pi-goal-list-loop-audit 0.31.0 → 0.31.2

package/extensions/goal-loop-forever.ts

@@ -547,7 +547,16 @@ export function listAuditFanoutItemText(finding: string): string {
   return `Fix audit finding: ${finding} — Done when: the fix is committed on the current branch with the repo's configured identity, and this finding's box in ${AUDIT_FINDINGS_REL} is checked ("- [x] … — fixed in <commit>").`;
 }
 
+/** v0.31.1: stacking-detection markers (junk-runner 2026-07-31: a held
+ * one-shot audit goal + a running audit loop = two stacked audit initiatives
+ * — the held one-shot read as "stalled" for 8h while the loop did all the
+ * work, and the agent conflated them). The guards in goal.ts match on these;
+ * the unit tests pin that the built targets still contain them.
+ */
+export const GOAL_AUDIT_ONESHOT_MARKER = "Run ONE project audit pass and leave the project in a known state";
+export const LOOP_AUDIT_MARKER = "iteration by iteration — FIX-FIRST";
+
 export function projectAuditTarget(focus?: string): string {
   const scope = focus && focus.trim() ? focus.trim() : "the whole project";
-  return `Run ONE project audit pass and leave the project in a known state. Scope: ${scope}. (1) Run a FRESH audit pass over the codebase — spawn Explore subagents for breadth — hunting real problems: bugs, broken flows, regressions, drift between docs and code, dead code, security holes. Not style nits, not speculative refactors. (2) Append every NEW finding to ${AUDIT_FINDINGS_REL} (create the file on the first finding; append-only — never delete, rewrite, or reorder existing lines; never re-report a finding already listed), classified: "- [ ] FIX: SEVERITY: short description (file:line)" for bugs and polish — whether to fix these is NOT a decision — and "- [?] DECIDE: short description (what the choice is, what each side costs)" for direction, trade-offs, and scope questions where two reasonable answers exist. (3) Fix every NEW FIX finding from this pass — real fixes, committed with the repo's configured identity on the current branch (no invented identities or branches) — then check the box: "- [x] … — fixed in <commit>". (4) Change NOTHING for DECIDE findings — present them in the completion report instead. (5) Honesty law: never fabricate findings to look busy; never check a box without the fix commit existing; never silently turn a DECIDE into a fix. Done when: the audit pass is complete, every new FIX finding has a fix commit and a checked box in ${AUDIT_FINDINGS_REL}, and every DECIDE finding is listed in the file and presented in the completion report.`;
+  return `${GOAL_AUDIT_ONESHOT_MARKER}. Scope: ${scope}. (1) Run a FRESH audit pass over the codebase — spawn Explore subagents for breadth — hunting real problems: bugs, broken flows, regressions, drift between docs and code, dead code, security holes. Not style nits, not speculative refactors. (2) Append every NEW finding to ${AUDIT_FINDINGS_REL} (create the file on the first finding; append-only — never delete, rewrite, or reorder existing lines; never re-report a finding already listed), classified: "- [ ] FIX: SEVERITY: short description (file:line)" for bugs and polish — whether to fix these is NOT a decision — and "- [?] DECIDE: short description (what the choice is, what each side costs)" for direction, trade-offs, and scope questions where two reasonable answers exist. (3) Fix every NEW FIX finding from this pass — real fixes, committed with the repo's configured identity on the current branch (no invented identities or branches) — then check the box: "- [x] … — fixed in <commit>". (4) Change NOTHING for DECIDE findings — present them in the completion report instead. (5) Honesty law: never fabricate findings to look busy; never check a box without the fix commit existing; never silently turn a DECIDE into a fix. Done when: the audit pass is complete, every new FIX finding has a fix commit and a checked box in ${AUDIT_FINDINGS_REL}, and every DECIDE finding is listed in the file and presented in the completion report.`;
 }

package/extensions/goal-settings.ts

@@ -112,9 +112,10 @@ export interface Settings {
 }
 
 export const DEFAULT_SETTINGS: Settings = {
-  // Unset = follow the pi session thinking level (user selects thinking in
-  // pi, auditor follows), floor "high" — the auditor is the verification
-  // gate, depth is worth more there than speed. /glla thinking= overrides.
+  // Unset = "high" at the call site (v0.31.2). The auditor is the
+  // verification gate: its depth must NOT ride the session's coding-speed
+  // thinking dial (user 2026-07-31: "we should also select its thinking
+  // level — we don't keep switching it"). /glla thinking= overrides.
   auditorThinkingLevel: undefined,
   // v0.24.6: subagents inherit the session model by default — one quota
   // pool, no surprise 403s from a pinned default agent's provider.

