserver REST ──────▶│ snapshot: 紧凑集群状态（字段白名单+top-N 截断）  │
  (schedulerclient)     └──────────────┬───────────────────────────────┘
                                       ▼
- SSE /xce/v1/events ─▶ detect: 确定性触发器（阈值→冷却→去重）
+ SSE /xce/v1/events ─▶ detect: 确定性触发器（去重；变化即触发）
  (事件只提前扫描)              │ 命中
                               ▼
                       provider.Evaluate(Scenario, State, Runbook)
@@ -31,13 +31,22 @@
                               ▼
                        幻觉防火墙 VerifyEvidence
                               │
-                              ▼
-                       declog: JSONL 决策日志（全上下文）
+                ┌─────────────┼──────────────────┐
+                ▼             ▼                  ▼
+        declog JSONL     push（P1）       （未来：CLI 直出）
+        决策日志     ─▶ POST /xce/v1/advisor/findings
+        （存档/审计）     └─ apiserver broker ─▶ SSE advisor.finding 事件
 ```
 
 - 代码位置：`src/frontend/advisor/`，独立 Go module（`replace schedulerclient`），
   `CGO_ENABLED=0`，不进 meson（与 mcp-server 一致，`go build` 直出）。
-- 部署：t1 上以 `XCE_API_KEY`（mcp_readonly 角色）运行 `advisor run`。
+- 部署：t1 上以 `XCE_API_KEY`（advisor_agent 角色）运行 `advisor run`。
+- 出口（P1）：决策日志是存档底账；push 是尽力而为的广播（失败只记日志，
+  永不影响观察循环），目标 = 同一 apiserver 的
+  `POST /xce/v1/advisor/findings`（`advisor:report` 权限），服务端原样
+  转成 SSE `advisor.finding` 事件分发给订阅者。`ADVISOR_PUSH=off` 关闭。
+  MCP 侧 `scheduler_advisor_findings` 工具直接读决策日志 JSONL
+  （`ADVISOR_LOG`），是被动读者、不经过 advisor 进程。
 
 ## 3. Provider 契约
 
@@ -108,7 +117,7 @@ tokens / findings / raw / state(+digest) / degraded`。JSONL、按大小轮转
 
 | 阶段 | 内容 |
 |---|---|
-| P1 | propose 档：findings 推送（SSE/MCP 工具/CLI）；runbook 上提 schedulerclient；评估 echo vs LLM 的对照数据 |
+| P1 ✅ | propose 档：findings 推送（SSE 广播 + MCP 工具，已实现；CLI 直出未做）；评估 echo vs LLM 的对照数据 |
 | P2 | act 档（仅当引入带校准置信度的 provider——见附录 A）：allowlist + 爆炸半径上限 + 审计 |
 | — | 自洽性伪置信度（k 次采样一致率）作为 act 档的非 Jev 替代——5 倍成本，只在关键判断用 |
 

src/frontend/advisor/cmd/main.go

@@ -25,6 +25,7 @@ import (
 	"advisor/internal/declog"
 	"advisor/internal/detect"
 	"advisor/internal/provider"
+	"advisor/internal/push"
 	"advisor/internal/runbook"
 
 	sclient "schedulerclient"
@@ -66,6 +67,7 @@ configuration (env):
   ADVISOR_INTERVAL     scan interval (default 120s)
   ADVISOR_LOG          decision log path (default /tmp/advisor-decisions.jsonl)
   ADVISOR_RUNBOOK      runbook YAML overriding the embedded one
+  ADVISOR_PUSH         findings push to the apiserver (default on; off/false/0 to disable)
 `)
 }
 
@@ -92,6 +94,18 @@ func pickProvider(cfg *config.Config) provider.Provider {
 	return provider.NewOpenAI(cfg.LLMBaseURL, cfg.LLMAPIKey, cfg.LLMModel, cfg.LLMTimeout)
 }
 
+// maybePusher wires the findings push (P1) when enabled. Push is best-effort
+// by design — a 403/timeout never affects the observe loop; the decision log
+// remains the store of record.
+func maybePusher(cfg *config.Config) *push.Pusher {
+	if !cfg.PushEnabled {
+		log.Printf("advisor: findings push disabled (ADVISOR_PUSH=off)")
+		return nil
+	}
+	log.Printf("advisor: findings push -> %s (same credential as read; role needs advisor:report)", cfg.APIURL)
+	return push.New(cfg.APIURL, cfg.APIKey)
+}
+
 func loadRunbook(cfg *config.Config) *runbook.Runbook {
 	rb, err := runbook.Load(cfg.RunbookPath)
 	if err != nil {
@@ -112,6 +126,7 @@ func cmdRun() {
 	a := agent.New(mustClient(cfg), detect.NewDetector(detect.DefaultLimits()),
 		pickProvider(cfg), dl, cfg.Interval)
 	a.RunbookText = rb.Render()
+	a.SetPusher(maybePusher(cfg))
 
 	ctx, stop := signal.NotifyContext(context.Background(), os.Interrupt, syscall.SIGTERM)
 	defer stop()

src/frontend/advisor/internal/agent/agent.go

@@ -17,6 +17,7 @@ import (
 	"advisor/internal/declog"
 	"advisor/internal/detect"
 	"advisor/internal/provider"
+	"advisor/internal/push"
 	"advisor/internal/snapshot"
 
 	sclient "schedulerclient"
@@ -30,6 +31,9 @@ type Agent struct {
 	primary  provider.Provider // may be nil: echo-only mode
 	echo     *provider.Echo
 	log      *declog.Log
+	// pusher reports findings to the apiserver's advisor endpoint (P1);
+	// may be nil (push disabled).
