Merge branch 'dev' into feature/workspace-star

6 files changed, 346 insertions, 1 deletions

diff --git a/packages/session-orchestrator/src/extension.ts b/packages/session-orchestrator/src/extension.ts
index 783d894..777feaa 100644
--- a/packages/session-orchestrator/src/extension.ts
+++ b/packages/session-orchestrator/src/extension.ts
@@ -65,6 +65,19 @@ export function activate(host: HostAPI): void {
     logger: host.logger,
     now: () => Date.now(),
     emit: (hook, payload) => host.emit(hook, payload),
+    // Injected process.memoryUsage() sampler — the production edge. Tests
+    // inject a fake to assert per-turn before/after telemetry. Wired in the
+    // shell (like `now: () => Date.now()`); pure decision logic is untouched.
+    sampleMemory: () => {
+      const m = process.memoryUsage();
+      return {
+        rss: m.rss,
+        heapUsed: m.heapUsed,
+        heapTotal: m.heapTotal,
+        external: m.external,
+        arrayBuffers: m.arrayBuffers,
+      };
+    },
     resolveQueue: () => {
       // Lazily resolve the message-queue service. Returns undefined when the
       // extension isn't loaded (feature degrades off) — checked via the
diff --git a/packages/session-orchestrator/src/index.ts b/packages/session-orchestrator/src/index.ts
index cc41fa8..3d3e816 100644
--- a/packages/session-orchestrator/src/index.ts
+++ b/packages/session-orchestrator/src/index.ts
@@ -39,6 +39,9 @@ export {
   defaultDispatchPolicy,
   delayFor,
   generateTurnId,
+  type MemorySample,
+  memoryDelta,
+  memorySampleAttributes,
   RETRY_BUDGET_MS,
   RETRY_SCHEDULE_MS,
   RETRY_TAIL_MS,
diff --git a/packages/session-orchestrator/src/orchestrator.test.ts b/packages/session-orchestrator/src/orchestrator.test.ts
index 076ad51..400ca0b 100644
--- a/packages/session-orchestrator/src/orchestrator.test.ts
+++ b/packages/session-orchestrator/src/orchestrator.test.ts
@@ -4,7 +4,10 @@ import type {
   AgentEvent,
   ChatMessage,
   EventHookDescriptor,
+  LogDeps,
   Logger,
+  LogRecord,
+  LogSink,
   ProviderContract,
   ProviderEvent,
   ProviderStreamOptions,
@@ -15,7 +18,7 @@ import type {
   ToolContract,
   TurnMetrics,
 } from "@dispatch/kernel";
-import { runTurn } from "@dispatch/kernel";
+import { createLogger, runTurn } from "@dispatch/kernel";
 import type { SystemPromptService } from "@dispatch/system-prompt";
 import { describe, expect, it } from "vitest";
 import {
@@ -27,6 +30,7 @@ import {
   type TurnLifecyclePayload,
   type WarmCompletedPayload,
 } from "./orchestrator.js";
+import type { MemorySample } from "./pure.js";
 import type { ToolAssembly } from "./tools-filter.js";
 
 function createInMemoryStore(): ConversationStore & {
@@ -3980,3 +3984,135 @@ describe("system prompt: compaction flow", () => {
     expect(capturedSystemPrompt?.startsWith("RECONSTRUCTED")).toBe(false);
   });
 });
+
+describe("per-turn memory telemetry", () => {
+  function capturingLogger(): { logger: Logger; records: LogRecord[] } {
+    let id = 0;
+    const deps: LogDeps = { now: () => 1000 + id++, newId: () => `id-${id++}` };
+    const records: LogRecord[] = [];
+    const sink: LogSink = { emit: (r) => records.push(r) };
+    return { logger: createLogger({ extensionId: "session-orchestrator" }, sink, deps), records };
+  }
+
+  it("logs before/after samples around the stream, tagged with conversationId + turnId", async () => {
+    const store = createInMemoryStore();
+    const provider: ProviderContract = { id: "p", stream: async function* () {} };
+    const { captured, captureRunTurn } = createCapturingRunTurn();
+    const { logger, records } = capturingLogger();
+
+    const samples: MemorySample[] = [
+      { rss: 100, heapUsed: 10, heapTotal: 20, external: 1, arrayBuffers: 0 },
+      { rss: 250, heapUsed: 30, heapTotal: 20, external: 1, arrayBuffers: 5 },
+    ];
+    let sampleIdx = 0;
+
+    const { orchestrator } = createSessionOrchestrator({
+      conversationStore: store,
+      resolveProvider: () => provider,
+      resolveTools: () => [],
+      applyToolsFilter: identityApplyToolsFilter,
+      runTurn: captureRunTurn,
+      logger,
+      sampleMemory: () => samples[sampleIdx++] as MemorySample,
+    });
+
+    await orchestrator.handleMessage({
+      conversationId: "conv-mem",
+      text: "hi",
+      onEvent: () => {},
+    });
+
+    expect(captured).toHaveLength(1);
+    const beforeLogs = records.filter((r) => r.kind === "log" && r.msg === "memory:turn:before");
+    const afterLogs = records.filter((r) => r.kind === "log" && r.msg === "memory:turn:after");
+    expect(beforeLogs).toHaveLength(1);
+    expect(afterLogs).toHaveLength(1);
+
+    // Both samples carry the turn's conversationId + turnId correlation.