package/extensions/loops/goal.ts

@@ -154,6 +154,7 @@ import {
 } from "../settings-menu.js";
 import {
   buildModelPickItems,
+  pickDiverseAuditorModel,
   ModelPickerComponent,
   type ModelPickItem,
 } from "../model-picker.js";
@@ -176,6 +177,8 @@ import {
   AUDIT_FINDINGS_REL,
   projectAuditTarget,
   LIST_AUDIT_COLLECT_MARKER,
+  GOAL_AUDIT_ONESHOT_MARKER,
+  LOOP_AUDIT_MARKER,
   listAuditCollectTarget,
   parseAuditFindingsForFanout,
   listAuditFanoutItemText,
@@ -1378,7 +1381,7 @@ async function retryStoredCompletionAudit(ctx: ExtensionContext, origin: "quota-
           completionSummary: claim.completionSummary,
           verificationSummary: claim.verificationSummary,
           model: auditorModel,
-          thinkingLevel: settings.auditorThinkingLevel ?? getSessionThinkingLevel(),
+          thinkingLevel: settings.auditorThinkingLevel ?? "high",
           onProgress: (progress) => {
             latestAuditProgress = { currentTool: progress.currentTool, label: progress.label, elapsedMs: progress.elapsedMs, lastEventAt: Date.now() };
             refreshUI(liveCtx);
@@ -1706,6 +1709,15 @@ async function cmdGoal(args: string, ctx: ExtensionContext): Promise<void> {
     // DECIDE findings presented, untouched. Runs as a normal goal through
     // cmdSet — the isolated auditor verifies the finish line.
     if (route.name === "audit") {
+      // v0.31.1: an active audit loop already owns auditing here — the
+      // one-shot duplicates it (same stacking confusion as junk-runner).
+      if (state.loop?.active && state.loop.target.includes(LOOP_AUDIT_MARKER)) {
+        appendLedger(ctx.cwd, "audit_stack_warn", { have: "loop", starting: "goal" });
+        ctx.ui.notify(
+          "An audit loop is already running here — a one-shot /goal audit duplicates its work. /loop status to see it; /loop stop first if you want the one-shot instead.",
+          "warning",
+        );
+      }
       return cmdSet(projectAuditTarget(route.rest || undefined), ctx, true);
     }
     // v0.28.27 (renamed /goal audit → /goal verify in v0.29.8): run the
@@ -2158,6 +2170,12 @@ async function cmdList(args: string, ctx: ExtensionContext): Promise<void> {
     // list drains them fix by fix, each with its own isolated audit.
     // Distinct from /goal audit (fix-in-pass, one audited unit) and
     // /loop audit (forever fix-first cadence).
+    // v0.31.1: an active audit loop is already draining this findings file —
+    // a collect pass would double-hunt the same ground.
+    if (state.loop?.active && state.loop.target.includes(LOOP_AUDIT_MARKER)) {
+      appendLedger(ctx.cwd, "audit_stack_warn", { have: "loop", starting: "list" });
+      ctx.ui.notify("An audit loop is already draining findings here — /list audit would double-hunt the same ground. /loop status to see it.", "warning");
+    }
     const objective = listAuditCollectTarget(rest || undefined);
     const n = enqueueItems(ctx, [objective], "/list audit");
     if (n === 0) return; // zombie-twin guard already explained itself
@@ -3021,6 +3039,18 @@ async function cmdLoop(args: string, ctx: ExtensionContext): Promise<void> {
       ctx.ui.notify("A goal is active — /goal cancel or /goal pause it before starting a loop.", "warning");
       return;
     }
+    // v0.31.1: a paused/active one-shot audit goal + this loop = two stacked
+    // audit initiatives (junk-runner 2026-07-31: the held one-shot read as
+    // "stalled" for 8h while the loop did all the work — the agent conflated
+    // them and proposed completing the goal for the loop's work). Warn, name
+    // the supersession, don't block — the user's agency, the user's call.
+    if (state.goal && state.goal.objective.includes(GOAL_AUDIT_ONESHOT_MARKER)) {
+      appendLedger(ctx.cwd, "audit_stack_warn", { have: "goal", starting: "loop", goalStatus: state.goal.status });
