pi-goal-list-loop-audit 0.38.34 → 0.38.36

package/CHANGELOG.md

@@ -1,5 +1,29 @@
 # Changelog
 
+## 0.38.36 — human-voice terminal summaries (2026-09-08)
+
+### Fixed
+
+- **Summaries read as machine receipts (field 20260908_223522/223523):** `clipSummaryValue` now cuts at clause boundaries and strips stranded punctuation (`…append/replace Apply…`, never `…playlist auto-add,…`); the counts line is verdict-only (`— audit: auditor approved (1 verdict).`, raw run stats stay in the archive); the AUDIT PENDING notice states facts instead of agent imperatives. Agent-written summaries get Codex-shape guidance in the continuation prompt (outcome-first, verifiable bullets, clause cuts). Researched from real Codex transcripts + agy binary strings; user chose Codex-style chat shape.
+- **Stale pins updated deliberately:** `— run:` telemetry assertions, `/detached auditor queued/` wordings, continuation-payload byte fixtures (+344/payload, growth exactly linear).
+
+### Verification
+
+- Full `release:check`: 2038 pass / 2 skip / 0 fail across 204 files; `tsc --noEmit` clean.
+- Evidence: `audit/HUMAN-VOICE-SUMMARY-2026-09-08.md`.
+
+## 0.38.35 — closed card + smart summary (2026-09-08)
+
+### Fixed
+
+- **Card no longer reads cut off (field 2026-09-08 220808):** the head objective now cuts at the last clause boundary inside the budget instead of mid-word (`truncateObjective`; character cut remains the fallback); the steady-state `model: primary … · inherited from session` row is dropped — it restated pi's own status line two rows below; and every card closes — the final `├─`/`│` row becomes `└─` whatever built it (active action row, lone orphan worker row). Provenance still renders on any news (pin, fallbacks, skips, failover); `/goal status` keeps the full chain.
+- **Stale card tests updated deliberately, purpose preserved:** the v0.33.0 slim-card `├─` tail, the v0.38.28 lone-row close, the loop-kind test's incidental model assertion, the narrow-width truncation fixture (now on a pinned row), and the worker-before-footer invariant (detail never lands *after* a footer).
+
+### Verification
+
+- Full `release:check`: 2037 pass / 2 skip / 0 fail across 204 files; `tsc --noEmit` clean.
+- Evidence: `audit/UI-CARD-CLOSED-2026-09-08.md`.
+
 ## 0.38.34 — completion communication: fair replay rotation + stale-test refresh (2026-09-08)
 
 ### Fixed

package/docs/INDEX.md

@@ -18,7 +18,7 @@ For shipped docs, the relevant entry points are:
   failback; v0.35.9 hardened cross-version npm tarball checks; v0.35.10
   handles multi-entry npm dry-run reports; v0.35.11 accepts both npm report
   shapes; v0.35.12 supports npm 12's keyed pack reports; v0.35.13 fixes stale-API recovery loops.
-  v0.35.14–v0.38.34 continue through the supervisor freeze (`/glla pause`),
+  v0.35.14–v0.38.36 continue through the supervisor freeze (`/glla pause`),
   load hold, auditor picker parity, Windows launch fix, zombie-watchdog
   subagent carve-out, due-wait backstop, the `/glla agents` visibility panel,
   durable state-root selection, blank-until-resume auditor context, frozen

