pi-goal-list-loop-audit 0.28.23 → 0.28.25

package/extensions/goal-loop-backoff.ts

@@ -162,6 +162,8 @@ export interface HeartbeatInput {
   /** Milliseconds since the last observed agent activity. */
   msSinceActivity: number;
   stallMs?: number;
+  /** v0.28.25: consecutive stall refires so far — spaces refires exponentially. */
+  consecutiveStalls?: number;
 }
 
 /** Should the heartbeat re-fire the continuation right now? */
@@ -169,7 +171,15 @@ export function shouldHeartbeatRefire(input: HeartbeatInput): boolean {
   if (!input.supervising) return false;
   if (!input.sessionIdle) return false;
   if (input.timerPending) return false;
-  return input.msSinceActivity >= (input.stallMs ?? HEARTBEAT_STALL_MS);
+  // v0.28.25: exponential spacing between stall refires — 1m, 2m, 4m, 8m
+  // (cap 8×). Field-observed in junk-runner: the flat 60s gate burned all
+  // 5 refires in ~4 minutes into a just-compacted session, pausing a
+  // resumable goal instead of giving the provider/queue time to recover.
+  // noteActivity() runs at each refire, so msSinceActivity measures the
+  // silence SINCE the last refire — scaling the threshold scales the gap.
+  const stallMs = input.stallMs ?? HEARTBEAT_STALL_MS;
+  const scale = 2 ** Math.min(input.consecutiveStalls ?? 0, 3);
+  return input.msSinceActivity >= stallMs * scale;
 }
 
 /**

package/extensions/loops/goal.ts

@@ -463,6 +463,25 @@ function startUITicker(): void {
 // and escalated loudly past 5 minutes.
 let continuationRearmStreak = 0;
 let loopRearmStreak = 0;
+// v0.28.24: post-compaction grace — a just-replaced session gets 3 minutes
+// to settle (queue drain, provider recovery) before stall counting resumes.
+// Field-observed in junk-runner: a 196k-token compact finished, then the
+// heartbeat burned all 5 stall refires in the next 5 minutes into a session
+// whose turn trigger was still dead — pausing a resumable goal 4 minutes
+// after the compact instead of giving pi room to recover.
+let compactionGraceUntil = 0;
+const COMPACTION_GRACE_MS = 3 * 60_000;
+// v0.28.25: provider-error retry cadence. Field-observed in dracon-utilities
+// (kimi, 19-session fleet on one provider account): a "concurrent request
+// limit" 403 storm got 5 retries BACK-TO-BACK (delay 0 after each errored
+// turn — the session is idle at agent_end, so scheduleContinuation fired
+// instantly) and the brake then cycled on a flat 60s cooldown for 1h 38m.
+// The condition clears on a minutes-to-fleet scale, not milliseconds:
+// ladder the inter-error retries (5s, 15s, 45s, 90s, 3m — the 5-retry
+// budget now spans ~5.5m) and escalate the brake cooldown per consecutive
+// brake (1m, 2m, 4m, 8m, 16m cap). A successful turn resets both.
+const ERROR_RETRY_LADDER_MS = [5_000, 15_000, 45_000, 90_000, 180_000];
+let errorBrakeStreak = 0;
 const SEND_REARM_LEDGER_EVERY = 600; // 600 × 50ms = 30s
 const SEND_REARM_ESCALATE_AT = 6000; // 6000 × 50ms = 5 minutes
 
@@ -545,6 +564,10 @@ function heartbeatTick(): void {
     return;
   }
   const sessionIdle = idle && !pending;
+  // v0.28.24: post-compaction grace — the whole stall/refire/watchdog
+  // machinery below stays quiet for 3 minutes while the replaced session
+  // settles (latch watchdog, wedge alert, refire counting all resume after).
+  if (Date.now() < compactionGraceUntil) return;
   // v0.26.5: pending-latch watchdog — a queued continuation whose turn
   // trigger was dropped (field-observed post-compaction: continuation
   // ACCEPTED at compact+0s, then 22 minutes of silence). The stuck latch
@@ -580,6 +603,7 @@ function heartbeatTick(): void {
     timerPending: continuationTimer !== null || loopTimer !== null,
     msSinceActivity: Date.now() - lastActivityAt,
     stallMs: HEARTBEAT_STALL_MS,
+    consecutiveStalls,
   });
   // Wedge alert (v0.23.2): session BUSY but silent for the threshold —
   // the classic hung-command case (a test suite that never exits holds
@@ -676,7 +700,7 @@ function freshCtx(): ExtensionContext | null {
   }
 }
 