+      ctx.ui.notify(
+        `Heads up: a ${state.goal.status} one-shot audit goal exists in this session — the audit loop SUPERSEDES it (one pass + fixes IS the loop's job). /goal cancel clears it; one audit initiative per session.`,
+        "warning",
+      );
+    }
     if (isLoopActive()) {
       ctx.ui.notify("A loop is already active. /loop stop first.", "warning");
       return;
@@ -3159,7 +3189,7 @@ function registerAgentTools(pi: any, ctx: ExtensionContext): void {
           completionSummary: p.completionSummary,
           verificationSummary: p.verificationSummary,
           model: auditorModel,
-          thinkingLevel: settings.auditorThinkingLevel ?? getSessionThinkingLevel(),
+          thinkingLevel: settings.auditorThinkingLevel ?? "high",
           signal: signal ?? undefined,
           onProgress: (progress) => {
             latestAuditProgress = {
@@ -4178,18 +4208,6 @@ function registerAgentTools(pi: any, ctx: ExtensionContext): void {
  * the thinking level the user selected in pi; if none is set, audits run at
  * "high" — the auditor is the verification gate, depth beats speed there.
  */
-function getSessionThinkingLevel(): "off" | "minimal" | "low" | "medium" | "high" | "xhigh" {
-  try {
-    const level = extensionApi?.getThinkingLevel?.();
-    if (level && ["off", "minimal", "low", "medium", "high", "xhigh"].includes(level)) {
-      return level as "off" | "minimal" | "low" | "medium" | "high" | "xhigh";
-    }
-  } catch {
-    // fall through to the floor
-  }
-  return "high";
-}
-
 /**
  * Resolve the auditor model (v0.6.2). The principle: **the user selects the
  * model in pi; the auditor uses it.** The plugin never picks a model itself.
@@ -4206,6 +4224,24 @@ function getSessionThinkingLevel(): "off" | "minimal" | "low" | "medium" | "high
 function resolveAuditorModel(ctx: ExtensionContext, ref?: string): { model: any; error?: string; via?: string } {
   if (ref && ref.trim()) {
     const trimmed = ref.trim();
+    // v0.31.2: "diverse" — the cross-vendor auditor (user design 2026-07-31:
+    // "there is benefit to have a different auditor"). Independent blind
+    // spots + a separate quota pool; the fresh auditor session shares no
+    // prompt cache anyway, so cross-vendor costs nothing extra.
+    if (trimmed.toLowerCase() === "diverse") {
+      const sessionModel = ctx.model as any;
+      const available = ctx.modelRegistry
+        .getAvailable()
+        .filter((m: any) => ctx.modelRegistry.hasConfiguredAuth(m));
+      const pick = pickDiverseAuditorModel(available, sessionModel?.provider);
+      if (pick) return { model: pick, via: "diverse" };
+      if (sessionModel) {
+        appendLedger(ctx.cwd, "auditor_model_fallback", { configured: "diverse", reason: "no configured-auth model outside the session's provider" });
+        ctx.ui.notify("Auditor model \"diverse\" found no model outside the session's provider — falling back to the session model (LOUD fallback; pin one via /glla → Auditor model).", "warning");
+        return { model: sessionModel, via: "session-fallback" };
+      }
+      return { model: undefined, error: "auditorModel \"diverse\": no configured-auth model at all" };
+    }
     // v0.29.17: an unavailable configured model (unknown id, or a provider
     // with no configured auth) falls back LOUDLY to the session model —
     // user request: "fall back to the session if unavailable". The v0.9.12
@@ -4326,6 +4362,7 @@ async function promptModelRef(
   ctx: ExtensionContext,
   title: string,
   emptyLabel: string,
+  extraTop: ModelPickItem[] = [],
 ): Promise<{ kind: "session" } | { kind: "ref"; ref: string } | undefined> {
   if (typeof (ctx.ui as { custom?: unknown }).custom !== "function" || !ctx.modelRegistry) {
     const v = await ctx.ui.input(title, "provider/model-id — empty keeps the default");
