pi-goal-list-loop-audit 0.21.1 → 0.22.0

package/README.md

@@ -88,7 +88,7 @@ grilling turn that ends in a Confirm dialog. For `/loop` specifically, the orche
 measure command once** and shows the real number in the dialog — you validate
 the metric before a single iteration burns tokens.
 
-With `branch=1`, all work lands on a scratch branch (`pi-gla-loop/<ts>-<slug>`):
+With `branch=1`, all work lands on a scratch branch (`pi-glla-loop/<ts>-<slug>`):
 improvements are committed, regressions are hard-reset (scratch branch only),
 and on stop you return to your original branch with merge instructions.
 Requires a clean working tree.
@@ -185,9 +185,10 @@ picks a model itself. Thinking follows the session too (floor `high`).
 ## Token guard
 
 Every goal tracks real token usage; crossing the budget pauses the goal.
-Default 10,000,000 per goal — a runaway threshold, not a big-goal threshold
-(real research/feature goals legitimately burn 2-4M). Tune with
-`/glla tokenlimit=<n>`. Loop 3 doesn't need this cap — it has its own brakes
+Off by default (opt-in) — set a budget with `/glla tokenlimit=<n>`. A high
+value like 10000000 is a runaway threshold, not a big-goal threshold
+(real research/feature goals legitimately burn 2-4M). Loop 3 doesn't need
+this cap — it has its own brakes
 (max iterations + plateau).
 
 ## Compatibility (what goes well, what conflicts)
@@ -234,7 +235,7 @@ prompts/
   goal-loop-forever-draft.md   # /loop drafting prompt
 scripts/
   smoke.sh                     # live integration harness (tmux + real models)
-tests/                         # 150 unit tests, no live pi required
+tests/                         # 168 unit tests, no live pi required
 docs/DESIGN.md                 # architectural decisions
 PLAN.md                        # milestones, decisions, gates
 ```

package/docs/DESIGN.md

@@ -51,8 +51,8 @@ below; later releases append addenda rather than rewrite history.
   auditor's evidence → disapproval, never a false approval.
 - **Token guard**: real accumulation from assistant-message `usage.totalTokens`
   (deduped across replayed `agent_end` history). Crossing `tokenlimit`
-  (default 1M) pauses the goal with a clear reason.
-- **Loop 3 `branch=1`**: scratch branch `pi-gla-loop/<ts>-<slug>`; commit per
+  (opt-in; off by default since v0.12.0) pauses the goal with a clear reason.
+- **Loop 3 `branch=1`**: scratch branch `pi-glla-loop/<ts>-<slug>`; commit per
   improvement, `git reset --hard` per regression — scratch branch only.
   Refuses non-git dirs and dirty trees; returns to the original branch on
   stop with merge instructions.
@@ -107,7 +107,7 @@ This is a deliberate trade-off. If the user wants drafting in v0.1.0, say so and
 
 Some alternatives considered:
 - Three packages: `pi-goal-list-loop-audit`, `pi-goal-list-loop-audit-list`, `pi-goal-list-loop-audit-loop`.
-- One package with three subcommands: `/goal`, `/pi-gla-list`, `/pi-gla-loop`.
+- One package with three subcommands: `/goal`, `/list`, `/loop`.
 
 We choose **one package with subcommands**. Reasoning:
 - Single install (`pi install npm:pi-goal-list-loop-audit`).
@@ -173,8 +173,8 @@ active → auditing          (complete_goal called)
 auditing → complete        (auditor <approved/>)
 auditing → active          (auditor <disapproved/>; reset iteration counter)
 active → paused            (pause_goal called, or stuck > 5 min, or empty turn)
-paused → active            (user /pi-gla-resume)
-active → aborted           (user /pi-gla-cancel)
+paused → active            (user /goal resume)
+active → aborted           (user /goal cancel)
 ```
 
 ### Decision 9: JSONL state (deterministic compaction)
@@ -189,9 +189,9 @@ This protects against model-generated summaries losing fidelity.
 |---|---|
 | `Esc` during auditor | Pause; user picks "complete without audit" or "continue" |
 | `Esc` during agent turn | Pause |
-| User `/pi-gla-pause` | Pause |
-| User `/pi-gla-cancel` | Abort (wipes active goal) |
-| Stuck > 5 min | Pause + notify |
+| User `/goal pause` | Pause |
+| User `/goal cancel` | Abort (wipes active goal) |
+| Stall watchdog (3 consecutive no-tool turns) | Pause + notify |
 | Empty turn (no tool calls) | Pause (no momentum) |
 