-function scheduleContinuation(ctx: ExtensionContext, force = false): void {
+function scheduleContinuation(ctx: ExtensionContext, force = false, delayMs?: number): void {
   if (!isActionableGoal()) return;
   rememberCtx(ctx);
   const goalId = state.goal!.id;
@@ -684,7 +708,7 @@ function scheduleContinuation(ctx: ExtensionContext, force = false): void {
   clearContinuationTimer();
   let delay = 0;
   try {
-    delay = ctx.isIdle() && !ctx.hasPendingMessages() ? 0 : BACKOFF_IDLE_RETRY_MS;
+    delay = delayMs ?? (ctx.isIdle() && !ctx.hasPendingMessages() ? 0 : BACKOFF_IDLE_RETRY_MS);
   } catch {
     return;
   }
@@ -1205,10 +1229,10 @@ async function cmdSet(args: string, ctx: ExtensionContext, skipDraft = false): P
     // v0.28.1 (S3): the goal is persisted — mark the interrupt so the next
     // fresh session LOADS it (held by default since v0.28.21), and tell the truth instead of "starting now".
     updateGoal({ interruptedAt: nowIso(), interruptedReason: "created in a stale session" }, ctx);
-    ctx.ui.notify(`Goal saved: ${shortObj(goal.objective)} — safe in .pi-glla/, but this stale process can't send continuations. Restart pi, then /goal resume (v0.28.21: session loads no longer auto-start by default). (id: ${goal.id})`, "warning");
+    ctx.ui.notify(`Goal saved: ${shortObj(goal.objective)} — safe in .pi-glla/, but this stale process can't send continuations. Restart pi, then /goal resume (v0.28.21: session loads no longer auto-start by default).`, "warning");
     return;
   }
-  ctx.ui.notify(`Goal started: ${shortObj(goal.objective)} — the auditor will verify on completion. (id: ${goal.id})`, "info");
+  ctx.ui.notify(`Goal started: ${shortObj(goal.objective)} — the auditor will verify on completion.`, "info");
   scheduleContinuation(ctx, true);
 }
 
@@ -1219,8 +1243,7 @@ async function cmdStatus(ctx: ExtensionContext): Promise<void> {
   }
   const g = state.goal;
   const lines = [
-    `[${g.id}] ${statusLabel(g.status)}`,
-    `Objective: ${g.objective}`,
+    `${statusLabel(g.status)}: ${g.objective}`,
     // v0.24.7: name WHERE the work came from — a queue item is not a goal.
     ...(g.policy === "list" ? [`Source: /list queue (${listQueue().length} waiting) — /list to manage`] : []),
     `Auto-continue: ${g.autoContinue ? "on" : "off"}`,
@@ -1240,10 +1263,10 @@ async function cmdPause(ctx: ExtensionContext): Promise<void> {
   // v0.22.7: name WHAT was paused — a list item resumes through /list.
   if (state.goal.policy === "list") {
     const queued = listQueue().length;
-    ctx.ui.notify(`List item ${state.goal.id} paused${queued > 0 ? ` (${queued} waiting in the list)` : ""}. /list resume to continue.`, "info");
+    ctx.ui.notify(`List item "${shortObj(state.goal.objective)}" paused${queued > 0 ? ` (${queued} waiting in the list)` : ""}. /list resume to continue.`, "info");
     return;
   }
-  ctx.ui.notify(`Goal ${state.goal.id} paused. /goal resume to continue.`, "info");
+  ctx.ui.notify(`Goal "${shortObj(state.goal.objective)}" paused. /goal resume to continue.`, "info");
 }
 
