pi-better-harness 0.31.0 → 0.31.1

package/node_modules/pi-better-goal/README.md

@@ -59,7 +59,7 @@ After every successful context compaction (Pi's `session_compact` event) while a
 
 ### Scheduler-backed workflow gates
 
-A coordinator that ships the resolve-issues scheduler (`scripts/next-actions.mjs` in its skill directory, or in a sibling `resolve-issues` skill) keeps its Goal from ending while the run still has work. The extension runs `node next-actions.mjs --run <run dir> --slots 1 --peek --json` (read-only, 20-second timeout), where the run directory is the folder of the task plan bound through `update_plan`. Every gate fails open: if the scheduler is missing, errors, times out, prints bad output, or no plan is bound yet, the call proceeds and its result carries a warning, so nothing deadlocks.
+A coordinator that ships the resolve-issues scheduler (`scripts/next-actions.mjs` in its skill directory, or in a sibling `resolve-issues` skill) keeps its Goal from ending while the run still has work. The extension first runs `node next-actions.mjs --run <run dir> --slots 1 --peek --json` (read-only, 20-second timeout), where the run directory is the folder of the task plan bound through `update_plan`. For completion and handback only, if the sole due action is `sync`, it reruns with `--sync` to obtain a fresh issue tree. This live check has a four-minute timeout to accommodate the tree fetcher's network attempts and follow-ups. Because `--peek` suppresses scheduler bookkeeping but not issue-tree cache writes, the live check uses a temporary copy of the scheduler's plan, tree, bookkeeping, review, and result inputs; it does not write the bound run. Heartbeat probes never request live synchronization. A failed or incomplete tree fetch that produces `sync` still refuses completion or handback; cached evidence is not treated as live. Every gate retains its fail-open behavior for execution failures: if the scheduler is missing, errors, times out, prints bad output, or no plan is bound yet, the call proceeds and its result carries a warning, so nothing deadlocks.
 
 - **Completion.** `update_goal` complete is refused, with each due action (action, unit, first 200 characters of detail), unless the scheduler reports `finish` and nothing else is due (an empty due list means workers are still running, so it is refused too). The Goal stays active.
 - **Pause.** `release_workflow` is refused unless every unfinished unit is quiet (nothing due for it) or blocked `human-decision` with `humanItem` 2-5. A unit blocked `retry-ci` or any other reason the run retries itself, or a `human-decision` item 1, is named in the refusal and the Goal stays active.

package/node_modules/pi-better-goal/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-better-goal",
-  "version": "0.10.0",
+  "version": "0.10.1",
   "description": "Pi extension for goal tracking with background-aware continuation.",
   "license": "MIT",
   "type": "module",

package/node_modules/pi-better-goal/src/workflow-gates.ts

@@ -1,5 +1,6 @@
 import { execFile } from "node:child_process";
-import { readFileSync, statSync } from "node:fs";
+import { cpSync, existsSync, mkdtempSync, readFileSync, rmSync, statSync } from "node:fs";
+import { tmpdir } from "node:os";
 import { dirname, join } from "node:path";
 
 import type { WorkflowOwner } from "./workflow.js";
@@ -8,6 +9,8 @@ import type { WorkflowOwner } from "./workflow.js";
 const WORKFLOW_PLAN_ENTRY = "pi-better-workflow-plan";
 const WORKFLOW_ENTRY = "pi-better-workflow";
 export const SCHEDULER_TIMEOUT_MS = 20_000;
+// Live tree queries allow three 60-second attempts plus follow-ups; cached probes stay cheap.
+const SCHEDULER_SYNC_TIMEOUT_MS = 240_000;
 const DETAIL_LIMIT = 200;
 
