pi-goal-list-loop-audit 0.23.3 → 0.23.5

package/extensions/goal-loop-core.ts

@@ -593,3 +593,37 @@ export function shouldAutoResumeOnSessionStart(reason: string | undefined, autoR
   if (autoResume === true) return true;
   return reason === "resume" || reason === "reload" || reason === "fork";
 }
+
+/**
+ * v0.23.5: normalize a drafter-supplied verification contract for the
+ * Confirm dialog AND for storage. Three cleanups, all mechanical:
+ *  1. Drop bare introducer lines ("Done when:", "Done when ALL of the
+ *     following are true:") — the dialog adds its own "Done when" header;
+ *     a model-supplied one renders doubled (field-observed) and pollutes
+ *     the shield's item list.
+ *  2. Strip a glued "Done when: " prefix on a content line.
+ *  3. Renumber bullet/numbered lines sequentially ("1.", "2.", ...) so the
+ *     dialog reads as a checklist and reject-feedback can cite item
+ *     numbers. Non-bullet prose lines pass through untouched.
+ */
+export function normalizeDraftContract(raw: string): string {
+  const lines = raw
+    .trim()
+    .split("\n")
+    .map((l) => l.trim())
+    .filter((l) => !/^(?:done when|verified when|verify|verification)\b[^:]*:\s*$/i.test(l))
+    .map((l) => l.replace(/^(?:done when|verified when)\s*:\s+/i, ""))
+    .filter((l) => l.length > 0);
+  let n = 0;
+  return lines
+    .map((l) => {
+      const m = l.match(/^(?:[-*•]\s+|\d+[.)]\s+)(.+)$/);
+      return m ? `${++n}. ${m[1]}` : l;
+    })
+    .join("\n");
+}
+
+/** Count the numbered checklist items in a normalized contract. */
+export function draftContractItemCount(normalized: string): number {
+  return normalized.split("\n").filter((l) => /^\d+\.\s/.test(l)).length;
+}

package/extensions/goal-loop-shield.ts

@@ -23,7 +23,16 @@ export function contractItems(contract: string): string[] {
     .filter((l) => l.length > 0)
     // Boundary lines ("Out of scope: ...") constrain the auditor's judgment;
     // they are not deliverables and have no evidence to quote (v0.22.6).
-    .filter((l) => !/^out of scope\b/i.test(l));
+    .filter((l) => !/^out of scope\b/i.test(l))
+    // Preamble lines are not checkable items (v0.23.4, darklord field bug:
+    // "Done when ALL of the following are true:" survived as an "item" —
+    // the prefix strip only fires when a colon directly follows "done
+    // when" — and the shield then blocked TWO genuine approvals forever,
+    // because no evidence can reference a preamble). Two mechanical
+    // predicates: a line still ending in a colon introduces a list, and a
+    // "(done when) (all of) the following ..." line IS the introducer.
+    .filter((l) => !l.endsWith(":"))
+    .filter((l) => !/^(?:done when\s+)?(?:all of\s+)?the following\b/i.test(l));
 }
 
 export interface RegressionShieldResult {

package/extensions/loops/goal.ts

@@ -53,6 +53,8 @@ import {
   newGoalId,
   nowIso,
   piGlaDir,
+  normalizeDraftContract,
+  draftContractItemCount,
   readState,
   renderGoalMarkdown,
   shouldAutoResumeOnSessionStart,
@@ -1630,8 +1632,10 @@ function registerAgentTools(pi: any, ctx: ExtensionContext): void {
         }
         return { content: [{ type: "text", text: `${n} items confirmed and queued (${listQueue().length} waiting).` }], details: {} };
       }
-      const contractBlock = p.verificationContract?.trim()
-        ? `\n\nDone when:\n${p.verificationContract.trim()}`
+      const normContract = p.verificationContract?.trim() ? normalizeDraftContract(p.verificationContract) : "";
+      const checkCount = normContract ? draftContractItemCount(normContract) : 0;
+      const contractBlock = normContract
+        ? `\n\nDone when${checkCount > 0 ? ` — ${checkCount} check${checkCount === 1 ? "" : "s"}` : ""}:\n${normContract}`
         : "\n\n(No verification contract — the auditor will infer done-criteria from the objective. Consider adding one.)";
       // v0.22.6: a list draft that will activate immediately must SAY so in
       // the Confirm dialog — "I started a list and ended up with a running
@@ -1657,7 +1661,7 @@ function registerAgentTools(pi: any, ctx: ExtensionContext): void {
       }
       const confirmedTarget = draftingTarget;
       draftingTarget = null;
-      const full = p.objective.trim() + (p.verificationContract?.trim() ? `\nDone when:\n${p.verificationContract.trim()}` : "");
+      const full = p.objective.trim() + (normContract ? `\nDone when:\n${normContract}` : "");
       // List drafting: the confirmed contract goes into the QUEUE, not active.
       if (confirmedTarget === "list") {
         const extracted = extractVerificationContract(full);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-goal-list-loop-audit",
-  "version": "0.23.3",
+  "version": "0.23.5",
   "description": "Goal. Loop. Audit. Done. — 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 — only the read tools needed to verify your goal.",
   "license": "MIT",
   "author": "dracon",

package/prompts/goal-loop-draft.md

@@ -19,6 +19,13 @@ request into a **confirmed goal contract**. Do NOT start substantive work yet.
    - **verification contract** — how an independent auditor can tell it is
      done, as checkable items (commands, file states, test outcomes).
      Strongly recommended; without it the auditor infers from the objective.
+     Write 3–8 mechanical checks, one per line, each verifiable with ONE
+     command or file check. The auditor must quote raw evidence for EVERY
+     item — a 17-item contract means a slow, expensive audit and more
+     regression-shield friction. Verify the artifact's integrity (the doc
+     exists, the table has N rows, the gates pass), not every sub-part.
+     Do NOT prefix the contract with "Done when:" — the Confirm dialog
+     adds that header itself.
    - **boundaries** — what is explicitly out of scope (fold into the
      objective text).
 4. Keep grilling until the objective and success criteria are concrete