+	pusher *push.Pusher
 	// RunbookText is the rendered runbook context handed to providers.
 	RunbookText string
 	// ScanInterval mirrors cfg.Interval; SSE events only pull the next scan
@@ -41,7 +45,8 @@ type Agent struct {
 	lastExits []snapshot.ExitState
 }
 
-// New assembles an Agent. primary may be nil (echo-only).
+// New assembles an Agent. primary may be nil (echo-only); pusher may be nil
+// (findings stay in the decision log only).
 func New(client sclient.Client, det *detect.Detector, primary provider.Provider, dl *declog.Log, scanInterval time.Duration) *Agent {
 	return &Agent{
 		client:       client,
@@ -54,6 +59,9 @@ func New(client sclient.Client, det *detect.Detector, primary provider.Provider,
 	}
 }
 
+// SetPusher enables reporting each round's findings to the apiserver.
+func (a *Agent) SetPusher(p *push.Pusher) { a.pusher = p }
+
 // Run blocks until ctx is cancelled: ticker scans + SSE event pull.
 // Event types only accelerate the next scan (debounced); detection and
 // evaluation always run on the same periodic path.
@@ -212,6 +220,18 @@ func mustJSON(v any) json.RawMessage {
 	return b
 }
 
+// snapshotTime extracts the snapshot's generated_at from the state JSON for
+// the findings push; "unknown" on a malformed state never blocks the report.
+func snapshotTime(stateJSON []byte) string {
+	var st struct {
+		GeneratedAt string `json:"generated_at"`
+	}
+	if err := json.Unmarshal(stateJSON, &st); err != nil || st.GeneratedAt == "" {
+		return "unknown"
+	}
+	return st.GeneratedAt
+}
+
 // evaluateTrigger runs the provider chain for one trigger and logs it.
 func (a *Agent) evaluateTrigger(ctx context.Context, tr detect.Trigger, stateJSON []byte) {
 	req := provider.EvaluateRequest{
@@ -257,6 +277,13 @@ func (a *Agent) evaluateTrigger(ctx context.Context, tr detect.Trigger, stateJSO
 		log.Printf("advisor: decision log append failed: %v", err)
 	}
 
+	// P1 propose tier: report the round to the apiserver's SSE fan-out so
+	// dashboards/MCP clients see diagnoses live. Best-effort; the decision
+	// log above already holds the record.
+	if a.pusher != nil {
+		a.pusher.BestEffort(ctx, tr.Scenario, snapshotTime(stateJSON), resp.Findings)
+	}
+
 	log.Printf("advisor: %s [%s] -> %d finding(s) (model=%s, %s, degraded=%v)",
 		tr.Scenario, tr.Fingerprint, len(resp.Findings), resp.Model, resp.Latency.Round(time.Millisecond), degraded)
 }

src/frontend/advisor/internal/config/config.go

@@ -41,6 +41,11 @@ type Config struct {
 	// RunbookPath optionally overrides the embedded runbook
 	// (ADVISOR_RUNBOOK).
 	RunbookPath string
+	// PushEnabled reports each round's findings to the apiserver's advisor
+	// endpoint (SSE fan-out, P1). On by default; ADVISOR_PUSH=off disables.
+	// The push target is APIURL with the same XCE_API_KEY — the key's role
+	// needs advisor:report.
+	PushEnabled bool
 }
 
 // Defaults.
@@ -55,6 +60,12 @@ const (
 // missing required values (XCE_API_KEY). LLM settings being empty is valid —
 // the advisor degrades to the echo provider.
 func FromEnv() (*Config, error) {
+	pushEnabled := true
+	switch os.Getenv("ADVISOR_PUSH") {
+	case "off", "false", "0":
+		pushEnabled = false
+	}
+
 	c := &Config{
 		APIURL:      getenv("XCE_API_URL", defaultAPIURL),
 		APIKey:      os.Getenv("XCE_API_KEY"),
@@ -65,6 +76,7 @@ func FromEnv() (*Config, error) {
 		RunbookPath: getenv("ADVISOR_RUNBOOK", ""),
 		Interval:    defaultInterval,
 		LLMTimeout:  defaultLLMTimeout,
+		PushEnabled: pushEnabled,
 	}
 
 	if c.APIKey == "" {

src/frontend/advisor/internal/push/push.go

@@ -0,0 +1,99 @@
+// Package push reports evaluation rounds' findings to the apiserver's
+// advisor endpoint, which fans them out to SSE subscribers as
+// advisor.finding events (P1 propose tier). Best-effort by design: a push
+// failure is logged and otherwise swallowed — the JSONL decision log on
+// this host is the store of record, and a monitoring loop must never die
+// because its telemetry is unreachable.