+    const turnId = captured[0]?.turnId;
+    expect(turnId).toMatch(/^turn-/);
+    const before = beforeLogs[0] as Extract<LogRecord, { kind: "log" }>;
+    const after = afterLogs[0] as Extract<LogRecord, { kind: "log" }>;
+    expect(before.conversationId).toBe("conv-mem");
+    expect(before.turnId).toBe(turnId);
+    expect(after.conversationId).toBe("conv-mem");
+    expect(after.turnId).toBe(turnId);
+
+    // Before sample carries absolute MB values.
+    expect(before.attributes?.rssMB).toBe(0); // 100 bytes rounds to 0 MB
+    // After sample carries absolute + delta (delta rss = 150 bytes → 0 MB).
+    expect(after.attributes?.rssMB).toBe(0);
+    // deltaRssMB is the rounded delta (150 bytes → 0 MB).
+    expect(after.attributes).toHaveProperty("deltaRssMB");
+  });
+
+  it("emits no memory logs when sampleMemory is not injected (degrades off)", async () => {
+    const store = createInMemoryStore();
+    const provider: ProviderContract = { id: "p", stream: async function* () {} };
+    const { captureRunTurn } = createCapturingRunTurn();
+    const { logger, records } = capturingLogger();
+
+    const { orchestrator } = createSessionOrchestrator({
+      conversationStore: store,
+      resolveProvider: () => provider,
+      resolveTools: () => [],
+      applyToolsFilter: identityApplyToolsFilter,
+      runTurn: captureRunTurn,
+      logger,
+      // sampleMemory intentionally omitted
+    });
+
+    await orchestrator.handleMessage({
+      conversationId: "conv-no-mem",
+      text: "hi",
+      onEvent: () => {},
+    });
+
+    const memLogs = records.filter(
+      (r) => r.kind === "log" && typeof r.msg === "string" && r.msg.startsWith("memory:turn"),
+    );
+    expect(memLogs).toHaveLength(0);
+  });
+
+  it("getActiveConversationCount tracks in-flight turns", async () => {
+    const store = createInMemoryStore();
+    const result: RunTurnResult = {
+      messages: [{ role: "assistant", chunks: [{ type: "text", text: "ok" }] }],
+      usage: { inputTokens: 1, outputTokens: 1 },
+      finishReason: "stop",
+    };
+    // A runTurn that blocks until the test releases it — keeps the turn
+    // active so getActiveConversationCount reflects an in-flight turn.
+    let release: () => void = () => {};
+    const blocked = new Promise<void>((resolve) => {
+      release = resolve;
+    });
+    const blockingRunTurn = async (_input: RunTurnInput): Promise<RunTurnResult> => {
+      await blocked;
+      return result;
+    };
+
+    const { orchestrator } = createSessionOrchestrator({
+      conversationStore: store,
+      resolveProvider: () => ({ id: "p", stream: async function* () {} }),
+      resolveTools: () => [],
+      applyToolsFilter: identityApplyToolsFilter,
+      runTurn: blockingRunTurn,
+    });
+
+    expect(orchestrator.getActiveConversationCount()).toBe(0);
+    const done = orchestrator.handleMessage({
+      conversationId: "conv-active",
+      text: "hi",
+      onEvent: () => {},
+    });
+    // Give the detached turn a tick to register as active.