@@ -4337,7 +4374,7 @@ async function promptModelRef(
   const models = ctx.modelRegistry
     .getAvailable()
     .filter((m: any) => ctx.modelRegistry.hasConfiguredAuth(m));
-  const items = buildModelPickItems(models, sessionLabel);
+  const items = buildModelPickItems(models, sessionLabel, extraTop);
   const pick = await ctx.ui.custom<ModelPickItem | undefined>((tui, theme, keybindings, done) => {
     return new ModelPickerComponent({ title, items }, () => tui.requestRender(), theme, keybindings, done);
   });
@@ -4389,10 +4426,21 @@ export async function handleSettingChoice(id: string, ctx: ExtensionContext): Pr
       return;
     }
     case "auditorModel": {
-      const pick = await promptModelRef(ctx, "Auditor model override", "provider/model-id — empty keeps the pi session model");
+      // v0.31.2: "diverse" sits at the top — the cross-vendor auditor
+      // (independent blind spots + a separate quota pool from the coding
+      // session). Picked like a model; resolution happens per-audit so
+      // provider availability is re-checked every time.
+      const diverseItem: ModelPickItem = {
+        kind: "model",
+        ref: "diverse",
+        label: "diverse — cross-vendor auditor: a different provider than the session, picked fresh per audit (Recommended)",
+        searchText: "diverse cross vendor independent different provider quota strategy recommended",
+      };
+      const pick = await promptModelRef(ctx, "Auditor model override", "provider/model-id — empty keeps the pi session model", [diverseItem]);
       if (pick === undefined) return;
       saveSettings("global", ctx.cwd, { auditorModel: pick.kind === "session" ? undefined : pick.ref });
       if (pick.kind === "session") ctx.ui.notify("Auditor model override cleared — the auditor follows the pi session model.", "info");
+      if (pick.kind === "ref" && pick.ref === "diverse") ctx.ui.notify("Auditor model: DIVERSE — each audit picks a configured model outside the session's provider (deepseek ↔ MiniMax first).", "info");
       return;
     }
     case "auditorThinkingLevel": {
@@ -5834,11 +5882,24 @@ export default function (pi: ExtensionAPI): void {
         const isListItem = state.goal.policy === "list";
         const resumeCmd = isListItem ? "/list resume" : "/goal resume";
         const resumeHint = `${resumeCmd} to continue${queued > 0 ? ` (+${queued} waiting in the list)` : ""} · /glla autoresume=on to auto-resume on load (global setting)`;
+        // v0.31.1: name the supersession — a held one-shot audit whose work a
+        // live audit loop now owns reads as "stalled" for HOURS otherwise
+        // (junk-runner: 8h21m of "held for explicit resume" on a goal the
+        // loop had superseded). The widget surface must say so.
+        const auditSuperseded =
+          state.goal.objective.includes(GOAL_AUDIT_ONESHOT_MARKER) &&
+          !!state.loop &&
+          (state.loop.active || state.loop.stopReason === HELD_ON_RESTORE) &&
+          !!state.loop.target?.includes(LOOP_AUDIT_MARKER);
         updateGoal({
           status: "paused",
           pauseKind: "blocked",
-          pauseReason: "restored on session load — held for explicit resume",
-          pauseSuggestedAction: resumeHint,
+          pauseReason: auditSuperseded
+            ? "restored on session load — SUPERSEDED by the audit loop in this session"
+            : "restored on session load — held for explicit resume",
+          pauseSuggestedAction: auditSuperseded
+            ? `/goal cancel clears it (the loop already owns the audit) · ${resumeHint} if you disagree`
+            : resumeHint,
         }, ctx);
         ctx.ui.notify(
           `${isListItem ? "List item" : "Goal"} held on restore: ${state.goal.objective.slice(0, 70)}${queued > 0 ? ` (+${queued} waiting in the list)` : ""} — ${resumeCmd} to continue.`,