 /** Scheduler locations relative to the owner's skill directory, in precedence order. */
@@ -100,9 +103,13 @@ function parseSchedulerOutput(stdout: string): SchedulerRun {
  * are due and 2 on a usage or plan error, so only the latter, a timeout, or bad output is a failure.
  */
 export function runScheduler(script: string, runDir: string, timeoutMs = SCHEDULER_TIMEOUT_MS): Promise<SchedulerRun> {
+  return executeScheduler(script, runDir, timeoutMs);
+}
+
+function executeScheduler(script: string, runDir: string, timeoutMs: number, live = false, cwd = runDir): Promise<SchedulerRun> {
   return new Promise((resolve) => {
-    execFile("node", [script, "--run", runDir, "--slots", "1", "--peek", "--json"],
-      { cwd: runDir, timeout: timeoutMs, encoding: "utf8", windowsHide: true, maxBuffer: 8 * 1024 * 1024 },
+    execFile("node", [script, "--run", runDir, "--slots", "1", "--peek", "--json", ...(live ? ["--sync"] : [])],
+      { cwd, timeout: timeoutMs, encoding: "utf8", windowsHide: true, maxBuffer: 8 * 1024 * 1024 },
       (error, stdout, stderr) => {
         if (error) {
           // An exit status of 1 means "actions due"; any other failure (exit 2, timeout, spawn error) is real.
@@ -117,6 +124,27 @@ export function runScheduler(script: string, runDir: string, timeoutMs = SCHEDUL
   });
 }
 
+async function settlementScheduler(script: string, runDir: string, timeoutMs?: number): Promise<SchedulerRun> {
+  const cached = await runScheduler(script, runDir, timeoutMs);
+  if (!cached.ok || cached.due.length !== 1 || cached.due[0]!.action !== "sync") return cached;
+
+  // --peek suppresses seen bookkeeping, not syncRun's issue-tree write. Give the live
+  // probe private inputs; keep cwd so relative workspace paths retain their meaning.
+  let snapshot: string | undefined;
+  try {
+    snapshot = mkdtempSync(join(tmpdir(), "pi-goal-scheduler-"));
+    for (const name of ["task-plan.json", "issue-tree.json", "next-actions-seen.json", "reviews", "results"]) {
+      const source = join(runDir, name);
+      if (existsSync(source)) cpSync(source, join(snapshot, name), { recursive: true, dereference: true });
+    }
+    return await executeScheduler(script, snapshot, timeoutMs ?? SCHEDULER_SYNC_TIMEOUT_MS, true, runDir);
+  } catch (error) {
+    return { ok: false, reason: `could not prepare the live scheduler check (${String(error).slice(0, DETAIL_LIMIT)})` };
+  } finally {
+    if (snapshot) rmSync(snapshot, { recursive: true, force: true });
+  }
+}
+
 interface PlannedRun {
   runDir: string;
   script: string;
@@ -147,7 +175,7 @@ export async function completionGate(
 ): Promise<GateVerdict> {
   const located = locateRun(entries, owner);
   if ("warning" in located) return unchecked(located.warning);
-  const run = await runScheduler(located.script, located.runDir, options.timeoutMs);
+  const run = await settlementScheduler(located.script, located.runDir, options.timeoutMs);
   if (!run.ok) return unchecked(`could not check the ${owner.name} run (${run.reason})`);
   const open = run.due.filter((action) => action.action !== "finish");
   if (open.length === 0 && run.due.length > 0) return { allow: true };
@@ -190,7 +218,7 @@ export async function pauseGate(
   if ("warning" in located) return unchecked(located.warning);
   const units = readUnits(located.runDir);
   if (!units) return unchecked(`could not read the ${owner.name} task plan`);
-  const run = await runScheduler(located.script, located.runDir, options.timeoutMs);
+  const run = await settlementScheduler(located.script, located.runDir, options.timeoutMs);
   if (!run.ok) return unchecked(`could not check the ${owner.name} run (${run.reason})`);
 
   const refusals: string[] = [];

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-better-harness",
-  "version": "0.31.0",
+  "version": "0.31.1",
   "description": "Pi extension bundle for a write sandbox, subagents, background tasks, SSH, goals, and structured plans.",
   "license": "MIT",
   "type": "module",
@@ -54,7 +54,7 @@
   },
   "dependencies": {
     "pi-better-background-tasks": "0.9.1",
-    "pi-better-goal": "0.10.0",
+    "pi-better-goal": "0.10.1",
     "pi-better-plan": "0.9.1",
     "pi-better-sandbox": "0.12.1",
     "pi-better-ssh": "0.1.3",