+    await Promise.resolve();
+    await Promise.resolve();
+    expect(orchestrator.getActiveConversationCount()).toBe(1);
+
+    release();
+    await done;
+    expect(orchestrator.getActiveConversationCount()).toBe(0);
+  });
+});
diff --git a/packages/session-orchestrator/src/orchestrator.ts b/packages/session-orchestrator/src/orchestrator.ts
index 4c6a673..aaf418a 100644
--- a/packages/session-orchestrator/src/orchestrator.ts
+++ b/packages/session-orchestrator/src/orchestrator.ts
@@ -30,6 +30,9 @@ import {
   defaultDispatchPolicy,
   delayFor,
   generateTurnId,
+  type MemorySample,
+  memoryDelta,
+  memorySampleAttributes,
   resolveModelName,
   resolveReasoningEffort,
 } from "./pure.js";
@@ -331,6 +334,13 @@ export interface SessionOrchestrator {
   subscribe(conversationId: string, listener: TurnEventListener): () => void;
   isActive(conversationId: string): boolean;
   /**
+   * The number of conversations currently driving a turn (in the
+   * `activeConversations` set). Used by host-bin's periodic memory telemetry
+   * to tag each RSS sample with the active-conversation count, so growth can
+   * be attributed to the streaming/turn path vs an idle baseline.
+   */
+  getActiveConversationCount(): number;
+  /**
    * Explicitly close a conversation (the user closed its tab — distinct from a
    * socket disconnect, which never touches the turn): aborts any in-flight turn
    * (the kernel finishes with `finishReason: "aborted"`, partial messages are
@@ -431,6 +441,16 @@ export interface SessionOrchestratorDeps {
   readonly logger?: Logger;
   /** Injected monotonic-ish clock (ms) forwarded to RunTurnInput for timing events. */
   readonly now?: () => number;
+  /**
+   * Optional process.memoryUsage() sampler, injected for testability. When
+   * present, the orchestrator captures a sample immediately before and after
+   * each turn's stream completes (`deps.runTurn`) and logs the per-turn delta
+   * tagged with conversationId + turnId — correlating RSS growth with the
+   * streaming path (the prime leak suspect). When absent (undefined), no
+   * per-turn memory telemetry is emitted (feature degrades off cleanly).
+   * Pure decision logic stays unchanged; this is additive observability.
+   */
+  readonly sampleMemory?: () => MemorySample;
   /** Emit a lifecycle event hook to subscribers. Injected from host. */
   readonly emit?: <TPayload>(hook: EventHookDescriptor<TPayload>, payload: TPayload) => void;
 }
@@ -886,6 +906,18 @@ export function createSessionOrchestrator(
         // FE to syncTail during generation (CR-6).
         await deps.conversationStore.append(conversationId, [userMsg]);
 
+        // Per-turn memory telemetry: capture a sample immediately BEFORE the
+        // stream starts. Paired with the post-stream sample below, this
+        // measures the streaming path's memory footprint per turn (the prime
+        // leak suspect — AI-SDK streaming buffers + per-turn message arrays).
+        // Tagged with conversationId + turnId via the turnLogger's correlation
+        // context. Additive observability only — does not alter the stream.
+        const sampleMem = deps.sampleMemory;
+        const memBefore = sampleMem?.();
+        if (memBefore !== undefined) {
+          turnLogger?.debug("memory:turn:before", memorySampleAttributes(memBefore));
+        }
+
         let stepsPersisted = false;
         const result = await deps.runTurn({
           ...opts,
@@ -899,6 +931,23 @@ export function createSessionOrchestrator(
           },
         });
 
+        // Per-turn memory telemetry: capture a sample immediately AFTER the
+        // stream completes and log the per-turn delta vs `memBefore`. A
+        // positive rss delta on a sealed turn flags memory retained by the
+        // streaming path (the leak we are localizing). No I/O beyond the
+        // injected sampler; pure delta computation via memoryDelta(). The
+        // delta attributes carry a `delta` prefix so the absolute "after"
+        // values and the per-turn delta coexist without key collision.