package/extensions/model-picker.ts

@@ -41,7 +41,7 @@ export interface RegistryModelLike {
 /** Build the picker's static item list from registry models (already
  * filtered to configured-auth providers by the caller). Session row first,
  * manual-entry row last; models sorted by provider then id. */
-export function buildModelPickItems(models: RegistryModelLike[], sessionLabel: string): ModelPickItem[] {
+export function buildModelPickItems(models: RegistryModelLike[], sessionLabel: string, extraTop: ModelPickItem[] = []): ModelPickItem[] {
   const sorted = [...models].sort((a, b) =>
     a.provider === b.provider ? a.id.localeCompare(b.id) : a.provider.localeCompare(b.provider),
   );
@@ -51,6 +51,7 @@ export function buildModelPickItems(models: RegistryModelLike[], sessionLabel: s
       label: `session model (${sessionLabel}) — clear the override`,
       searchText: "session model default clear override follow",
     },
+    ...extraTop,
     ...sorted.map((m) => {
       const ref = `${m.provider}/${m.id}`;
       return {
@@ -206,3 +207,47 @@ export class ModelPickerComponent {
 
 // Re-export for callers that only need the width helper's type signature.
 export { visibleWidth };
+
+/** v0.31.2: provider preference for the "diverse" auditor strategy (user
+ * design 2026-07-31: "there is benefit to have a different auditor — M3's
+ * auditor could be deepseek and vice versa"). A cross-vendor auditor reads
+ * the executor's claims with INDEPENDENT blind spots (same-family models
+ * share failure modes), and it spends a DIFFERENT provider's quota pool —
+ * audits stop eating the coding session's MiniMax window. The auditor is
+ * already a fresh session (no prompt-cache sharing), so cross-vendor costs
+ * nothing in cache terms.
+ *
+ * Order: deepseek-via-openrouter first, MiniMax second, then the other
+ * providers seen on this rig. The session's own provider is EXCLUDED
+ * entirely (openrouter hosts everything — family-level reasoning about its
+ * catalogue isn't reliable, so exclusion is at provider granularity).
+ */
+export const DIVERSE_AUDITOR_PREFERENCE: Array<{ provider: string; match?: string }> = [
+  { provider: "openrouter", match: "deepseek/deepseek-chat" },
+  { provider: "openrouter", match: "deepseek" },
+  { provider: "minimax", match: "MiniMax-M3" },
+  { provider: "minimax" },
+  { provider: "kimi-coding" },
+  { provider: "kimi" },
+  { provider: "xai-auth" },
+  { provider: "opencode" },
+  { provider: "zenmux" },
+];
+
+/** Pure selection: the first preference entry (outside the session's
+ * provider) with at least one available model wins; within an entry the
+ * models arrive pre-filtered (auth checked by the caller) and sorted.
+ * Returns undefined when nothing outside the session provider is available
+ * — the caller falls back LOUDLY to the session model.
+ */
+export function pickDiverseAuditorModel(
+  models: RegistryModelLike[],
+  sessionProvider: string | undefined,
+): RegistryModelLike | undefined {
+  for (const pref of DIVERSE_AUDITOR_PREFERENCE) {
+    if (pref.provider === sessionProvider) continue;
+    const hit = models.find((m) => m.provider === pref.provider && (!pref.match || m.id.includes(pref.match)));
+    if (hit) return hit;
+  }
+  return undefined;
+}

package/extensions/settings-menu.ts

@@ -173,9 +173,9 @@ export function buildSettingsRows(
       id: "auditorThinkingLevel",
       section: "auditor",
       label: "Auditor thinking",
-      valueText: show("auditorThinkingLevel", "session, floor high"),
+      valueText: show("auditorThinkingLevel", "high (fixed — never the session coding dial)"),
       sourceText: src("auditorThinkingLevel"),
-      description: "thinking level for the auditor session",
+      description: "thinking level for the auditor session — unset = high (the verification gate does not ride the session thinking dial)",
     },
     {
       id: "auditCap",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-goal-list-loop-audit",
-  "version": "0.31.0",
+  "version": "0.31.2",
   "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",