+package push
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"log"
+	"net/http"
+	"time"
+
+	"advisor/internal/provider"
+)
+
+// Pusher posts one round's findings to the apiserver.
+type Pusher struct {
+	baseURL string
+	token   string
+	hc      *http.Client
+}
+
+// New returns a Pusher for the apiserver at baseURL (e.g.
+// http://t1:8080) authenticated with a token (JWT or xce_ak_ key) whose
+// role grants advisor:report.
+func New(baseURL, token string) *Pusher {
+	return &Pusher{
+		baseURL: baseURL,
+		token:   token,
+		hc:      &http.Client{Timeout: 10 * time.Second},
+	}
+}
+
+// Report uploads one evaluation round. scenario names the trigger's
+// scenario; generatedAt is the snapshot time the findings evaluate against;
+// degraded marks the round as answered by the echo fallback.
+func (p *Pusher) Report(ctx context.Context, scenario, generatedAt string, findings []provider.Finding) error {
+	if p == nil || p.baseURL == "" {
+		return nil // push not configured: a no-op, not an error
+	}
+	if len(findings) == 0 {
+		return nil
+	}
+	body, err := json.Marshal(map[string]any{
+		"generated_at": generatedAt,
+		"scenario":     scenario,
+		"findings":     findings,
+	})
+	if err != nil {
+		return fmt.Errorf("marshal findings: %w", err)
+	}
+
+	// One retry: the apiserver may be restarting (or momentarily slow). A
+	// second failure is final for this round — the decision log still has
+	// the record, and the next diagnosis re-reports.
+	var lastErr error
+	for attempt := 0; attempt < 2; attempt++ {
+		req, err := http.NewRequestWithContext(ctx, http.MethodPost,
+			p.baseURL+"/xce/v1/advisor/findings", bytes.NewReader(body))
+		if err != nil {
+			return fmt.Errorf("build request: %w", err)
+		}
+		req.Header.Set("Authorization", p.token)
+		req.Header.Set("Content-Type", "application/json")
+
+		resp, err := p.hc.Do(req)
+		if err != nil {
+			lastErr = err
+		} else if resp.StatusCode == http.StatusOK {
+			resp.Body.Close()
+			return nil
+		} else {
+			lastErr = fmt.Errorf("HTTP %d", resp.StatusCode)
+			resp.Body.Close()
+		}
+		if attempt == 0 {
+			select {
+			case <-ctx.Done():
+				return ctx.Err()
+			case <-time.After(time.Second):
+			}
+		}
+	}
+	return fmt.Errorf("report findings: %w", lastErr)
+}
+
+// BestEffort wraps Report, logging failures instead of returning them. The
+// agent loop calls this — an unreportable finding is a degraded monitoring
+// setup, never a reason to stop observing.
+func (p *Pusher) BestEffort(ctx context.Context, scenario, generatedAt string, findings []provider.Finding) {
+	if err := p.Report(ctx, scenario, generatedAt, findings); err != nil {
+		log.Printf("advisor: findings push failed (non-fatal): %v", err)
+	}
+}

src/frontend/advisor/internal/push/push_test.go

@@ -0,0 +1,99 @@
+package push
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"sync/atomic"
+	"testing"
+
+	"advisor/internal/provider"
+)
+
+func sampleFindings() []provider.Finding {
+	return []provider.Finding{
+		{Kind: "diagnosis", Severity: provider.SevDegraded, Summary: "1 job stalled",
+			Evidence: []string{"73937053183639552", "unknown(0)"}, Confidence: 1.0},
+	}
+}
+
+func TestReport_PostsShapeAndAuth(t *testing.T) {
+	var gotAuth, gotPath string
+	var body map[string]any
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotAuth = r.Header.Get("Authorization")
+		gotPath = r.URL.Path
+		_ = json.NewDecoder(r.Body).Decode(&body)
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"code":200,"msg":"Success","data":{"published":1}}`))
+	}))
+	defer srv.Close()
+
+	p := New(srv.URL, "xce_ak_test")
+	if err := p.Report(context.Background(), "pend_stall", "2026-10-19T08:00:00Z", sampleFindings()); err != nil {
+		t.Fatalf("Report: %v", err)
+	}
+	if gotPath != "/xce/v1/advisor/findings" {
+		t.Errorf("path = %q", gotPath)
+	}
+	if gotAuth != "xce_ak_test" {
+		t.Errorf("auth = %q", gotAuth)
+	}
+	if body["scenario"] != "pend_stall" {
+		t.Errorf("scenario = %v", body["scenario"])
+	}
+	fs, ok := body["findings"].([]any)
+	if !ok || len(fs) != 1 {
+		t.Fatalf("findings = %v", body["findings"])
+	}
+}
+
+func TestReport_RetriesOnceThenSucceeds(t *testing.T) {
+	var calls atomic.Int32
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if calls.Add(1) == 1 {
+			w.WriteHeader(http.StatusBadGateway)
+			return
+		}
+		_, _ = w.Write([]byte(`{"code":200,"msg":"Success","data":{}}`))
+	}))
+	defer srv.Close()
+
+	p := New(srv.URL, "tok")
+	if err := p.Report(context.Background(), "pend_stall", "now", sampleFindings()); err != nil {
+		t.Fatalf("retry after 502 should succeed: %v", err)
+	}
+	if got := calls.Load(); got != 2 {
+		t.Fatalf("calls = %d, want 2", got)
+	}
+}
+
+func TestReport_GivesUpAfterTwoFailures(t *testing.T) {
+	var calls atomic.Int32
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		calls.Add(1)
+		w.WriteHeader(http.StatusServiceUnavailable)
+	}))
+	defer srv.Close()
+
+	p := New(srv.URL, "tok")
+	if err := p.Report(context.Background(), "pend_stall", "now", sampleFindings()); err == nil {
+		t.Fatal("expected error after both attempts failed")
+	}
+	if got := calls.Load(); got != 2 {
+		t.Fatalf("calls = %d, want exactly 2 (one retry, no loop)", got)
+	}
+}
+
+func TestReport_EmptyFindingsAndUnconfiguredAreNoOps(t *testing.T) {
+	// nil pusher and empty findings must not touch the network.
