perf(session): coalesce per-turn persistence to ~2 disk-touching writes

scripts/turn-persistence-census.ts

@@ -0,0 +1,277 @@
+#!/usr/bin/env bun
+/**
+ * CL-10667 before/after measurement: per-turn session persistence write ops.
+ *
+ * Mirrors the reactor's persist sequence for one agentic turn: one inference
+ * cycle (writePrompt + writeResponse) and two tool batches followed by the
+ * cycle terminal. Every write-kind fs.promises call of the store stack is
+ * counted (split by whether the path is under .git), so the coalescing becomes
+ * an honest before/after delta on the same host and scenario.
+ *
+ * The flow itself changed shape with CL-10667, so the same script models both:
+ * --flow=legacy replicates the pre-coalescing reactor (tool-batch commitCycle
+ * per batch); --flow=coalesced replicates the post-coalescing one (writeTurns
+ * per tool batch, full commitCycle only at the cycle terminal).
+ *
+ * Usage:
+ *   bun scripts/turn-persistence-census.ts [turnCount] [--flow=legacy|coalesced] [--json]
+ */
+
+import fs from "node:fs";
+import path from "node:path";
+import type { ConversationTurn, TokenUsage } from "@intx/types/runtime";
+import { TransformRecord } from "@intx/types/runtime";
+import { createOptimizedContextStore } from "../src/session/optimized-context-store.js";
+
+const WRITE_OPS = [
+  "writeFile",
+  "appendFile",
+  "rename",
+  "unlink",
+  "truncate",
+  "mkdir",
+  "copyFile",
+] as const;
+
+const counts: Record<string, number> = {};
+const gitPaths: Record<string, number> = {};
+
+function bump(table: Record<string, number>, op: string): void {
+  table[op] = (table[op] ?? 0) + 1;
+}
+
+function installCounters(): void {
+  const target = fs.promises as unknown as Record<string, unknown>;
+  for (const op of WRITE_OPS) {
+    const original = target[op];
+    if (typeof original !== "function") continue;
+    target[op] = function counted(this: unknown, ...args: unknown[]) {
+      const first = args[0];
+      if (typeof first === "string" && first.includes(`/.git/`)) {
+        bump(gitPaths, op);
+      }
+      bump(counts, op);
+      return (original as (...args: unknown[]) => unknown).apply(this, args);
+    };
+  }
+}
+
+function turn(text: string): ConversationTurn {
+  return { role: "user", content: [{ type: "text", text }], timestamp: 1 };
+}
+
+function assistantTurn(text: string): ConversationTurn {
+  return {
+    role: "assistant",
+    content: [{ type: "text", text }],
+    timestamp: 2,
+  };
+}
+
+function toolResultTurn(name: string): ConversationTurn {
+  return {
+    role: "user",
+    content: [
+      {
+        type: "tool_result",
+        callId: `call-${name}`,
+        content: [{ type: "text", text: `${name} result` }],
+      },
+    ],
+    timestamp: 3,
+  };
+}
+
+function turnBody(index: number): string {
+  return (
+    `turn ${index}: ` + "census padding. ".repeat(40) + `end-of-turn ${index}`
+  );
+}
+
+const EMPTY_USAGE: TokenUsage = {
+  input: 0,
+  output: 0,
+  cacheRead: 0,
+  cacheWrite: 0,
+  thinking: 0,
+};
+
+function usageForTurn(turnCount: number): TokenUsage {
+  return {
+    ...EMPTY_USAGE,
+    input: 1000 * turnCount,
+    output: 500 * turnCount,
+  };
+}
+
+async function runScenario(
+  dir: string,
+  turnCount: number,
+  flow: string,
+): Promise<{ turns: number; persistCalls: Record<string, number> }> {
+  const base = await createOptimizedContextStore(dir);
+  const PERSIST_METHODS = [
+    "writePrompt",
+    "writeResponse",
+    "writeTurns",
+    "writeManifest",
+    "writeMetadata",
+    "commit",
+  ] as const;
+  const persistCalls: Record<string, number> = {};
+  const store = Object.assign(Object.create(base), base) as typeof base;
+  for (const method of PERSIST_METHODS) {
+    const original = base[method] as (
+      ...callArgs: unknown[]
+    ) => Promise<unknown>;
+    (store as never as Record<string, unknown>)[method] = (
+      ...callArgs: unknown[]
+    ) => {
+      bump(persistCalls, method);
+      return original.apply(base, callArgs);
+    };
+  }
+  const history: ConversationTurn[] = [];
+  const record: () => TransformRecord = () => ({
+    strategy: "test.transform",
+    version: "1",
+    parameters: {},
+    reason: "census",
+    decisions: {},
+  });
+  for (let t = 0; t < turnCount; t++) {
+    // Inference cycle: prompt + assistant turn.
