bermudis-pi-goodies 0.23.0 → 0.23.2

package/README.md

@@ -25,7 +25,7 @@ extensions. One entry point, twelve independent features.
 After publishing the package to npm:
 
 ```bash
-pi install npm:bermudis-pi-goodies@0.23.0
+pi install npm:bermudis-pi-goodies@0.23.2
 ```
 
 Remove any old `bermudis-pi-goodies.ts` symlink before reloading Pi. Each

package/clean-tui.ts

@@ -44,7 +44,12 @@ import { Box, Container, Text } from "@earendil-works/pi-tui";
 import { homedir } from "node:os";
 import { readFileSync } from "node:fs";
 import { completeSimple } from "@earendil-works/pi-ai/compat";
-import type { Api, Model, ThinkingLevel } from "@earendil-works/pi-ai";
+import type {
+  Api,
+  AssistantMessage,
+  Model,
+  ThinkingLevel,
+} from "@earendil-works/pi-ai";
 import { logGoodiesEvent, setGoodiesLogPathForTesting } from "./goodies-log.ts";
 import { describeError } from "./json-file.ts";
 import {
@@ -596,8 +601,8 @@ async function summarizeViaProvider(
   signal: AbortSignal,
 ): Promise<string> {
   const t = await resolveSummaryTransport();
-  const response = await completeSimple(
-    t.model,
+  const response = await completeSummaryTurn(
+    t,
     {
       messages: [
         {
@@ -609,13 +614,7 @@ async function summarizeViaProvider(
         },
       ],
     },
-    {
-      apiKey: t.apiKey,
-      headers: t.headers,
-      maxTokens: SUMMARY_MAX_TOKENS,
-      signal,
-      reasoning: summaryReasoning(t.model),
-    },
+    signal,
   );
   return convertSummaryResponse(response, t.label);
 }
@@ -625,8 +624,8 @@ async function summarizeThinkingViaProvider(
   signal: AbortSignal,
 ): Promise<string> {
   const t = await resolveSummaryTransport();
-  const response = await completeSimple(
-    t.model,
+  const response = await completeSummaryTurn(
+    t,
     {
       messages: [
         {
@@ -636,13 +635,7 @@ async function summarizeThinkingViaProvider(
         },
       ],
     },
-    {
-      apiKey: t.apiKey,
-      headers: t.headers,
-      maxTokens: SUMMARY_MAX_TOKENS,
-      signal,
-      reasoning: summaryReasoning(t.model),
-    },
+    signal,
   );
   return convertSummaryResponse(response, t.label);
 }
@@ -654,6 +647,25 @@ async function summarizeThinkingViaProvider(
  * thing a raw endpoint could never do. Safe to call fire-and-forget
  * (pi-codex makes OAuth-refreshing calls the same way).
  */
+/**
+ * Summaries are best-effort: an unconfigured model (or a session whose model
+ * registry never arrived) means the feature is OFF, not failing. Those
+ * conditions throw this sentinel so every failure path can drop them
+ * silently — no log line, no backoff, no pause widget. Config can flip
+ * underneath a session at any moment (another pi session rewrites
+ * goodies.json on every write), so this must be checked at the failure
+ * boundary, not only at the enqueue gates.
+ */
+function summariesOffError(reason: string): Error {
+  const err = new Error(reason);
+  err.name = "SummariesOffError";
+  return err;
+}
+
+function isSummariesOffError(err: unknown): boolean {
+  return (err as Error | undefined)?.name === "SummariesOffError";
+}
+
 async function resolveSummaryTransport(): Promise<{
   model: Model<Api>;
   label: string;
@@ -661,10 +673,10 @@ async function resolveSummaryTransport(): Promise<{
   headers?: Record<string, string | null>;
 }> {
   const configured = getSummaryModel();
-  if (!configured) throw new Error("no summary model configured");
+  if (!configured) throw summariesOffError("no summary model configured");
   const registry = summaryModelRegistry;
   if (!registry)
-    throw new Error("model registry not captured yet this session");
+    throw summariesOffError("model registry not captured yet this session");
   const found = findSummaryModel(registry, configured);
   if (!found)
     throw new Error(`summary model "${configured}" not found in registry`);
@@ -748,6 +760,46 @@ function summaryReasoning(model: Model<Api>): ThinkingLevel | undefined {
   return "minimal";
 }
 