+	p := New("", "tok")
+	if err := p.Report(context.Background(), "s", "now", nil); err != nil {
+		t.Fatalf("empty findings: %v", err)
+	}
+	var q *Pusher
+	if err := q.Report(context.Background(), "s", "now", sampleFindings()); err != nil {
+		t.Fatalf("nil pusher: %v", err)
+	}
+}

src/frontend/apiserver/backend/broker.go

@@ -12,16 +12,38 @@ const (
 	EventJobUpdated  = "job.updated"
 	EventJobDeleted  = "job.deleted"
 	EventJobFinished = "job.finished"
+	// EventAdvisorFinding is one scheduler-advisor diagnosis (P1 propose
+	// tier). Carried in AdvisorFindings; the advisor's JSONL decision log on
+	// its own host remains the store of record — the broker only fans the
+	// finding out live.
+	EventAdvisorFinding = "advisor.finding"
 )
 
-// Event is a single lifecycle event. Exactly one of Job / (JobID, ExitCode)
-// is populated depending on Type.
+// AdvisorFinding is the advisor-facing shape of one diagnosis. Deliberately
+// decoupled from the advisor module (separate go modules); the advisor's
+// provider.Finding marshals to exactly this shape (kind/severity/summary/
+// evidence/confidence/suggestion/unverified).
+type AdvisorFinding struct {
+	Kind       string   `json:"kind"`
+	Severity   string   `json:"severity"`
+	Summary    string   `json:"summary"`
+	Evidence   []string `json:"evidence,omitempty"`
+	Confidence float64  `json:"confidence,omitempty"`
+	Suggestion string   `json:"suggestion,omitempty"`
+	Unverified bool     `json:"unverified,omitempty"`
+}
+
+// Event is a single lifecycle event. Exactly one of Job / (JobID, ExitCode) /
+// AdvisorFindings is populated depending on Type.
 type Event struct {
 	Type     string          `json:"type"`
 	JobID    uint64          `json:"job_id,omitempty"`
 	ExitCode int             `json:"exit_code,omitempty"`
 	Status   string          `json:"job_status,omitempty"` // JobStateStr of the job (PEND/RUN/DONE/EXIT/...)
 	Job      *models.JobInfo `json:"job,omitempty"`
+	// AdvisorFindings carries the diagnoses of one advisor evaluation round
+	// (Type EventAdvisorFinding).
+	AdvisorFindings []AdvisorFinding `json:"advisor_findings,omitempty"`
 }
 
 // Broker is a thread-safe fan-out of lifecycle events to subscribed channels.

src/frontend/apiserver/config/rbac.yaml

@@ -18,6 +18,7 @@ permissions:
   - usergroup:view     # view usergroup info
   - resource:view      # view resource info
   - event:view    # subscribe to the live job-event SSE stream (scheduler-agent Phase 2)
+  - advisor:report     # advisor service account reports findings (broadcast-only; scheduler-advisor P1)
   - ar:view            # view advanced reservation info
   - ar:submit          # submit advanced reservation
   - ar:control         # delete, modify, create
@@ -211,3 +212,33 @@ roles:
       - action: event:view
         domain:
           - type: all
+
+  # advisor_agent: service-account role for the observe-only scheduler-advisor
+  # (P1 propose tier). Cluster visibility equal to mcp_readonly (the snapshot
+  # needs job/queue/host/event views) plus ONE write action, advisor:report —
+  # POST /xce/v1/advisor/findings, which only broadcasts diagnoses to SSE
+  # subscribers. No job:control/queue:control/cluster:control: the advisor is
+  # architecturally incapable of mutating cluster state.