package/extensions/completion-summary.ts

@@ -78,18 +78,33 @@ export function isUsefulCompletionSummary(text: string | undefined): boolean {
  * Missing labels are retained as `not recorded` so a compact notification
  * cannot accidentally imply evidence that the durable recap did not contain.
  */
-/** Cut a summary value at a word boundary, never mid-word (the
- * Screenshot_20260903_204003/204005 complaint: `0 o…`, `qu…`, `belo…`).
- * Falls back to a hard cut only when the head holds no space past the
- * halfway mark — a long token such as a commit hash must not eviscerate
- * the whole value. Short values pass through untouched (no `…`). */
+/** Cut a summary value at a clause boundary, never mid-word (field
+ * complaints 2026-09-03 `0 o…` and 2026-09-08 `playlist auto-add,…`:
+ * a word-boundary cut that strands dangling punctuation still reads as
+ * clipping, not summarizing). Prefer the last clause boundary
+ * (`, ; : · — – ( [`) past a floor so the cut reads intentional; fall
+ * back to the word break, then to a hard cut only when the head holds
+ * no space past the halfway mark — a long token such as a commit hash
+ * must not eviscerate the whole value. Trailing punctuation is stripped
+ * before the ellipsis. Short values pass through untouched (no `…`). */
 export function clipSummaryValue(value: string, limit: number): string {
   const clean = value.replace(/\s+/g, " ").trim();
   const capped = Number.isFinite(limit) ? Math.max(8, Math.floor(limit)) : 72;
   if (clean.length <= capped) return clean;
   const head = clean.slice(0, capped - 1);
+  const floor = Math.max(16, Math.floor((capped - 1) * 0.4));
+  let boundary = -1;
+  for (let i = 0; i < head.length; i++) {
+    if (/[,;:·—–(\[]/.test(head[i]!)) boundary = i;
+  }
+  if (boundary >= floor) {
+    const cut = head.slice(0, boundary).replace(/[,;:·—–(\[\s]+$/u, "");
+    if (cut.length >= Math.min(floor, 16)) return `${cut}…`;
+  }
   const space = head.lastIndexOf(" ");
-  const kept = (space > capped / 2 ? head.slice(0, space) : head).trimEnd();
+  const kept = (space > capped / 2 ? head.slice(0, space) : head)
+    .trimEnd()
+    .replace(/[,;:·—–(\[]$/u, "");
   return `${kept}…`;
 }
 
@@ -192,15 +207,12 @@ export function buildApprovalChatLines(notice: {
   ];
 }
 
-/** v0.38.25: the audit-goal counts line. Compact execution + audit proof
- * built ONLY from durable goal state — turns/file-writes/bash from
- * telemetry, verdict counts from auditHistory. Absent facts are named as
- * absent, never invented (the recorded-facts honesty rule). */
+/** v0.38.25: the audit-goal counts line — verdict proof ONLY, built from
+ * durable goal state. Raw run stats (turns/file-writes/bash calls) read as
+ * a machine receipt in user chat (field 2026-09-08 223522/223523); the
+ * full execution evidence stays in the six-label archive record. Absent
+ * facts are named as absent, never invented (recorded-facts honesty). */
 export function buildAuditCountsLine(goal: Goal, auditNote?: string): string {
-  const telemetry = goal.telemetry;
-  const run = telemetry
-    ? `${telemetry.turns} turns · ${telemetry.fileWrites} file writes · ${telemetry.bashCalls} bash calls`
-    : "no execution telemetry was recorded";
   const history = goal.auditHistory ?? [];
   const latest = history.length > 0 ? history[history.length - 1] : undefined;
   const audit = auditNote ?? (latest === undefined
@@ -209,7 +221,7 @@ export function buildAuditCountsLine(goal: Goal, auditNote?: string): string {
       const verdict = latest.approved ? "approved" : latest.impossible ? "impossible" : latest.disapproved ? "disapproved" : "no verdict";
       return `auditor ${verdict} (${history.length} verdict${history.length === 1 ? "" : "s"})`;
     })());
-  return `— run: ${run} · ${audit}.`;
+  return `— audit: ${audit}.`;
 }
 
 export interface TerminalApprovalRenderInput {

package/extensions/goal-loop-display.ts

@@ -83,6 +83,35 @@ export function truncateCells(s: string, max: number): string {
   return `${out}…`;
 }
 
+/** Clause-aware objective shortening for the card head. A raw character cut
+ * ends objectives mid-word ("…restyle the …") which reads as a clipped
+ * card rather than a summary (field 2026-09-08 220808). When the text
+ * overruns, prefer cutting at the last clause boundary (: ; · — – ( [)
+ * inside the budget so the head reads as an intentional summary; fall back
+ * to the character cut when no boundary clears the floor. Display-only. */
+export function truncateObjective(s: string, max: number): string {
+  const clean = compactDisplayText(s);
+  if (tuiVisibleWidth(clean) <= max) return clean;
+  if (max <= 1) return truncateCells(clean, max);
+  const budget = max - 1;
+  const floor = Math.max(16, Math.floor(budget * 0.4));
+  let out = "";
+  let w = 0;
+  let boundaryLen = -1;
+  for (const ch of clean) {
+    const cw = tuiVisibleWidth(ch);
+    if (w + cw > budget) break;
+    out += ch;
+    w += cw;
+    if (/[:;·—–(\[]/.test(ch)) boundaryLen = out.length;
+  }
+  if (boundaryLen > 0) {
+    const cut = [...out].slice(0, boundaryLen).join("").replace(/[:;·—–(\[\s]+$/u, "");
+    if (tuiVisibleWidth(cut) >= Math.min(floor, 16)) return `${cut}…`;
+  }
+  return `${out}…`;
+}
+
 function displayPauseReason(reason: string): string {
   return compactDisplayText(sanitizeProviderDisplayText(reason));
 }
@@ -353,26 +382,34 @@ function compactMainModelRecoveryLine(recovery: MainModelRecovery | undefined, c
 