 ## Open follow-ups (post-v0.1.0)

package/extensions/goal-loop-auditor.ts

@@ -5,13 +5,11 @@
  * Isolated completion auditor. Runs in a fresh pi agent session with no
  * extensions, no skills, no prompts, no themes, no editor. Only read tools.
  *
- * In v0.1.0 the auditor behaves the same as pi-goal-x's auditor: it must
- * call at least one read tool before <approved/>.
- *
- * In v0.2.0 (NOT YET IMPLEMENTED) we will add regression_shield: the
- * auditor's report must include raw output (cat / grep / bash) for every
- * must-verify item in the verification contract. The orchestrator will
- * reject <approved/> without that evidence.
+ * Two enforced floors: the auditor must call at least one read tool before
+ * <approved/>, and regression_shield (goal-loop-shield.ts) requires the
+ * report to include raw output (cat / grep / bash) for every must-verify
+ * item in the verification contract — the orchestrator rejects
+ * evidence-free approvals.
  */
 
 import type { Model } from "@earendil-works/pi-ai";

package/extensions/goal-loop-core.ts

@@ -181,7 +181,7 @@ export function routeGoalArgs(raw: string): GoalRoute {
 /**
  * Parse a bulk list-import file (v0.8.1): markdown checklists (`- [ ]`,
  * `- [x]`), bullets (`-`, `*`, `•`), numbered items (`1.`, `2)`), and plain
- * lines all become queue items. Headings (`# …`), blank lines, and HTML
+ * lines all become list items. Headings (`# …`), blank lines, and HTML
  * comments are skipped. A sisyphus-style plan file should import clean.
  */
 export function parseListImport(content: string): string[] {
@@ -357,7 +357,7 @@ export function takeAt<T>(items: T[], n: number): [T, T[]] | null {
 
 export interface State {
   goal: Goal | null;
-  /** Loop 2: queue of pending goal items. Activated one at a time. */
+  /** Loop 2: list of pending goal items. Activated one at a time. */
   list?: ListItem[];
   /** Loop 3: metric-driven forever loop. */
   loop?: import("./goal-loop-forever.js").LoopState;

package/extensions/goal-loop-display.ts

@@ -4,6 +4,9 @@
  *
  * Pure display builders for the live TUI (status line + above-editor widget).
  * No pi imports — unit tests exercise these directly. The orchestrator calls
+ * No RUNTIME imports at all: tests run under `node --experimental-strip-types`,
+ * which does not rewrite `.js` → `.ts` specifiers — a value import from
+ * ./goal-loop-core.js breaks the suite (type-only imports are erased, safe).
  * ctx.ui.setStatus/setWidget with whatever these return.
  */
 
@@ -140,8 +143,11 @@ function goalLines(g: Goal, state: State, audit: AuditDisplayProgress | null | u
         : paint(theme, "success", "●");
   const head = `${icon} ${truncate(g.objective.replace(/\s+/g, " "), 64)}`;
   const statusWord = g.status === "active" ? paint(theme, "success", "active") : g.status;
-  const tokens = paint(theme, "dim", `${fmtTokens(g.usage?.tokensUsed ?? 0)}/${fmtTokens(g.usage?.tokensLimit ?? 10_000_000)} tok`);
-  const lines = [head, `├─ ${statusWord} · ${fmtElapsed(now - Date.parse(g.createdAt))} · ${tokens}`];
+  // Token segment only when a budget is set (v0.22.0): the guard is opt-in,
+  // and "0/0 tok" carried no information when off.
+  const tokenLimit = g.usage?.tokensLimit ?? 0;
+  const tokens = tokenLimit > 0 ? ` · ${paint(theme, "dim", `${fmtTokens(g.usage?.tokensUsed ?? 0)}/${fmtTokens(tokenLimit)} tok`)}` : "";
+  const lines = [head, `├─ ${statusWord} · ${fmtElapsed(now - Date.parse(g.createdAt))}${tokens}`];
   if (g.status === "auditing") {
     lines.push(`├─ auditor: ${audit?.label ?? "running"}${audit?.currentTool ? ` · ${truncate(audit.currentTool, 30)}` : ""}`);
     if (audit?.elapsedMs) lines.push(`└─ ${paint(theme, "dim", `${fmtElapsed(audit.elapsedMs)} in isolated session`)}`);

package/extensions/loops/goal.ts

@@ -71,15 +71,11 @@ import {
 } from "../goal-loop-forever.js";
 import {
   accountTurnForNudges,
-  BACKOFF_HARD_CAP_MS,
   BACKOFF_IDLE_RETRY_MS,
-  backoffMs,
   HEARTBEAT_INTERVAL_MS,
   HEARTBEAT_MAX_NUDGES,
   HEARTBEAT_STALL_MS,
-  humanMs,
   shouldHeartbeatRefire,
-  shouldPauseAfterBackoff,
 } from "../goal-loop-backoff.js";
 