+    history.push(turn(turnBody(t)));
+    history.push(assistantTurn(`reply to ${t}`));
+    await store.writePrompt(history);
+
+    if (flow === "legacy") {
+      const last = history[history.length - 1];
+      if (last !== undefined && last.role === "assistant") {
+        await store.writeResponse({
+          ...last,
+          role: "assistant",
+          model: last.model ?? "census-model",
+        });
+      }
+    }
+
+    for (let batch = 0; batch < 2; batch++) {
+      history.push(toolResultTurn(`tool-${t}-${batch}`));
+      if (flow === "legacy") {
+        // Old tool-batch path: a full commitCycle per batch.
+        await store.writeTurns(history);
+        await store.writeManifest([record()]);
+        await store.writeMetadata({
+          pendingOperations: [],
+          tokenUsage: usageForTurn(t + 1),
+        });
+        await store.commit({ message: "checkpoint: tool-execution" });
+      } else {
+        // Coalesced tool-batch path: segments on disk only.
+        await store.writeTurns(history);
+      }
+    }
+
+    // Cycle terminal (reply/wait/done): the remaining cycle persists.
+    if (flow !== "legacy") {
+      await store.writeTurns(history);
+      // Two transforms happened; the terminal commit carries them.
+      await store.writeManifest([record(), record()]);
+      await store.writeMetadata({
+        pendingOperations: [],
+        tokenUsage: usageForTurn(t),
+      });
+      await store.commit({ message: "Cycle: done" });
+    }
+  }
+  const fresh = await createOptimizedContextStore(dir);
+  const turns = (await fresh.load()).turns.length;
+  return { turns, persistCalls };
+}
+
+function sortedEntries(table: Record<string, number>): [string, number][] {
+  return Object.entries(table).sort((a, b) => b[1] - a[1]);
+}
+
+async function main(argv: string[]): Promise<number> {
+  const arg = argv[2];
+  const flow = argv.find((a) => a.startsWith("--flow="))?.split("=")[1];
+  if (flow !== "legacy" && flow !== "coalesced") {
+    console.error(
+      "usage: bun scripts/turn-persistence-census.ts [turnCount] --flow=legacy|coalesced [--json]",
+    );
+    return 1;
+  }
+  const turnCount =
+    arg !== undefined && !arg.startsWith("-") ? Number(arg) : 40;
+  if (!Number.isFinite(turnCount) || turnCount < 1) {
+    console.error(`invalid turnCount: ${arg}`);
+    return 1;
+  }
+  installCounters();
+
+  const prefix = `census-${flow}-`;
+  const dir = await fs.promises.mkdtemp(path.join(process.cwd(), prefix));
+  const started = performance.now();
+  const { turns: loadedTurns, persistCalls } = await runScenario(
+    dir,
+    turnCount,
+    flow,
+  );
+  const wallMs = Math.round(performance.now() - started);
+
+  const totalOps = Object.values(counts).reduce((sum, n) => sum + n, 0);
+  const gitOps = Object.values(gitPaths).reduce((sum, n) => sum + n, 0);
+
+  const asJson = argv.includes("--json");
+  if (asJson) {
+    console.log(
+      JSON.stringify(
+        {
+          flow,
+          turnCount,
+          wallMs,
+          loadedTurns,
+          totalWriteOps: totalOps,
+          gitPathWriteOps: gitOps,
+          byName: counts,
+          gitByPath: gitPaths,
+          perTurn: {
+            wallMs: +(wallMs / turnCount).toFixed(2),
+            writeOps: +(totalOps / turnCount).toFixed(2),
+          },
+          persistCalls,
+          persistCallsPerTurn: Object.fromEntries(
+            Object.entries(persistCalls).map(([k, v]) => [
+              k,
+              +(v / turnCount).toFixed(3),
+            ]),
+          ),
+        },
+        null,
+        2,
+      ),
+    );
+  } else {
+    console.log(
+      `flow=${flow} ${turnCount} turns; verified ${loadedTurns} turns at the end`,
+    );
+    console.log(
+      `wall: ${wallMs} ms (${(wallMs / turnCount).toFixed(2)} ms/turn)`,
+    );
+    console.log(
+      `write ops: ${totalOps} total (${(totalOps / turnCount).toFixed(2)}/turn), of which ${gitOps} under .git`,
+    );
+    for (const [op, n] of sortedEntries(counts)) {
+      console.log(`  ${op}: ${n}`);
+    }
+    for (const [op, n] of sortedEntries(gitPaths)) {
+      console.log(`  (.git) ${op}: ${n}`);
+    }
+  }
+  await fs.promises.rm(dir, { recursive: true, force: true });
+  return 0;
+}
+
+process.exit(await main(process.argv));