 function modelProvenanceLines(provenance: ModelProvenanceDisplay | undefined, width?: number): string[] {
   if (!provenance) return [];
+  const primary = typeof provenance.primary === "string" ? provenance.primary.trim() : "";
+  const fallbacks = uniqueModelRefs(provenance.fallbackRefs);
+  const skipped = uniqueModelRefs(provenance.skippedForbiddenRefs);
+  const handledTurn = typeof provenance.handledTurn === "string" ? provenance.handledTurn.trim() : "";
+  const handledTurnNews = !!handledTurn && handledTurn.toLowerCase() !== (primary || "").toLowerCase();
+  const handledAudit = typeof provenance.handledAudit === "string" ? provenance.handledAudit.trim() : "";
+  // The card sits just above pi's own status line, which already names the
+  // session model. A lone `model: primary X · inherited from session` row
+  // restates that indicator for zero new information while deepening the
+  // tree (field 2026-09-08 220808). Show the block only when it carries
+  // news: a pinned selection, fallbacks, skips, or a failover that
+  // handled work. `/goal status` keeps the full chain regardless.
+  if (primary && provenance.primarySource !== "pinned" && fallbacks.length === 0 && skipped.length === 0 && !handledTurnNews && !handledAudit) return [];
   const budget = budgetFor(width, 3, 60);
   const lines: string[] = [];
-  const primary = typeof provenance.primary === "string" ? provenance.primary.trim() : "";
   if (primary) {
     const source = provenance.primarySource === "pinned" ? "pinned" : "inherited from session";
     lines.push(`model: primary ${truncate(primary, budget)} · ${source}`);
   }
-  const fallbacks = uniqueModelRefs(provenance.fallbackRefs);
   if (fallbacks.length > 0) lines.push(`fallbacks: ${truncate(fallbacks.join(" → "), budget)}`);
-  const skipped = uniqueModelRefs(provenance.skippedForbiddenRefs);
   if (skipped.length > 0) lines.push(`skipped forbidden: ${truncate(skipped.join(", "), budget)}`);
-  const handledTurn = typeof provenance.handledTurn === "string" ? provenance.handledTurn.trim() : "";
   // A normal turn is handled by the configured primary model. Repeating the
   // same ref immediately below it adds no information and, when it is the
   // last row, leaves a misleading continuation glyph in the card. Keep the
   // row only when a recovery/failover actually handled the turn.
-  if (handledTurn && handledTurn.toLowerCase() !== primary.toLowerCase()) {
+  if (handledTurnNews) {
     lines.push(`handled turn: ${truncate(handledTurn, budget)}`);
   }
-  const handledAudit = typeof provenance.handledAudit === "string" ? provenance.handledAudit.trim() : "";
   if (handledAudit) {
     const source = provenance.handledAuditSource?.trim() ? modelSourceLabel(provenance.handledAuditSource) : "";
     const via = source ? ` · via ${truncate(source, 24)}` : "";
@@ -1450,15 +1487,18 @@ export function buildWidgetLines(state: State, audit?: AuditDisplayProgress | nu
       withAgents = orphanHead ? [orphanHead, ...agentLines] : [...agentLines];
     }
   }
-  // Keep the final tree row visually closed when the compact judgment block
-  // ends without a queue footer. Without this, its last `├─` reads like a
-  // missing continuation and makes an already dense card feel unfinished.
-  if (withAgents && inner && detailedAgents.length === 0 && extras?.durableDeferRecommendation && withAgents.length > 0) {
+  // The card always reads as a finished block: a final `├─`/`│` row
+  // promises continuation rows that never come (field 2026-09-08 220808 —
+  // the active card ended on its action row, reading as cut off). Close
+  // the tail whatever built it; rows that already close (`└─`, the head,
+  // the spacer) are untouched. Tree prefixes are literal — paint wraps
+  // content, never the glyph — so this holds themed too.
+  if (withAgents && withAgents.length > 0) {
     const tailIndex = withAgents.length - 1;
     const tail = withAgents[tailIndex]!;
-    if (tail.startsWith("├─ ") || tail.startsWith("│ ")) {
-      withAgents[tailIndex] = `└─ ${tail.slice(3)}`;
-    }
+    const m = tail.match(/^(├─ |│  |│ )/);
+    const prefix = m?.[1] ?? "";
+    if (prefix) withAgents[tailIndex] = `└─ ${tail.slice(prefix.length)}`;
   }
   // v0.28.6 (E1): a persistence failure outranks everything — first line,
   // on every render, until a write lands again.
@@ -1663,7 +1703,7 @@ function goalLines(g: Goal, state: State, audit: AuditDisplayProgress | null | u
   const headIcon = headLive
     ? paint(theme, headLive.band === "fresh" ? "success" : headLive.band === "aging" ? "warning" : "error", headLive.breath)
     : icon;
-  const head = `${headIcon} ${truncate(displayObjective(g.objective), objBudget)} ${paint(theme, "dim", "·")} ${segsText}`;
+  const head = `${headIcon} ${truncateObjective(displayObjective(g.objective), objBudget)} ${paint(theme, "dim", "·")} ${segsText}`;
   const lines = [head];
   // v0.38.8: durable verdict tally as a first-class card row — the widget
   // is the glance surface, and stored verdicts are the progress evidence