+        if (memBefore !== undefined) {
+          const memAfter = sampleMem?.();
+          if (memAfter !== undefined) {
+            turnLogger?.info("memory:turn:after", {
+              ...memorySampleAttributes(memAfter),
+              ...memorySampleAttributes(memoryDelta(memBefore, memAfter), "delta"),
+            });
+          }
+        }
+
         // Fallback: if onStepComplete was never called (e.g., a fake
         // runTurn in tests), persist all result messages as a batch.
         if (!stepsPersisted && result.messages.length > 0) {
@@ -1042,6 +1091,10 @@ export function createSessionOrchestrator(
       return activeTurns.has(conversationId);
     },
 
+    getActiveConversationCount() {
+      return activeConversations.size;
+    },
+
     closeConversation(conversationId) {
       const turn = activeTurns.get(conversationId);
       const abortedTurn = turn !== undefined;
diff --git a/packages/session-orchestrator/src/pure.test.ts b/packages/session-orchestrator/src/pure.test.ts
index 7a574f1..d4fe84a 100644
--- a/packages/session-orchestrator/src/pure.test.ts
+++ b/packages/session-orchestrator/src/pure.test.ts
@@ -6,6 +6,9 @@ import {
   defaultDispatchPolicy,
   delayFor,
   generateTurnId,
+  type MemorySample,
+  memoryDelta,
+  memorySampleAttributes,
   RETRY_BUDGET_MS,
   RETRY_SCHEDULE_MS,
   RETRY_TAIL_MS,
@@ -194,3 +197,71 @@ describe("retry backoff schedule (delayFor)", () => {
     expect(schedule.slice(8).every((d) => d === RETRY_TAIL_MS)).toBe(true);
   });
 });
+
+describe("memorySampleAttributes", () => {
+  const sample: MemorySample = {
+    rss: 100 * 1024 * 1024, // 100 MB
+    heapUsed: 40 * 1024 * 1024,
+    heapTotal: 60 * 1024 * 1024,
+    external: 5 * 1024 * 1024,
+    arrayBuffers: 2 * 1024 * 1024,
+  };
+
+  it("formats fields as rounded MB with no prefix", () => {
+    const attrs = memorySampleAttributes(sample);
+    expect(attrs).toEqual({
+      rssMB: 100,
+      heapUsedMB: 40,
+      heapTotalMB: 60,
+      externalMB: 5,
+      arrayBuffersMB: 2,
+    });
+  });
+
+  it("namespaces keys with the given prefix", () => {
+    const attrs = memorySampleAttributes(sample, "delta");
+    expect(attrs).toEqual({
+      deltaRssMB: 100,
+      deltaHeapUsedMB: 40,
+      deltaHeapTotalMB: 60,
+      deltaExternalMB: 5,
+      deltaArrayBuffersMB: 2,
+    });
+  });
+
+  it("rounds fractional MB", () => {
+    const attrs = memorySampleAttributes({ ...sample, rss: 100.6 * 1024 * 1024 });
+    expect(attrs.rssMB).toBe(101);
+  });
+});
+
+describe("memoryDelta", () => {
+  const before: MemorySample = {
+    rss: 200 * 1024 * 1024,
+    heapUsed: 100 * 1024 * 1024,
+    heapTotal: 150 * 1024 * 1024,
+    external: 10 * 1024 * 1024,
+    arrayBuffers: 4 * 1024 * 1024,
+  };
+  const after: MemorySample = {
+    rss: 350 * 1024 * 1024,
+    heapUsed: 120 * 1024 * 1024,
+    heapTotal: 150 * 1024 * 1024,
+    external: 10 * 1024 * 1024,
+    arrayBuffers: 8 * 1024 * 1024,
+  };
+
+  it("computes signed after - before per field", () => {
+    const delta = memoryDelta(before, after);
+    expect(delta.rss).toBe(150 * 1024 * 1024);
+    expect(delta.heapUsed).toBe(20 * 1024 * 1024);
+    expect(delta.heapTotal).toBe(0);
+    expect(delta.external).toBe(0);
+    expect(delta.arrayBuffers).toBe(4 * 1024 * 1024);
+  });
+
+  it("is negative when memory dropped", () => {