+  advisor_agent:
+    permissions:
+      - action: advisor:report
+        domain:
+          - type: all
+      - action: job:view
+        domain:
+          - type: all
+      - action: queue:view
+        domain:
+          - type: all
+      - action: host:view
+        domain:
+          - type: all
+      - action: resource:view
+        domain:
+          - type: all
+      - action: cluster:view
+        domain:
+          - type: all
+      - action: event:view
+        domain:
+          - type: all

src/frontend/apiserver/controllers/events.go

@@ -8,6 +8,7 @@ import (
 	"github.com/gin-gonic/gin"
 
 	"apiserver/backend"
+	"apiserver/common"
 	logger "apiserver/utils"
 
 	"go.uber.org/zap"
@@ -106,3 +107,39 @@ func (e *EventsController) SubscribeJobs(ctx *gin.Context) {
 		}
 	}
 }
+
+// ReportFindings accepts one advisor evaluation round's diagnoses and fans
+// them out to SSE subscribers as advisor.finding events (P1 propose tier).
+// The advisor's JSONL decision log on its own host remains the store of
+// record; this endpoint only broadcasts. Auth: advisor:report (granted to
+// the advisor_agent service role; see rbac.yaml).
+func (e *EventsController) ReportFindings(ctx *gin.Context) {
+	var req struct {
+		// GeneratedAt is the snapshot time the findings were evaluated
+		// against (RFC3339); echoed to subscribers unchanged.
+		GeneratedAt string                   `json:"generated_at"`
+		Scenario    string                   `json:"scenario"`
+		Findings    []backend.AdvisorFinding `json:"findings"`
+	}
+	if err := ctx.ShouldBindJSON(&req); err != nil {
+		ReturnError(ctx, common.INVALID_PARAMS_CODE, "Invalid request format: "+err.Error(), nil)
+		return
+	}
+	if len(req.Findings) == 0 {
+		ReturnError(ctx, common.INVALID_PARAMS_CODE, "findings must be non-empty", nil)
+		return
+	}
+	if req.Scenario == "" {
+		ReturnError(ctx, common.INVALID_PARAMS_CODE, "scenario is required", nil)
+		return
+	}
+
+	backend.DefaultBroker.Publish(backend.Event{
+		Type:            backend.EventAdvisorFinding,
+		Status:          req.Scenario,
+		AdvisorFindings: req.Findings,
+	})
+	ReturnSuccess(ctx, common.SUCCESS_CODE, common.SUCCESS_MSG, map[string]interface{}{
+		"published": len(req.Findings),
+	})
+}

src/frontend/apiserver/controllers/rbac_mcp_readonly_test.go

@@ -50,3 +50,46 @@ func hasViewSuffix(action string) bool {
 	}
 	return action[len(action)-len(suffix):] == suffix
 }
+
+// TestRBACYAML_AdvisorAgentRole_PinsReportSurface verifies the advisor_agent
+// role (P1 propose tier): it exists, grants advisor:report (its ONE write
+// action — POST /xce/v1/advisor/findings), and everything else it grants is
+// a plain *:view. The advisor must never hold a control/submit permission.
+func TestRBACYAML_AdvisorAgentRole_PinsReportSurface(t *testing.T) {
+	resetRBACState()
+
+	path, err := filepath.Abs("../config/rbac.yaml")
+	if err != nil {
+		t.Fatalf("resolve rbac.yaml path: %v", err)
+	}
+	if _, err := os.Stat(path); err != nil {
+		t.Skipf("config/rbac.yaml not found at %s: %v", path, err)
+	}
+	if err := LoadRBACConfig(path); err != nil {
+		t.Fatalf("LoadRBACConfig: %v", err)
+	}
+
+	if !store.Roles.Has("advisor_agent") {
+		t.Fatal("expected advisor_agent role to be loaded from config/rbac.yaml")
+	}
+
+	perms := store.GetRolePermissions("advisor_agent")
+	if len(perms) == 0 {
+		t.Fatal("advisor_agent should have at least one permission")
+	}
+
+	reports := 0
+	for _, p := range perms {
+		switch {
+		case p.Action == "advisor:report":
+			reports++
+		case hasViewSuffix(p.Action):
+			// fine — read-only visibility
+		default:
+			t.Errorf("advisor_agent grants forbidden permission %q — the advisor is observe-only; only advisor:report and *:view are allowed", p.Action)
+		}
+	}
+	if reports != 1 {
+		t.Errorf("advisor_agent should grant exactly one advisor:report action, got %d", reports)
+	}
+}

src/frontend/apiserver/middleware/whitelist.go

@@ -48,6 +48,7 @@ var WhitelistPOSTURL = []string{
 	"/xce/v1/roles/create",   // create role
 	"/xce/v1/roles/*",        // add permissions for role
 	"/xce/v1/users/*",        // add roles for user
+	"/xce/v1/advisor/findings", // advisor reports findings (scheduler-advisor P1; RBAC advisor:report applies downstream)
 }
 
 var WhitelistDELETEURL = []string{

src/frontend/apiserver/middleware/whitelist_test.go

@@ -63,6 +63,11 @@ func TestPathMatchesPattern(t *testing.T) {
 		{"/xce/v1/roles/*", "/xce/v1/roles/cluster_admin", true},
 		{"/xce/v1/users/*", "/xce/v1/users/current", true},
 		{"/xce/v1/users/*", "/xce/v1/users/admin/basic_user", true},
+		// advisor findings (scheduler-advisor P1): the POST whitelist must
+		// admit the advisor's report path or the URL middleware 403s before
+		// the router (and its advisor:report check) is ever reached.
