pi-goal-list-loop-audit 0.38.39 → 0.38.41

package/CHANGELOG.md

@@ -1,5 +1,27 @@
 # Changelog
 
+## 0.38.41 — grilling standard codified (2026-09-09)
+
+### Fixed
+
+- **Questionnaire discipline is now GLLA behavior, not an agreement:** all three drafting prompts (`goal-loop-draft.md`, `goal-loop-forever-draft.md`, `goal-loop-plan.md`) carry the roadmap-then-stages rule (batch independent questions, sequence dependent ones with the roadmap stated first), the rich-option standard (concrete-consequence descriptions + preview pane where end-states matter, never padded to four), and the explicit Esc-correction invitation.
+- **Pin test:** `tests/drafting-questionnaire.test.ts` (9 pins) fails if any prompt drops any part of the rule.
+
+### Verification
+
+- New pin test green; full `release:check` 0 failures; `tsc --noEmit` clean.
+
+## 0.38.40 — provider overflow errors count as starvation (2026-09-09)
+
+### Fixed
+
+- **PR #48 (Bjynt, merged):** `isContextStarvedLengthStop` now detects explicit provider context-overflow errors (`exceed_context_size_error`, `exceeds the available context size`, `context window exceeded`, …) arriving as `stopReason="error"`, firing the starvation ladder + emergency compactor (≥90% usage gate kept). Case 1 refactor is behavior-identical. Wiring confirmed live via the raw session message's `errorMessage`.
+- Follow-up noted on the PR: OpenAI-style `maximum context length` wording still falls through the matcher set.
+
+### Verification
+
+- Contributor's Case 2 tests + full `release:check` 0 failures; `tsc --noEmit` clean.
+
 ## 0.38.39 — Codex-like terminal summary (2026-09-09)
 
 ### 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–v0.38.39 continue through the supervisor freeze (`/glla pause`),
+  v0.35.14–v0.38.41 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/length-continue.ts

@@ -46,6 +46,10 @@ export interface AssistantLengthMessageLike {
   usage?: {
     output?: number;
   };
+  /** Provider error message, when stopReason is "error". */
+  errorMessage?: string;
+  /** Error type from provider (e.g., "exceed_context_size_error"). */
+  errorType?: string;
 }
 
 export const LENGTH_CONTINUE_CONTEXT_STARVED_PERCENT = 90;
@@ -60,21 +64,49 @@ export const LENGTH_CONTINUE_CONTEXT_STARVED_MAX_OUTPUT = 8;
  * here queues another 1-token request before pi's post-agent_end compaction
  * check and delays the actual cure (field: darklord 2026-08-02, 198,116 /
  * 198,179 total tokens of a 200,000 window, output=1 twice).
+ *
+ * v0.38.36: also detect explicit context-overflow provider errors
+ * (e.g., "exceed_context_size_error", "exceeds the available context size")
+ * which arrive as stopReason="error" with an errorMessage, not as
+ * stopReason="length". These are also context starvation — the request
+ * exceeded the model's context window before generation could start.
  */
 export function isContextStarvedLengthStop(
   message: AssistantLengthMessageLike | null | undefined,
   contextUsage: ContextUsageLike | null | undefined,
 ): boolean {
-  if (message?.stopReason !== "length") return false;
-  const output = message.usage?.output;
-  if (typeof output !== "number" || !Number.isFinite(output)) return false;
-  if (output > LENGTH_CONTINUE_CONTEXT_STARVED_MAX_OUTPUT) return false;
-  const percent = typeof contextUsage?.percent === "number"
-    ? contextUsage.percent
-    : typeof contextUsage?.tokens === "number" && typeof contextUsage?.contextWindow === "number" && contextUsage.contextWindow > 0
-      ? (contextUsage.tokens / contextUsage.contextWindow) * 100
-      : null;
-  return percent !== null && Number.isFinite(percent) && percent >= LENGTH_CONTINUE_CONTEXT_STARVED_PERCENT;
+  // Case 1: pi's context-safety clamp (stopReason="length" with tiny output)
+  if (message?.stopReason === "length") {
+    const output = message.usage?.output;
+    if (typeof output === "number" && Number.isFinite(output) && output <= LENGTH_CONTINUE_CONTEXT_STARVED_MAX_OUTPUT) {
+      const percent = typeof contextUsage?.percent === "number"
+        ? contextUsage.percent
+        : typeof contextUsage?.tokens === "number" && typeof contextUsage?.contextWindow === "number" && contextUsage.contextWindow > 0
+          ? (contextUsage.tokens / contextUsage.contextWindow) * 100
+          : null;
+      return percent !== null && Number.isFinite(percent) && percent >= LENGTH_CONTINUE_CONTEXT_STARVED_PERCENT;
+    }
+    return false; // Real overlong response, not context starvation
+  }
+  // Case 2: explicit provider context-overflow error (stopReason="error")
+  if (message?.stopReason === "error") {
+    const errMsg = (message.errorMessage ?? message.errorType ?? "").toLowerCase();
+    if (
+      errMsg.includes("exceed_context_size") ||
+      errMsg.includes("exceeds the available context") ||
+      errMsg.includes("exceeds the context window") ||
+      (errMsg.includes("exceeds the model") && errMsg.includes("context")) ||
+      (errMsg.includes("context window") && errMsg.includes("exceed"))
+    ) {
+      const percent = typeof contextUsage?.percent === "number"
+        ? contextUsage.percent
+        : typeof contextUsage?.tokens === "number" && typeof contextUsage?.contextWindow === "number" && contextUsage.contextWindow > 0
+          ? (contextUsage.tokens / contextUsage.contextWindow) * 100
+          : null;
+      return percent !== null && Number.isFinite(percent) && percent >= LENGTH_CONTINUE_CONTEXT_STARVED_PERCENT;
+    }
+  }
+  return false;
 }
 