 // =================================================================
@@ -87,7 +83,6 @@ import {
 // =================================================================
 
 const GOAL_EVENT_ENTRY = "goal-event";
-const STATE_ENTRY = "goal-state";
 /** stopReason marker for a loop held (not stopped) by the fresh-session restore gate. */
 const HELD_ON_RESTORE = "held: restored in a fresh session";
 
@@ -201,7 +196,6 @@ let continuationTimer: NodeJS.Timeout | null = null;
 let continuationScheduledFor: string | null = null;
 let iterationCounter = 0;
 let toolCallsThisTurn = 0;
-let consecutiveStuckIterations = 0;
 let consecutiveErrorIterations = 0;
 let consecutiveNoToolIterations = 0;
 
@@ -420,7 +414,6 @@ function activateNextListItem(ctx: ExtensionContext, n = 1): boolean {
   if (next.verificationContract) goal.verificationContract = next.verificationContract;
   setGoal(goal, ctx);
   iterationCounter = 0;
-  consecutiveStuckIterations = 0;
   consecutiveErrorIterations = 0;
   ctx.ui.notify(`List item #${n} activated (${rest.length} remaining): ${goal.objective.slice(0, 80)}`, "info");
   scheduleContinuation(ctx, true);
@@ -534,7 +527,6 @@ async function cmdSet(args: string, ctx: ExtensionContext, skipDraft = false): P
   setGoal(goal, ctx);
   // Reset counters
   iterationCounter = 0;
-  consecutiveStuckIterations = 0;
   consecutiveErrorIterations = 0;
   consecutiveNoToolIterations = 0;
   ctx.ui.notify(`Goal ${goal.id} created — starting now. Auditor will verify on completion.`, "info");
@@ -1233,7 +1225,7 @@ async function cmdLoop(args: string, ctx: ExtensionContext): Promise<void> {
     return;
   }
 
-  ctx.ui.notify("Usage: /loop [status] | /loop start \"<target>\" measure=\"<cmd>\" direction=min|max [done=<value>] [window=5] [max=50] | /loop stop", "info");
+  ctx.ui.notify("Usage: /loop [status] | /loop start \"<target>\" measure=\"<cmd>\" direction=min|max [window=5] [max=50] | /loop stop", "info");
 }
 
 // =================================================================
@@ -1545,7 +1537,6 @@ function registerAgentTools(pi: any, ctx: ExtensionContext): void {
       const goal = createGoal(full, liveCtx);
       setGoal(goal, liveCtx);
       iterationCounter = 0;
-      consecutiveStuckIterations = 0;
       consecutiveErrorIterations = 0;
       scheduleContinuation(liveCtx, true);
       return {
@@ -1774,7 +1765,7 @@ function registerAgentTools(pi: any, ctx: ExtensionContext): void {
 
   pi.registerTool(defineTool({
     name: "list_status",
-    label: "Queue status",
+    label: "List status",
     description: "Show the active goal and the /list list (loop 2) as text: what's running, what's waiting.",
     parameters: Type.Object({}),
     async execute() {
@@ -1856,7 +1847,8 @@ interface Settings {
   auditorThinkingLevel?: "off" | "minimal" | "low" | "medium" | "high" | "xhigh";
   /** Shell command run on goal complete / goal pause / loop stop; message passed as $1. */
   notifyCmd?: string;
-  /** Per-goal token budget; crossing it pauses the goal. Default 1,000,000. */
+  /** Per-goal token budget; crossing it pauses the goal. Off by default
+   * (opt-in guard, v0.12.0): unset/0 = no budget. */
   tokenLimit?: number;
   /** on → restored goals/loops/lists auto-resume even in fresh sessions
    * (unattended rigs). Default off: restore holds until /goal resume. */
@@ -2147,10 +2139,13 @@ async function cmdSettings(args: string, ctx: ExtensionContext): Promise<void> {
 const OUR_COMMANDS = ["goal", "glla", "list", "loop"];
 let collisionWarned = false;
 