+    const delta = memoryDelta(after, before);
+    expect(delta.rss).toBe(-150 * 1024 * 1024);
+  });
+});
diff --git a/packages/session-orchestrator/src/pure.ts b/packages/session-orchestrator/src/pure.ts
index 0d2068f..489b140 100644
--- a/packages/session-orchestrator/src/pure.ts
+++ b/packages/session-orchestrator/src/pure.ts
@@ -1,4 +1,5 @@
 import type {
+  Attributes,
   ChatMessage,
   Chunk,
   ImageInput,
@@ -134,3 +135,71 @@ export function defaultDispatchPolicy(): ToolDispatchPolicy {
 export function generateTurnId(): string {
   return `turn-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
 }
+
+// ── Memory telemetry (leak localization) ────────────────────────────────────
+//
+// Pure helpers for process.memoryUsage() sampling. The orchestrator owns the
+// sample SHAPE (this type) so its per-turn sampling and the host-bin periodic
+// timer share one contract without a cross-package import of an
+// implementation — host-bin imports this type, the orchestrator never imports
+// host-bin. Pure: inputs → attributes/delta, no I/O, no clock.
+
+/**
+ * A snapshot of process.memoryUsage() at one instant. Mirrors the subset of
+ * Node/Bun's MemoryUsage we log for leak localization (rss, heapUsed,
+ * heapTotal, external, arrayBuffers). Owned here so the orchestrator's
+ * per-turn sampling and the host-bin periodic timer agree on the shape.
+ */
+export interface MemorySample {
+  readonly rss: number;
+  readonly heapUsed: number;
+  readonly heapTotal: number;
+  readonly external: number;
+  readonly arrayBuffers: number;
+}
+
+const BYTES_PER_MB = 1024 * 1024;
+
+function mb(bytes: number): number {
+  return Math.round(bytes / BYTES_PER_MB);
+}
+
+/**
+ * Pure: format a {@link MemorySample} as flat logger {@link Attributes}
+ * (values in MB, rounded). Flat scalars are serializable (D3) and queryable
+ * (D9) in the journal. No I/O.
+ *
+ * Pass a `prefix` to namespace the keys — e.g. `memorySampleAttributes(delta,
+ * "delta")` yields `deltaRssMB`, so an "after" log can carry both the absolute
+ * sample (`rssMB`) and the per-turn delta (`deltaRssMB`) without key collision.
+ * The first letter of each field is capitalized after the prefix for
+ * readability (`deltaRssMB`, not `deltarssMB`).
+ */
+export function memorySampleAttributes(sample: MemorySample, prefix?: string): Attributes {
+  const p = prefix === undefined ? "" : prefix;
+  const cap = (s: string): string =>
+    s.length === 0 ? s : `${s[0]?.toUpperCase() ?? ""}${s.slice(1)}`;
+  const field = (name: string): string => (p.length === 0 ? name : `${p}${cap(name)}`);
+  return {
+    [field("rssMB")]: mb(sample.rss),
+    [field("heapUsedMB")]: mb(sample.heapUsed),
+    [field("heapTotalMB")]: mb(sample.heapTotal),
+    [field("externalMB")]: mb(sample.external),
+    [field("arrayBuffersMB")]: mb(sample.arrayBuffers),
+  };
+}
+
+/**
+ * Pure: compute the signed per-field delta `after - before`. A positive
+ * `rss` delta on a sealed turn flags memory retained by the streaming path
+ * (the prime leak suspect). No I/O.
+ */
+export function memoryDelta(before: MemorySample, after: MemorySample): MemorySample {
+  return {
+    rss: after.rss - before.rss,
+    heapUsed: after.heapUsed - before.heapUsed,
+    heapTotal: after.heapTotal - before.heapTotal,
+    external: after.external - before.external,
+    arrayBuffers: after.arrayBuffers - before.arrayBuffers,
+  };
+}