 export function makeLengthContinueTracker(max: number = LENGTH_CONTINUE_MAX) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-goal-list-loop-audit",
-  "version": "0.38.39",
+  "version": "0.38.41",
   "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-draft.md

@@ -78,6 +78,14 @@ request into a **confirmed goal contract**. Do NOT start substantive work yet.
    If an `ask_user_question` tool is available in this session, prefer it for
    structured choices (it renders proper option lists); plain conversation is
    fine otherwise and for free-form answers.
+   **Questionnaire discipline — roadmap, then stages:** batch independent
+   questions upfront in one `ask_user_question` call; when a later question
+   depends on an earlier answer, state the full roadmap first, then ask
+   stage-by-stage so every option makes sense when answered. Every option
+   description states its concrete consequence, plus a preview pane wherever
+   what the option produces matters — two honest options beat four padded
+   guesses. Close every batch by inviting correction: if any of this is off,
+   Esc and say what's wrong, and you re-ask.
 2. Targeted read-only research is allowed when it helps define a better
    contract (read a file, check the repo layout). Do NOT implement anything.
    **Default to subagents for research**: use the `subagent` tool to spawn a `scout` subagent

package/prompts/goal-loop-forever-draft.md

@@ -56,6 +56,15 @@ stops it.
      genuinely endless (an ever-improving spec, continuous hardening).
 3. Clarify the **direction**: is lower better (min) or higher better (max)?
    (Skip this for a metricless loop — there is no direction without a metric.)
+   **Questionnaire discipline — roadmap, then stages:** batch independent
+   questions upfront in one `ask_user_question` call; when a later question
+   depends on an earlier answer (bounds depend on the metric, metricless
+   depends on no-number), state the full roadmap first, then ask
+   stage-by-stage so every option makes sense when answered. Every option
+   description states its concrete consequence, plus a preview pane wherever
+   what the option produces matters — two honest options beat four padded
+   guesses. Close every batch by inviting correction: if any of this is off,
+   Esc and say what's wrong, and you re-ask.
 4. Optional tuning: `window` (plateau stop after N non-improving iterations,
    default 5 — meaningless for metricless), `max` (iteration cap: default 50
    for METRIC loops; default UNBOUNDED (`max=0`) for metricless loops,

package/prompts/goal-loop-plan.md

@@ -27,7 +27,14 @@ confirms.
    - **Verification strategy** — how each milestone proves itself done.
    Ask one focused question at a time within a round; offer a recommended
    default with each question so the user can answer "yes". Prefer
-   `ask_user_question` for structured choices when available.
+   `ask_user_question` for structured choices when available. The rounds
+   ARE the stages: state the full round roadmap first, then work it
+   round-by-round (later rounds depend on earlier answers — never batch a
+   later round's questions upfront). Every option description states its
+   concrete consequence, plus a preview pane wherever what the option
+   produces matters — two honest options beat four padded guesses. Close
+   every round by inviting correction: if any of this is off, Esc and say
+   what's wrong, and you re-ask.
 3. **The proposal is a structured expanded objective.** Use this shape
    inside the objective text:
    - **Current state** — what exists today, with file/module references.