src/exec/runner.test.ts

@@ -1,5 +1,6 @@
 import { describe, expect, spyOn, test } from "bun:test";
 import type { ToolCall } from "@intx/types/runtime";
+import { defined } from "../../testkit/defined.js";
 import type { AgentTool } from "@intx/agent";
 import { submitOutputDefinition } from "../agent/director.js";
 import { DIRECTOR_REGISTRY } from "../agent/directors/registry.js";
@@ -564,6 +565,99 @@ describe("runExec", () => {
       cleanup();
     }
   });
+
+  test("heartbeat forces identical-payload writes; a failed save does not take the baseline", async () => {
+    // CL-10667 review fix: the exec heartbeat tick must bypass the
+    // identical-payload skip (mtime freshness is a liveness contract read
+    // cross-process by ageStaleRunningState), and the baseline must record
+    // only successes so a transient save failure cannot freeze run.json.
+    const previous = getActiveRun();
+    clearActiveRun();
+    const sessionId = "exec-heartbeat-fail";
+    const { cwd, home, cleanup } = createTempDirs(
+      "corbits-exec-heartbeat-cwd-",
+      "corbits-exec-heartbeat-home-",
+    );
+    try {
+      const saveCalls: { ok: boolean; status?: string }[] = [];
+      await withMockedHomedir(home, () =>
+        withMockedModuleDuring(
+          import.meta.resolve("../session/state.js"),
+          (real: typeof import("../session/state.js")) =>
+            ({
+              ...real,
+              saveState(
+                cwd: string,
+                sessionId: string,
+                state: import("../session/state.js").RunState,
+                dir?: string,
+              ) {
+                if (saveCalls.length === 0) {
+                  saveCalls.push({ ok: false, status: state.status });
+                  // Transient failure on the first running write.
+                  return Promise.reject(new Error("ENOSPC-like"));
+                }
+                saveCalls.push({ ok: true, status: state.status });
+                return real.saveState(cwd, sessionId, state, dir);
+              },
+            }) as typeof real,
+          () =>
+            withMockedModuleDuring(
+              import.meta.resolve("../session/run-liveness.js"),
+              (real: typeof import("../session/run-liveness.js")) =>
+                ({
+                  ...real,
+                  RUN_HEARTBEAT_INTERVAL_MS: 10,
+                }) as typeof real,
+              () =>
+                withMockedModuleDuring(
+                  import.meta.resolve("../session/assemble-runtime.js"),
+                  (inner: typeof import("../session/assemble-runtime.js")) =>
+                    ({
+                      ...inner,
+                      assembleInferenceBase: () =>
+                        // Delay the bootstrap failure so heartbeat ticks run
+                        // while the run is unsettled.
+                        new Promise<never>((_, reject) => {
+                          setTimeout(
+                            () => reject(new Error("bootstrap failed")),
+                            80,
+                          );
+                        }),
+                    }) as typeof inner,
+                  async () => {
+                    const { runExec: runExecUnderMock } =
+                      await import("./runner.js");
+                    const result = await runExecUnderMock({
+                      ...bareConfig("do the thing"),
+                      cwd,
+                      sessionId,
+                    });
+                    expect(result.status).toBe("failed");
+                  },
+                ),
+            ),
+        ),
+      );
+
+      // Write 1 (running) failed; the heartbeat retried the identical payload
+      // and forced a disk write past any baseline; the terminal wrote last.
+      expect(saveCalls[0]).toEqual({ ok: false, status: "running" });
+      const runningOk = saveCalls.findIndex((call) => call.ok);
+      expect(runningOk).toBeGreaterThan(0);
+      expect(defined(saveCalls[runningOk]).status).toBe("running");
+      expect(defined(saveCalls[saveCalls.length - 1]).status).toBe("failed");
+      const finalPersisted = await loadState(cwd, sessionId, home);
+      expect(finalPersisted).toMatchObject({
+        kind: "ok",
+        state: { status: "failed" },
+      });
+    } finally {
+      if (previous !== null) setActiveRun(previous);
+      else clearActiveRun();
+      cleanup();
+    }
+  });
 });
 
 describe("resolveExecDirectorOverlay", () => {