 async function cmdResume(ctx: ExtensionContext): Promise<void> {
@@ -2732,7 +2755,7 @@ function registerAgentTools(pi: any, ctx: ExtensionContext): void {
   pi.registerTool(defineTool({
     name: "pause_goal",
     label: "Pause goal",
-    description: "Pause the active goal with a reason and suggested action. Use when blocked on user input or unable to make progress. When the user must CHOOSE between options, pass kind=\"decision\" with the options list (recommended = 1-based index of the best one) — decision pauses render as a prominent DECISION NEEDED card. Time-gated waits (retry at a specific time) use kind=\"wait\" with resumeAt (ISO). Operational failures use kind=\"error\".",
+    description: "Pause the active goal with a reason and suggested action. Use when blocked on user input or unable to make progress. When the user must CHOOSE between options, pass kind=\"decision\" with the options list (recommended = 1-based index of the best one) — decision pauses render as a prominent DECISION NEEDED card. Time-gated waits (retry at a specific time) use kind=\"wait\" with resumeAt (ISO). Operational failures use kind=\"error\". VOCABULARY (v0.28.24): decision options and reasons must reference REAL commands only — /goal resume, /goal cancel, /goal tweak \"<new text>\", /list remove N, /list next, /list resume, /loop stop, /loop resume. These all act on the ACTIVE goal/item: there is NO /goal drop and NO command takes a goal id. Never show goal ids to the user — name the thing ('the active goal', 'list item \"<short name>\"'); ids are internal plumbing the user cannot act on.",
     parameters: Type.Object({
       reason: Type.String({ description: "Why the work is paused" }),
       suggestedAction: Type.Optional(Type.String({ description: "What the user should do next" })),
@@ -4420,6 +4443,12 @@ export default function (pi: ExtensionAPI): void {
     rememberCtx(ctx);
     if (!isSupervising()) return;
     appendLedger(ctx.cwd, "session_compact", {});
+    // v0.28.24: a compaction is LEGITIMATE busy time — reset the send-rearm
+    // storm streaks (π-web nearly escalated a "send-retry storm" pause during
+    // a 3.5-minute compact) and open the post-compaction stall grace.
+    continuationRearmStreak = 0;
+    loopRearmStreak = 0;
+    compactionGraceUntil = Date.now() + COMPACTION_GRACE_MS;
     const settle = setTimeout(() => {
       const c = freshCtx();
       if (!c) return;
@@ -4592,7 +4621,7 @@ export default function (pi: ExtensionAPI): void {
         // the auto-resume the marker promised.
         if (wasInterrupted) updateGoal({ interruptedAt: undefined, interruptedReason: undefined }, ctx);
         ctx.ui.notify(
-          `Resuming ${state.goal.policy === "list" ? "list item" : "goal"} [${state.goal.id}]: ${state.goal.objective.slice(0, 70)}${listQueue().length > 0 ? ` (+${listQueue().length} queued)` : ""}${wasInterrupted ? " — auto-resumed after the stale-handle interrupt" : ""}`,
+          `Resuming ${state.goal.policy === "list" ? "list item" : "goal"}: ${state.goal.objective.slice(0, 70)}${listQueue().length > 0 ? ` (+${listQueue().length} queued)` : ""}${wasInterrupted ? " — auto-resumed after the stale-handle interrupt" : ""}`,
           "info",
         );
         // v0.28.4 (P3): skip nudge accounting for the first recovery turns.
@@ -4611,14 +4640,14 @@ export default function (pi: ExtensionAPI): void {
           pauseSuggestedAction: resumeHint,
         }, ctx);
         ctx.ui.notify(
-          `${isListItem ? "List item" : "Goal"} held on restore [${state.goal.id}]: ${state.goal.objective.slice(0, 70)}${queued > 0 ? ` (+${queued} waiting in the list)` : ""} — ${resumeCmd} to continue.`,
+          `${isListItem ? "List item" : "Goal"} held on restore: ${state.goal.objective.slice(0, 70)}${queued > 0 ? ` (+${queued} waiting in the list)` : ""} — ${resumeCmd} to continue.`,
           "info",
         );
       }
     } else if (state.goal && state.goal.status === "active") {
       // Active but autoContinue off: nothing auto-fires — just surface it.
       ctx.ui.notify(
-        `Restored ${state.goal.policy === "list" ? "list item" : "goal"} [${state.goal.id}]: ${state.goal.objective.slice(0, 70)}${listQueue().length > 0 ? ` (+${listQueue().length} queued)` : ""}`,
+        `Restored ${state.goal.policy === "list" ? "list item" : "goal"}: ${state.goal.objective.slice(0, 70)}${listQueue().length > 0 ? ` (+${listQueue().length} queued)` : ""}`,
         "info",
       );
