@kal-elsam/kairo-runtime 0.23.0 → 0.23.2

package/CHANGELOG.md

@@ -5,6 +5,45 @@ Historical entries below may reference the legacy `@kal-elsam/harness` package n
 
 ## Unreleased
 
+## 0.23.2 — 2026-09-18 (Kairo Runtime)
+
+Patch release.
+
+### Fixed
+
+- Routing eligibility (`checkCandidate`, shared by real execution
+  routing, ASK routing, and the FIT widget) and ASK's own ordering
+  only ever read the primary (5h) quota window — a provider whose
+  weekly window was nearly exhausted still got picked first and was
+  never excluded, as long as its 5h window looked healthy. Now takes
+  the worse of the two windows everywhere, fail-closed, matching the
+  same principle already applied to OpenCode Go's own windows.
+- `askCodex`/`askClaude` (the ASK-mode question path) used a fixed 30s
+  deadline from process start. Codex's ASK path always uses the
+  provider's single default model regardless of question complexity
+  (no effort-based tiering for Codex today), so a heavier default
+  model plus a cold sandboxed `codex exec` start can genuinely exceed
+  30s with real quota to spare — a real, slow answer, not a hang. Both
+  ask calls now reset their timeout on every real stdout/stderr chunk
+  instead, never an absolute one, so only a genuine hang still times
+  out.
+
+## 0.23.1 — 2026-09-18 (Kairo Runtime)
+
+Patch release.
+
+### Fixed
+
+- A plain chat question's action label and any failure message said
+  generic "Asking Kairo" throughout the wait, even though one specific
+  real adapter (Codex or Claude) is what's actually being asked and
+  can actually fail — Kairo is the system, never the one answering.
+  Adds `service.planAsk`, a read-only preview of ASK mode's routing
+  decision (reusing the same cached probes `submitTask` already uses),
+  so the cockpit now names the real provider/model from the start of
+  the wait instead of only leaking it incidentally through a raw error
+  message on failure.
+
 ## 0.23.0 — 2026-09-18 (Kairo Runtime)
 
 Minor release. First slice of PROJECT TEAM's automatic model-fallback

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@kal-elsam/kairo-runtime",
-  "version": "0.23.0",
+  "version": "0.23.2",
   "description": "Kairo Runtime — local agent operating system for Codex, Cursor, Claude, Pi, Engram, and Graphify.",
   "type": "module",
   "homepage": "https://github.com/Kal-elSam/harness#readme",

package/src/global/cockpit/app.js

@@ -248,7 +248,7 @@ export async function runCockpitApp({
     });
   }
 
-  editor.onSubmit = (text) => {
+  editor.onSubmit = async (text) => {
     const task = text.trim();
     if (!task) return;
     if (task.startsWith("/")) {
@@ -435,7 +435,38 @@ export async function runCockpitApp({
     // PLAN/AGENT always create a plan (see service.submitTask) — replacing
     // the old isLikelyQuestion guess with what the user explicitly told
     // Kairo they're doing (Shift+Tab / /plan).
-    return runAction("Asking Kairo", async () => {
+    //
+    // Kairo is the system, never the thing actually answering — a real
+    // adapter (Codex, Claude, …) always is, and that real name should be
+    // visible from the moment the wait starts, not just in a successful
+    // result or leaked incidentally through an error message. PLAN/AGENT
+    // is deterministically Codex (see submitArchitecture/createPlan), so
+    // that label needs no extra call; ASK's real provider depends on
+    // current quota/eligibility, so a cheap planAsk preview (the same
+    // routing submitTask itself will use, its probes cached — see
+    // service.js's planAsk doc) resolves it first. A caller that predates
+    // planAsk (or a preview that itself fails) falls back to the old
+    // generic label rather than ever blocking the real submit on it.
+    const mode = view.workMode;
+    let askLabel = "Asking Kairo";
+    if (mode === "plan" || mode === "agent") {
+      askLabel = "Asking Codex to plan";
+    } else {
+      try {
+        const preview = await service.planAsk?.({ cwd, task });
+        if (preview?.decision?.decision === "ROUTED") {
+          const provider = preview.decision.provider;
+          const providerLabel = provider ? provider.charAt(0).toUpperCase() + provider.slice(1) : null;
+          if (providerLabel) {
+            askLabel = `Asking ${providerLabel}${preview.decision.model ? ` · ${preview.decision.model}` : ""}`;
+          }
+        }
+      } catch {
+        // Best-effort label only — a real routing failure still surfaces
+        // through submitTask itself below, never swallowed here.
+      }
+    }
+    return runAction(askLabel, async () => {
       const result = await service.submitTask({ cwd, task, mode: view.workMode });
       if (result.kind === "answer") {
         pushTranscript("kairo", `${result.provider}${result.model ? ` · ${result.model}` : ""}: ${result.answer}`);