package/extensions/loops/goal-tools.ts

@@ -1876,7 +1876,7 @@ function registerAgentTools(pi: any): void {
         if (notifyFailure) current.ui.notify(`Completion auditor worker failed to settle (infrastructure, not a verdict). The stored claim is safe; ${activeGoalSurfaceCommand("resume")} retries it.`, "warning");
       });
       return {
-        content: [{ type: "text", text: `AUDIT PENDING — nonterminal. Completion claim persisted; detached auditor queued (model: ${via ?? "setting"}). The goal is not approved or complete. Do not claim completion or give a final success summary. Do not wait or poll: GLLA will post the concrete final summary after verified approval and durable archive; rejection resumes work.` }],
+        content: [{ type: "text", text: `AUDIT PENDING — nonterminal. The completion claim is stored and a detached auditor is settling it (model: ${via ?? "setting"}). The goal is not approved or complete yet. The final summary posts here after verified approval and durable archive; rejection resumes work automatically. Ending this turn now is correct — no waiting or polling needed.` }],
         details: { status: "audit-pending", terminal: false },
         // End this tool batch, not the goal; the detached worker owns settlement.
         terminate: true,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-goal-list-loop-audit",
-  "version": "0.38.34",
+  "version": "0.38.36",
   "description": "Mission control for autonomous pi: interview-drafted goals, an audited task queue, and forever-loops (metric, spec, project-audit) that run for hours. A detached extension-less auditor process re-verifies every completion with raw evidence without holding the main pi turn; confirmed drafts, decision pauses and consent gates keep you in charge.",
   "license": "AGPL-3.0-only",
   "author": "dracon",

package/prompts/goal-loop-continuation.md

@@ -196,3 +196,5 @@ The orchestrator's backstop is the stall watchdog: three consecutive turns with
 ### Completion communication
 
 `complete_goal` submits a nonterminal claim, not a completion verdict. While the detached audit is pending, do not say the goal is done/complete/approved or present a final success summary. Do not wait or poll. If a response is needed, give one brief pending-status sentence. GLLA posts the concrete outcome-first summary into chat after verified approval and durable archive, without needing another user prompt. Do not duplicate that summary. A rejected claim remains unfinished and repair work continues.
+
+The agent-written completionSummary is user-facing evidence, not a step log: open with the outcome in plain words, then one verifiable result per label with its proof inline (test counts, file paths, commit refs) — never raw run stats. Cut values at clause boundaries, never mid-word; state what was deliberately left out when it matters.