+		{"/xce/v1/advisor/findings", "/xce/v1/advisor/findings", true},
+		{"/xce/v1/advisor/findings", "/xce/v1/advisor/findings/extra", false},
 		// exact patterns
 		{"/xce/v1/hosts", "/xce/v1/hosts", true},
 		{"/xce/v1/hosts", "/xce/v1/host", false},

src/frontend/apiserver/router/routers.go

@@ -167,6 +167,17 @@ func Router() *gin.Engine {
 		eventsApi.GET("", controllers.RequirePermission("event:view"), eventsHandler.SubscribeJobs)
 	}
 
+	// ################ case: advisor route (P1 propose tier)
+	// The scheduler-advisor posts its evaluation round's findings here; the
+	// apiserver fans them out to SSE subscribers as advisor.finding events.
+	// Write-scoped to advisor:report (mcp_readonly) — the advisor's key can
+	// only report diagnoses, not read or mutate anything else.
+	advisorApi := r.Group("/xce/v1/advisor")
+	{
+		eventsHandler := &controllers.EventsController{}
+		advisorApi.POST("/findings", controllers.RequirePermission("advisor:report"), eventsHandler.ReportFindings)
+	}
+
 	// ################ case: resources route
 	resApi := r.Group("/xce/v1/resources")
 	{

src/frontend/mcp-server/cmd/main.go

@@ -170,6 +170,14 @@ func main() {
 		mcp.WithDestructiveHintAnnotation(false),
 	), app.HandlePendReasons)
 
+	// ── scheduler_advisor_findings (observe-mode advisor outlet) ──
+	s.AddTool(mcp.NewTool("scheduler_advisor_findings",
+		mcp.WithDescription("Recent findings from the XCE scheduler-advisor — an observe-only LLM ops copilot that periodically diagnoses cluster state (pend stalls, worker health). Read from the advisor's decision log; each record has ts, scenario, provider, and structured findings (kind/severity/summary/evidence/confidence/suggestion). Findings are advisory for a human — never executed. Requires the ADVISOR_LOG env var pointing at the advisor's decision log."),
+		mcp.WithNumber("tail", mcp.Description("Max recent decision records to return (default 10, max 100)")),
+		mcp.WithReadOnlyHintAnnotation(true),
+		mcp.WithDestructiveHintAnnotation(false),
+	), app.HandleAdvisorFindings)
+
 	// ── write tools (Phase 1) ──
 	// Write tools are annotated destructive so the MCP client (e.g. Claude
 	// Code) prompts the user for confirmation before executing them — the

src/frontend/mcp-server/internal/advisor.go

@@ -0,0 +1,86 @@
+// advisor.go reads the scheduler-advisor's JSONL decision log for the
+// scheduler_advisor_findings tool. The log is the advisor's store of record
+// (see src/frontend/advisor/internal/declog); the MCP tool is a passive
+// reader — it never talks to the advisor process.
+package internal
+
+import (
+	"bufio"
+	"encoding/json"
+	"fmt"
+	"os"
+	"strings"
+)
+
+// AdvisorRecord mirrors one declog.Record line. Parsed leniently: unknown
+// fields are ignored, missing fields render as empty.
+type AdvisorRecord struct {
+	TS        string          `json:"ts"`
+	Scenario  string          `json:"scenario"`
+	Trigger   string          `json:"trigger"`
+	Provider  string          `json:"provider"`
+	Model     string          `json:"model"`
+	LatencyMS int64           `json:"latency_ms"`
+	Findings  json.RawMessage `json:"findings"`
+	Degraded  bool            `json:"degraded,omitempty"`
+	Error     string          `json:"error,omitempty"`
+}
+
+// AdvisorFinding mirrors one provider.Finding inside the findings array.
+type AdvisorFinding struct {
+	Kind       string   `json:"kind"`
+	Severity   string   `json:"severity"`
+	Summary    string   `json:"summary"`
+	Evidence   []string `json:"evidence"`
+	Confidence float64  `json:"confidence"`
+	Suggestion string   `json:"suggestion,omitempty"`
+	Unverified bool     `json:"unverified,omitempty"`
+}
+
+// AdvisorFindingsOptions bounds the read.
+type AdvisorFindingsOptions struct {
+	// Path is the decision log path (ADVISOR_LOG).
+	Path string
+	// Tail is the max number of records returned (most recent last).
+	Tail int
+}
+
+// ReadAdvisorFindings reads up to Tail records from the advisor decision
+// log, returning them in file order (oldest first, newest last). It returns
+// a missing-file error sentinel the handler translates into a friendly
+// message, and skips corrupt lines rather than failing the whole read.