     } else if ((!state.goal || state.goal.status === "complete" || state.goal.status === "aborted") && listQueue().length > 0) {
@@ -4644,7 +4673,7 @@ export default function (pi: ExtensionAPI): void {
         pauseSuggestedAction: "/loop to work the loop, or /loop stop then /goal resume to work the goal",
       }, ctx);
       ctx.ui.notify(
-        `Goal [${state.goal.id}] held — a loop also exists; one active thing at a time. /loop to resume the loop, or /loop stop then /goal resume.`,
+        `Goal held — a loop also exists; one active thing at a time. /loop to resume the loop, or /loop stop then /goal resume.`,
         "info",
       );
       maybeDecisionPopup(ctx);
@@ -4804,33 +4833,46 @@ export default function (pi: ExtensionAPI): void {
       if (consecutiveErrorIterations >= 5) {
         // v0.28.5 (E8): carry the REAL error text — the pause used to say
         // literally "5 consecutive errors: error" (stopReason, not the
-        // provider error). And give transient flakes ONE capped auto-resume
-        // (60s, reason re-checked) — the E8 incident lost 1.5h to a
+        // provider error). And give transient flakes ONE auto-resume per brake
+        // (escalating cooldown, reason re-checked) — the E8 incident lost 1.5h to a
         // 60-second provider hiccup waiting on a manual /goal resume.
         const detail = text.trim() ? ` (last: ${text.trim().replace(/\s+/g, " ").slice(0, 160)})` : "";
         const reason = `5 consecutive errors${detail}`;
+        // v0.28.25: the cooldown escalates per CONSECUTIVE brake — a fleet-wide
+        // 403 window is not cleared by re-braking every 60 seconds.
+        const cooldownMs = 60_000 * 2 ** Math.min(errorBrakeStreak, 4);
+        const cooldownMin = Math.round(cooldownMs / 60_000);
+        errorBrakeStreak++;
         updateGoal({
           status: "paused",
           pauseKind: "wait",
-          pauseResumeAt: new Date(Date.now() + 60_000).toISOString(),
+          pauseResumeAt: new Date(Date.now() + cooldownMs).toISOString(),
           pauseReason: reason,
-          pauseSuggestedAction: "Transient provider flake? The goal auto-resumes once in 60s if still paused for this reason — or /goal resume now.",
+          pauseSuggestedAction: `Transient provider flake? The goal auto-resumes once in ${cooldownMin}m if still paused for this reason — or /goal resume now.`,
         }, ctx);
         ctx.ui.notify(`Goal paused: ${reason}.`, "warning");
         notifyExternal(ctx, `Goal paused: ${reason}.`);
         appendLedger(ctx.cwd, "goal_paused", { reason });
-        scheduleQuotaRetry(ctx, 60, reason, () => {
+        scheduleQuotaRetry(ctx, cooldownMs / 1000, reason, () => {
           // Re-check: only auto-resume if STILL paused for the error brake
           // (a user /goal pause during the window is not stomped).
           if (state.goal && state.goal.status === "paused" && (state.goal.pauseReason ?? "").startsWith("5 consecutive errors")) {
             updateGoal({ status: "active" }, ctx);
             appendLedger(ctx.cwd, "goal_resumed", { via: "error-brake-retry" });
-            ctx.ui.notify("Auto-resumed after the 5-error brake (60s cooldown).", "info");
+            ctx.ui.notify("Auto-resumed after the 5-error brake (cooldown elapsed).", "info");
             scheduleContinuation(ctx, true);
           }
         }, "5 consecutive errors — auto-retry");
         return;
       }
+      // v0.28.25: under the brake, the retry rides the exponential ladder —
+      // NOT the immediate scheduleContinuation at the bottom of this handler
+      // (an errored turn leaves the session idle, so the default delay is 0:
+      // exactly how 5 retries fired back-to-back in dracon-utilities).
+      const retryDelayMs = ERROR_RETRY_LADDER_MS[Math.min(consecutiveErrorIterations - 1, ERROR_RETRY_LADDER_MS.length - 1)];
+      appendLedger(ctx.cwd, "error_retry_backoff", { attempt: consecutiveErrorIterations, delayMs: retryDelayMs });
+      scheduleContinuation(ctx, true, retryDelayMs);
+      return;
     } else if (stopReason === "aborted") {
       // v0.28.5 (E8): user aborts are not provider errors. Separate brake,
       // honest message, and NO auto-resume — aborting five turns in a row
@@ -4851,6 +4893,7 @@ export default function (pi: ExtensionAPI): void {
     } else {
       consecutiveErrorIterations = 0;
       consecutiveAbortIterations = 0;
+      errorBrakeStreak = 0; // v0.28.25: a healthy turn clears the brake cooldown
     }
 