-// Providers verified to exist in a bare (extension-less) session. The auditor
-// spawns exactly such a session, so extension-registered providers (kilocode,
-// zenmux on this rig) fail inside it. Unknown providers get a soft one-time
-// notice — not an error, since the built-in set grows over time.
+// Providers known to pi core. The auditor inherits the already-resolved
+// Model object from this session (in-process createAgentSession), so a
+// provider defined in ~/.pi/agent/models.json with auth.json credentials
+// works even though it is not "built-in". Unknown providers get a soft
+// one-time conditional notice: if audits error with auth failures, an
+// explicit /glla model= override is the fix. (v0.22.0: reworded from the
+// stale "extension-registered → auditor fails auth" premise.)
 const KNOWN_BUILTIN_PROVIDERS = new Set([
   "anthropic", "google", "google-vertex", "google-gemini-cli", "openai", "openai-codex",
   "openrouter", "opencode", "azure-openai-responses", "groq", "cerebras", "xai", "zai",
@@ -2167,7 +2162,7 @@ function warnIfAuditorProviderRisky(ctx: ExtensionContext): void {
     const provider = (ctx.model as any)?.provider as string | undefined;
     if (!provider || KNOWN_BUILTIN_PROVIDERS.has(provider)) return;
     ctx.ui.notify(
-      `pi-goal-list-loop-audit: session provider "${provider}" is extension-registered — the auditor (extension-less session) will fail auth with it. If audits error, set a built-in override once: /glla model=provider/id`,
+      `pi-goal-list-loop-audit: session provider "${provider}" is not a known built-in. The auditor inherits the resolved model in-process, so this usually works — but if audits error with auth/provider failures, set an explicit override once: /glla model=provider/id`,
       "info",
     );
   } catch {
@@ -2410,10 +2405,9 @@ export default function (pi: ExtensionAPI): void {
       consecutiveErrorIterations = 0;
     }
 
-    // Hard 5-min cap: if no successful continue for >5 min, pause.
-    // For v0.1.0 we only detect this on agent_end; a tighter watchdog lands in v0.2.0.
-    // We schedule the next continuation; if it produces nothing useful in 5 min,
-    // the next agent_end iteration of THIS branch will trigger the cap.
+    // No wall-clock cap by design: a goal ends via completion, explicit
+    // pause/cancel, the stall watchdog, the 5-consecutive-errors pause, or
+    // the token guard — never via an elapsed-time cutoff.
 
     scheduleContinuation(ctx, false);
   });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-goal-list-loop-audit",
-  "version": "0.21.1",
+  "version": "0.22.0",
   "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-continuation.md

@@ -73,11 +73,11 @@ pause_goal({reason: "...", suggestedAction: "..."})
 
 ## HARD RULES
 
-- **Do not modify the objective autonomously.** The objective is the user's; if it has drifted from what makes sense, call `pause_goal` and propose a `/pi-gla-tweak` instead.
+- **Do not modify the objective autonomously.** The objective is the user's; if it has drifted from what makes sense, call `pause_goal` and propose a `/goal tweak` instead.
 - **Do not pretend completion.** If verification evidence is missing, call `pause_goal` instead of `complete_goal`.
 - **Do not polish doorknobs.** If you are out of work and the goal is satisfied, call `complete_goal` instead of inventing a side-improvement.
 - **Do not give up early.** If a task is hard, run it down properly. The auditor will catch doorknobs; the agent's job is to do the real work.
 
-## BACKOFF
+## STALLS
 
-If the orchestrator schedules a 5-minute pause between iterations, that's the safety net: it means the agent has not made meaningful progress in a while. Use the pause to surface what is blocking, not to keep iterating in a way that wastes effort.
+The orchestrator's backstop is the stall watchdog: three consecutive turns with no tool calls pause the goal. If you feel yourself spinning — repeating the same approach, no new evidence — stop early instead: call `pause_goal` with what is blocking and a concrete suggested action, rather than burning the remaining watchdog turns.

package/schemas/goal.schema.json

@@ -2,7 +2,7 @@
   "$schema": "http://json-schema.org/draft-07/schema#",
   "$id": "https://raw.githubusercontent.com/DraconDev/pi-goal-list-loop-audit/main/schemas/goal.schema.json",
   "title": "pi-goal-list-loop-audit goal state",
-  "description": "Active goal state for pi-goal-list-loop-audit v0.4.0 (loops 1-3). Persisted as markdown at .pi-gla/goals/<id>.md and mirrored into .pi-gla/active.jsonl state events.",
+  "description": "Active goal state for pi-goal-list-loop-audit (loops 1-3). Persisted as markdown at .pi-glla/goals/<id>.md and mirrored into .pi-glla/active.jsonl state events.",
   "type": "object",
   "required": ["id", "objective", "status", "policy", "autoContinue", "usage", "createdAt", "updatedAt"],
   "properties": {