+func ReadAdvisorFindings(opts AdvisorFindingsOptions) ([]AdvisorRecord, error) {
+	if opts.Path == "" {
+		return nil, fmt.Errorf("advisor log path not configured (set ADVISOR_LOG for the mcp-server)")
+	}
+	f, err := os.Open(opts.Path)
+	if err != nil {
+		if os.IsNotExist(err) {
+			return nil, fmt.Errorf("advisor decision log not found at %s — the advisor may not be running (ADVISOR_LOG)", opts.Path)
+		}
+		return nil, fmt.Errorf("open advisor log: %w", err)
+	}
+	defer f.Close()
+
+	var recs []AdvisorRecord
+	sc := bufio.NewScanner(f)
+	sc.Buffer(make([]byte, 0, 64*1024), 8<<20) // decision records embed the full state snapshot; can be large
+	for sc.Scan() {
+		line := strings.TrimSpace(sc.Text())
+		if line == "" {
+			continue
+		}
+		var r AdvisorRecord
+		if err := json.Unmarshal([]byte(line), &r); err != nil {
+			continue // corrupt line — skip, keep reading
+		}
+		recs = append(recs, r)
+	}
+	if err := sc.Err(); err != nil {
+		return nil, fmt.Errorf("read advisor log: %w", err)
+	}
+	if opts.Tail > 0 && len(recs) > opts.Tail {
+		recs = recs[len(recs)-opts.Tail:]
+	}
+	return recs, nil
+}

src/frontend/mcp-server/internal/advisor_test.go

@@ -0,0 +1,105 @@
+package internal
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/mark3labs/mcp-go/mcp"
+)
+
+// textOf extracts the text of the first content element.
+func textOf(res *mcp.CallToolResult) string {
+	if res == nil || len(res.Content) == 0 {
+		return ""
+	}
+	tc, ok := res.Content[0].(mcp.TextContent)
+	if !ok {
+		return ""
+	}
+	return tc.Text
+}
+
+// writeAdvisorLog writes a small JSONL decision log and returns its path.
+func writeAdvisorLog(t *testing.T, lines ...string) string {
+	t.Helper()
+	p := filepath.Join(t.TempDir(), "advisor-decisions.jsonl")
+	if err := os.WriteFile(p, []byte(strings.Join(lines, "\n")+"\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	return p
+}
+
+func TestReadAdvisorFindings_ParsesTailSkipsCorrupt(t *testing.T) {
+	p := writeAdvisorLog(t,
+		`{"ts":"2026-01-01T00:00:00Z","scenario":"pend_stall","trigger":"fp-1","provider":"echo","latency_ms":1,"findings":[{"kind":"anomaly","severity":"severe","summary":"2 pending job(s) stalled","evidence":["job 101","no_resource"],"confidence":0.9,"suggestion":"check queue limits"}]}`,
+		`{corrupt line`,
+		`{"ts":"2026-01-01T00:01:00Z","scenario":"worker_health","trigger":"fp-2","provider":"openai","model":"glm-5-3-flash","latency_ms":24000,"findings":[{"kind":"diagnosis","severity":"degraded","summary":"worker w2 unreachable","evidence":["w2"],"confidence":0.7,"unverified":true}]}`,
+	)
+	recs, err := ReadAdvisorFindings(AdvisorFindingsOptions{Path: p, Tail: 10})
+	if err != nil {
+		t.Fatalf("read: %v", err)
+	}
+	if len(recs) != 2 {
+		t.Fatalf("want 2 records (corrupt line skipped), got %d", len(recs))
+	}
+	if recs[0].Scenario != "pend_stall" || recs[1].Scenario != "worker_health" {
+		t.Fatalf("scenario order wrong: %s, %s", recs[0].Scenario, recs[1].Scenario)
+	}
+	if recs[0].Provider != "echo" || recs[1].Model != "glm-5-3-flash" {
+		t.Fatalf("provider/model parse wrong: %+v", recs)
+	}
+	if recs[0].LatencyMS != 1 || recs[1].LatencyMS != 24000 {
+		t.Fatalf("latency parse wrong: %+v", recs)
+	}
+
+	// Tail bounds the result to the most recent records.