     // No wall-clock cap by design: a goal ends via completion, explicit

package/extensions/reviewer.ts

@@ -111,17 +111,84 @@ export function stripCodeSpans(text: string): string {
     .replace(/`[^`\n]*`/g, " ");
 }
 
+/** v0.28.24: join hard-wrapped lines before classification. Completion
+ * summaries and transcripts arrive wrapped at ~70 cols, and line-by-line
+ * extraction sliced findings at the wrap point (field-observed in
+ * hellhunter: a list item whose ENTIRE objective was "Run a post-completion
+ * regression scan on the hellhunter codebase to" — the first visual line of
+ * a wrapped paragraph, enqueued by the convert-findings-to-list cascade).
+ * A line that doesn't end a sentence continues on the next line unless that
+ * line starts a new list item or heading. */
+export function unwrapHardWrappedLines(text: string): string {
+  const lines = text.split("\n");
+  const out: string[] = [];
+  for (const line of lines) {
+    const prev = out[out.length - 1];
+    const startsNewItem = /^\s*([-*•>]|\d+[.)]|#)/.test(line);
+    // v0.28.24: join only when the continuation starts LOWERCASE — the
+    // mid-sentence signal. Punctuation-less standalone items ("TODO: fix x")
+    // start uppercase/keyword and must NOT merge with the next item.
+    const continuesSentence = /^[a-z]/.test(line.trimStart());
+    if (
+      prev !== undefined &&
+      prev.trim().length > 0 &&
+      line.trim().length > 0 &&
+      !startsNewItem &&
+      continuesSentence &&
+      !/[.!?:;)"'\]]$/.test(prev.trimEnd())
+    ) {
+      out[out.length - 1] = `${prev.trimEnd()} ${line.trimStart()}`;
+    } else {
+      out.push(line);
+    }
+  }
+  return out.join("\n");
+}
+
+/** v0.28.24: a candidate ending in a dangling connector is a wrap/parse
+ * fragment, not a finding ("…codebase to", "…the settings and"). */
+const DANGLING_END = /\s(to|and|or|but|the|a|an|of|for|with|in|on|at|that|which|into|from|by|is|are|was|were|be|been|so|if|then|than|as|nor|yet|per|via)$/i;
+
+/** v0.28.24: cut at a clause boundary, never mid-word — the finding text IS
+ * the item's user-facing name once enqueued. */
+export function cutAtClauseBoundary(s: string, max: number): string {
+  if (s.length <= max) return s;
+  const window = s.slice(0, max);
+  const clause = Math.max(
+    window.lastIndexOf(". "),
+    window.lastIndexOf("! "),
+    window.lastIndexOf("? "),
+    window.lastIndexOf("; "),
+    window.lastIndexOf(", "),
+    window.lastIndexOf(" — "),
+    window.lastIndexOf(": "),
+  );
+  if (clause >= Math.floor(max * 0.4)) return window.slice(0, clause + 1).trimEnd();
+  const space = window.lastIndexOf(" ");
+  return (space > 0 ? window.slice(0, space) : window).trimEnd();
+}
+
 /** Scan source texts line-by-line for finding-shaped content. Code
- * spans are stripped first (v0.26.4) — findings live in prose. */
-export function extractFindings(sources: Array<{ name: string; text: string }>, max: number): Finding[] {
+ * spans are stripped first (v0.26.4) — findings live in prose. Hard-wrapped
+ * lines are joined (v0.28.24) — findings are sentence-shaped, not
+ * visual-line-shaped. `completedObjective` (v0.28.24) dedupes findings that
+ * merely restate the just-completed goal (exact-match dedupe at v0.28.16 was
+ * too narrow — duplicates arrive as prefixes/paraphrases). */
+export function extractFindings(sources: Array<{ name: string; text: string }>, max: number, completedObjective?: string): Finding[] {
   const out: Finding[] = [];
   const seen = new Set<string>();
+  const completedNorm = completedObjective ? normalizeObjective(completedObjective) : "";
   for (const { name, text } of sources) {
-    for (const line of stripCodeSpans(text).split("\n")) {
+    for (const line of unwrapHardWrappedLines(stripCodeSpans(text)).split("\n")) {
       const cls = classifyFindingText(line);
       if (!cls) continue;
-      const clean = line.trim().replace(/^[-*>\s\[\]x]+/, "").slice(0, 200);
+      const clean = cutAtClauseBoundary(line.trim().replace(/^[-*>\s\[\]x]+/, ""), 200);
       if (clean.length < 8 || seen.has(clean)) continue;
+      if (DANGLING_END.test(clean) || /[,;:\u2014-]$/.test(clean)) continue; // v0.28.24: wrap/parse fragment
+      if (completedNorm) {
+        const nf = normalizeObjective(clean);
+        if (nf.length >= 24 && (completedNorm.startsWith(nf) || nf.startsWith(completedNorm))) continue; // v0.28.24: restates the completed goal
+      }
       seen.add(clean);
       out.push({ text: clean, source: name, class: cls });
       if (out.length >= max) return out;
@@ -260,7 +327,7 @@ export function runReviewer(
     }
   }
 