+/**
+ * Run one summary completion through pi-ai, shared by both summary kinds so
+ * the reasoning_effort compatibility retry lives in exactly one place.
+ *
+ * Some OpenAI-compatible endpoints validate reasoning_effort against their
+ * own enum and reject "minimal" outright — observed on Command Code
+ * (commandcode/poolside/*): 400 invalid_request_error, param:"reasoning_effort",
+ * accepted values low|medium|high|xhigh|max. "low" is the floor of every known
+ * enum, so retry once there before surfacing the failure; an endpoint that
+ * rejects "low" too would pause as before.
+ */
+async function completeSummaryTurn(
+  t: Awaited<ReturnType<typeof resolveSummaryTransport>>,
+  context: Parameters<typeof completeSimple>[1],
+  signal: AbortSignal,
+): Promise<AssistantMessage> {
+  const options = (reasoning: ThinkingLevel | undefined) => ({
+    apiKey: t.apiKey,
+    headers: t.headers,
+    maxTokens: SUMMARY_MAX_TOKENS,
+    signal,
+    reasoning,
+  });
+  // completeSimple does NOT throw for HTTP errors — it returns an
+  // AssistantMessage with stopReason:"error" + errorMessage, so the
+  // compatibility check inspects the response, not a catch block.
+  const response = await completeSimple(
+    t.model,
+    context,
+    options(summaryReasoning(t.model)),
+  );
+  if (
+    response.stopReason === "error" &&
+    (response.errorMessage ?? "").includes("reasoning_effort")
+  ) {
+    return await completeSimple(t.model, context, options("low"));
+  }
+  return response;
+}
+
 // Summaries are best-effort polish over the heuristic hint, but failures must
 // not be silent: every request lands in the structured log with its outcome,
 // so request volume and a broken provider/key/model choice are queryable
@@ -935,8 +987,14 @@ function startSummaryRequest(cmd: string): void {
       // concurrency slot with zero log output.
       if (!signal.aborted) pendingSummaries.delete(cmd);
       // Switching sessions aborts in-flight summaries deliberately: that is
-      // not a provider failure — neither penalize nor log it.
-      if (signal.aborted || (result.err as Error)?.name === "AbortError")
+      // not a provider failure — neither penalize nor log it. Same for the
+      // feature being switched off underneath the request (config rewrite by
+      // another session): drop silently, no backoff, no pause widget.
+      if (
+        signal.aborted ||
+        (result.err as Error)?.name === "AbortError" ||
+        isSummariesOffError(result.err)
+      )
         return;
       const pauseMs = noteSummaryFailure(result.err);
       logSummaryFailure(
@@ -1069,6 +1127,14 @@ async function summarizeWithRetries(job: {
 
 /** Start queued requests while capacity allows and no backoff is active. */
 function drainSummaryQueue(): void {
+  // The feature can be switched off between enqueue and drain — config is
+  // re-read from disk on every write and any pi session can rewrite it. Off
+  // means off: drop the deferred requests instead of draining them into
+  // requests that resolveSummaryTransport can only refuse.
+  if (!getSummaryModel()) {
+    summaryRequestQueue.length = 0;
+    return;
+  }
   while (
     summaryRequestQueue.length > 0 &&
     pendingSummaries.size < SUMMARY_MAX_INFLIGHT &&
@@ -1321,7 +1387,11 @@ function maybeRequestThinkingSummary(text: string): void {
         }
         return;
       }
-      if (!signal.aborted && (result.err as Error)?.name !== "AbortError") {
+      if (
+        !signal.aborted &&
+        (result.err as Error)?.name !== "AbortError" &&
+        !isSummariesOffError(result.err)
+      ) {
         const pauseMs = noteSummaryFailure(result.err);
         logSummaryFailure(
           text,

package/goodies.ts

@@ -341,13 +341,18 @@ export function completeGoodiesArguments(
   if (verb === "summary-model") {
     // Single value token (model ids contain no spaces): "off"/"default"
     // clear the setting; anything else matches the catalogue, ranked.
+    // Values MUST carry the verb: pi applies a selection by replacing the
+    // whole argument-text span with item.value (both its slash-argument
+    // path and the forced-Tab wrapper use prefix: argumentText), so bare
+    // values wipe "summary-model" off the line. enable/disable and
+    // thinking-summaries already follow this pattern.
     const q = valuePrefix.toLowerCase();
     const items: AutocompleteItem[] = [];
     for (const w of ["off", "default"]) {
-      if (w.startsWith(q)) items.push({ value: w, label: w });
+      if (w.startsWith(q)) items.push({ value: `${verb} ${w}`, label: w });
     }
     for (const c of rankCandidates(completionModels, q).slice(0, 20)) {
-      items.push({ value: c, label: c });
+      items.push({ value: `${verb} ${c}`, label: c });
     }
     return items.length ? items : null;
   }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "bermudis-pi-goodies",
-  "version": "0.23.0",
+  "version": "0.23.2",
   "repository": {
     "type": "git",
     "url": "git+https://github.com/bermudi/agent-extensions.git",