+	recs, err = ReadAdvisorFindings(AdvisorFindingsOptions{Path: p, Tail: 1})
+	if err != nil {
+		t.Fatalf("read tail=1: %v", err)
+	}
+	if len(recs) != 1 || recs[0].Scenario != "worker_health" {
+		t.Fatalf("tail=1 want only newest worker_health, got %+v", recs)
+	}
+}
+
+func TestReadAdvisorFindings_MissingLog(t *testing.T) {
+	_, err := ReadAdvisorFindings(AdvisorFindingsOptions{Path: "/nonexistent/log.jsonl"})
+	if err == nil || !strings.Contains(err.Error(), "not found") {
+		t.Fatalf("want friendly not-found error, got %v", err)
+	}
+	_, err = ReadAdvisorFindings(AdvisorFindingsOptions{Path: ""})
+	if err == nil || !strings.Contains(err.Error(), "not configured") {
+		t.Fatalf("want not-configured error, got %v", err)
+	}
+}
+
+func TestHandleAdvisorFindings_EmptyLogMeansQuietCluster(t *testing.T) {
+	p := writeAdvisorLog(t) // exists, zero records
+	app := &App{advisorLogPath: p}
+	res, err := app.HandleAdvisorFindings(t.Context(), mcp.CallToolRequest{})
+	if err != nil {
+		t.Fatalf("handler: %v", err)
+	}
+	if res == nil || len(res.Content) == 0 {
+		t.Fatal("want text content")
+	}
+	if txt := textOf(res); !strings.Contains(txt, "no evaluation records") {
+		t.Fatalf("want quiet-cluster message, got %q", txt)
+	}
+}
+
+func TestHandleAdvisorFindings_LogMissingReportsFriendlyError(t *testing.T) {
+	app := &App{advisorLogPath: "/nonexistent/log.jsonl"}
+	res, err := app.HandleAdvisorFindings(t.Context(), mcp.CallToolRequest{})
+	if err != nil {
+		t.Fatalf("handler: %v", err)
+	}
+	if res == nil || !res.IsError {
+		t.Fatal("want tool error result")
+	}
+	if txt := textOf(res); !strings.Contains(txt, "not found") {
+		t.Fatalf("want not-found message, got %q", txt)
+	}
+}

src/frontend/mcp-server/internal/app.go

@@ -5,6 +5,7 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"os"
 	"strconv"
 	"time"
 
@@ -16,12 +17,15 @@ import (
 type App struct {
 	client    sclient.Client
 	templates *TemplateRegistry
+	// advisorLogPath is the scheduler-advisor's decision log (ADVISOR_LOG,
+	// empty disables the scheduler_advisor_findings tool's data source).
+	advisorLogPath string
 }
 
 // NewApp creates a new App with a schedulerclient connected to the given base URL.
 func NewApp(baseURL, token string) (*App, error) {
 	hc := sclient.NewClient(baseURL, token)
-	return &App{client: hc, templates: NewTemplateRegistry()}, nil
+	return &App{client: hc, templates: NewTemplateRegistry(), advisorLogPath: os.Getenv("ADVISOR_LOG")}, nil
 }
 
 // ── helpers ─────────────────────────────────────────────────────────────────
@@ -552,3 +556,57 @@ func (a *App) HandleEstimateWaitTime(ctx context.Context, req mcp.CallToolReques
 		"estimate_basis":        "Rough heuristic from current pending count and free cores; not a guarantee.",
 	})
 }
+
+// HandleAdvisorFindings serves the scheduler_advisor_findings tool: the
+// scheduler-advisor's recent findings, read from its JSONL decision log
+// (ADVISOR_LOG). Read-only and passive — the log is the store of record.
+func (a *App) HandleAdvisorFindings(ctx context.Context, req mcp.CallToolRequest) (*mcp.CallToolResult, error) {
+	tail := int(getNumber(req, "tail"))
+	if tail <= 0 {
+		tail = 10
+	}
+	if tail > 100 {
+		tail = 100
+	}
+	recs, err := ReadAdvisorFindings(AdvisorFindingsOptions{Path: a.advisorLogPath, Tail: tail})
+	if err != nil {
+		return mcp.NewToolResultError(err.Error()), nil
+	}
+	if len(recs) == 0 {
+		return textResultStr("The advisor decision log exists but has no evaluation records yet — the advisor has not fired any detectors since it started (a healthy cluster produces no decisions)."), nil
+	}
+
+	// Compact projection: one summary block per record; the raw decision
+	// log stays on disk for full detail.
+	type recordOut struct {
+		TS        string           `json:"ts"`
+		Scenario  string           `json:"scenario"`
+		Provider  string           `json:"provider"`
+		Model     string           `json:"model,omitempty"`
+		LatencyMS int64            `json:"latency_ms,omitempty"`
+		Degraded  bool             `json:"degraded,omitempty"`
+		Error     string           `json:"error,omitempty"`
+		Findings  []AdvisorFinding `json:"findings"`
+	}
+	out := make([]recordOut, 0, len(recs))
+	for _, r := range recs {
+		o := recordOut{
+			TS:        r.TS,
+			Scenario:  r.Scenario,
+			Provider:  r.Provider,
+			Model:     r.Model,
+			LatencyMS: r.LatencyMS,
+			Degraded:  r.Degraded,
+			Error:     r.Error,
+		}
+		if len(r.Findings) > 0 {
+			_ = json.Unmarshal(r.Findings, &o.Findings) // findings are advisory text; a malformed array renders as null
+		}
+		out = append(out, o)
+	}
+	return textResult(map[string]any{
+		"log_path": a.advisorLogPath,
+		"records":  out,
+		"note":     "Findings are observe-only diagnostics from the scheduler-advisor; suggestions are advisory for a human and never executed. unverified=true means the finding cited evidence not present in the cluster state (downgrade trust accordingly).",
+	})
+}