-  const findings = extractFindings(deps.sources, config.maxFindingsPerReview);
+  const findings = extractFindings(deps.sources, config.maxFindingsPerReview, source.objective);
   const bugs = findings.filter((f) => f.class === "bug" || f.class === "refactor");
   const architectural = findings.filter((f) => f.class === "architectural");
   const strategic = findings.filter((f) => f.class === "strategic");

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-goal-list-loop-audit",
-  "version": "0.28.23",
+  "version": "0.28.25",
   "description": "Goal. Loop. Audit. Done. \u2014 a pi-coding-agent extension that supervises long-running work, with isolated auditor on each completion. Beat bamboozling by design: the auditor runs in a fresh session with no extensions, no skills, no editor \u2014 only the read tools needed to verify your goal.",
   "license": "MIT",
   "author": "dracon",

package/prompts/goal-loop-continuation.md

@@ -144,6 +144,8 @@ When the goal is genuinely blocked and you cannot make progress without user inp
 pause_goal({reason: "...", suggestedAction: "..."})
 ```
 
+When the user must CHOOSE between paths, use `pause_goal` with `kind="decision"`, an `options` list, and `recommended` (1-based index) — a prominent decision card renders and the user picks. **Vocabulary rules for reasons and options (v0.28.24):** reference only REAL commands — `/goal resume`, `/goal cancel`, `/goal tweak "<new text>"`, `/list remove N`, `/list next`, `/list resume`, `/loop stop`, `/loop resume` — all act on the ACTIVE goal/item; there is **no `/goal drop`** and **no command takes a goal id**. Never show goal ids (`20260729065635-gbtxsm`) in user-facing text — name the thing instead ("the active goal", "list item 'regression scan'"); ids are internal plumbing the user cannot act on.
+
 ## HARD RULES
 
 - **Do not modify the objective silently.** The objective is the user's; if it has drifted from what makes sense, use `complete_goal`'s `newObjective` at completion time, or `pause_goal` and propose a `/goal tweak` mid-flight — never just work on something else and claim the original.