package/src/global/conversation/service.js

@@ -651,12 +651,16 @@ export function createConversationService(deps = {}) {
       return { ...publicPlan(result.status), reused: result.reused === true, projectRoot };
     },
     /**
-     * Real read-only question -> real answer, via whichever provider is
-     * actually available/quota-healthy — no task, no plan, no approval
-     * gate. Throws (never returns a fabricated answer) if no provider can
-     * answer or the call itself fails.
+     * Read-only preview of who ASK mode would actually ask right now —
+     * never calls a provider. Lets a caller (the cockpit's action label)
+     * show the real provider/model BEFORE the potentially slow real call
+     * starts, instead of a generic "Asking Kairo" that stays true no
+     * matter which real provider ends up answering (or timing out).
+     * Cheap to call again right after: the same underlying usage/catalog
+     * probes askQuestion itself uses are cached (see readCodexUsageCached
+     * etc.), so there's no real duplicate provider I/O.
      */
-    async askQuestion({ cwd, task }) {
+    async planAsk({ cwd, task }) {
       const projectRoot = await root(cwd);
       const adapters = inspectAdapters({ cwd: projectRoot });
       let codexUsage = null;
@@ -672,6 +676,16 @@ export function createConversationService(deps = {}) {
         claudeCatalog = readClaudeModelsImpl();
       }
       const decision = routeAsk({ adapters, codexUsage, claudeUsage, catalogs: { codex: codexCatalog, claude: claudeCatalog }, taskText: task });
+      return { decision, projectRoot };
+    },
+    /**
+     * Real read-only question -> real answer, via whichever provider is
+     * actually available/quota-healthy — no task, no plan, no approval
+     * gate. Throws (never returns a fabricated answer) if no provider can
+     * answer or the call itself fails.
+     */
+    async askQuestion({ cwd, task }) {
+      const { decision, projectRoot } = await this.planAsk({ cwd, task });
       if (decision.decision !== "ROUTED") throw new Error(`Cannot answer: ${decision.why}`);
       const result = await askProviderImpl({ provider: decision.provider, question: task, model: decision.model, cwd: projectRoot });
       if (result.status !== "answered") throw new Error(result.error ?? `${decision.provider} gave no answer.`);

package/src/global/intelligence/execution-router.js

@@ -118,9 +118,21 @@ function findAdapter(adapterId, adapters) {
   return adapters.find((adapter) => adapter.id === baseId) ?? null;
 }
 
-/** @param {object|null} usageEntry - a codex/claude usage-probe result (primary/secondary windows) */
+/**
+ * The WORST remaining headroom across both real windows (5h "primary" and
+ * weekly "secondary") — never just the primary one. A provider whose
+ * weekly quota is nearly gone must be treated that way everywhere
+ * (exclusion AND ordering) even while its 5h window still looks healthy;
+ * otherwise Kairo keeps routing to it, burning through the one budget
+ * that's actually about to run out. Mirrors the same fail-closed
+ * principle checkCandidate already applies to OpenCode Go's windows (any
+ * one window being real trouble is real trouble, full stop).
+ * @param {object|null} usageEntry - a codex/claude usage-probe result (primary/secondary windows)
+ */
 function remainingPercent(usageEntry) {
-  return usageEntry?.primary?.remainingPercent ?? null;
+  const known = [usageEntry?.primary?.remainingPercent, usageEntry?.secondary?.remainingPercent]
+    .filter((value) => typeof value === "number");
+  return known.length > 0 ? Math.min(...known) : null;
 }
 
 // Below this real remaining-quota percentage, a provider is treated as

package/src/global/intelligence/quick-ask.js

@@ -37,6 +37,32 @@ function unknown(error) {
   return { status: "error", answer: null, error: String(error) };
 }
 
+/**
+ * Resets on every real stdout/stderr chunk from the child, never an
+ * absolute deadline from process start — the same real distinction
+ * execution-adapters/opencode.js's own idle timeout draws: a real, live
+ * answer that's just taking a while (a heavier reasoning model, a cold
+ * sandbox start) must never be killed for merely being slow, only a
+ * process that's produced nothing at all for `timeoutMs` really looks
+ * hung. A single-shot ask call, so this stays local rather than reusing
+ * run-supervisor.js's own detached-run mechanism.
+ * @param {import("node:child_process").ChildProcess} child
+ * @param {number} timeoutMs
+ * @param {() => void} onIdle
+ * @returns {() => void} call to clear the timer once the call finishes
+ */
+function armIdleTimeout(child, timeoutMs, onIdle) {
+  let handle = null;
+  const reset = () => {
+    if (handle) clearTimeout(handle);
+    handle = setTimeout(onIdle, timeoutMs);
+  };
+  child.stdout?.on("data", reset);
+  child.stderr?.on("data", reset);
+  reset();
+  return () => { if (handle) clearTimeout(handle); };
+}
+
 /** @param {{question:string, model:string|null, cwd:string, spawn:Function, timeoutMs:number, env:object}} args */
 function askClaude({ question, model, cwd, spawn, timeoutMs, env }) {
   // --restricted: removes Bash/code-execution tools and WebFetch, ignores
@@ -57,11 +83,11 @@ function askClaude({ question, model, cwd, spawn, timeoutMs, env }) {
     }
     let stdout = "";
     let finished = false;
-    const timer = setTimeout(() => finish(unknown("claude -p timed out")), timeoutMs);
+    const clearIdleTimer = armIdleTimeout(child, timeoutMs, () => finish(unknown(`claude -p idle-timed out after ${timeoutMs}ms with no output`)));
     function finish(result) {
       if (finished) return;
       finished = true;
-      clearTimeout(timer);
+      clearIdleTimer();
       try { child.kill?.(); } catch { /* best effort */ }
       resolve(result);
     }
@@ -106,11 +132,11 @@ async function askCodex({ question, model, cwd, spawn, timeoutMs, env }) {
         return;
       }
       let finished = false;
-      const timer = setTimeout(() => finish(unknown("codex exec timed out")), timeoutMs);
+      const clearIdleTimer = armIdleTimeout(child, timeoutMs, () => finish(unknown(`codex exec idle-timed out after ${timeoutMs}ms with no output`)));
       function finish(result) {
         if (finished) return;
         finished = true;
-        clearTimeout(timer);
+        clearIdleTimer();
         try { child.kill?.(); } catch { /* best effort */ }
         resolve(result);
       }

package/src/global/observability/codex-models.js

@@ -89,7 +89,7 @@ export async function readCodexModels({
     child.once?.("close", () => { if (!finished) finish(unknown("codex app-server closed before model list")); });
 
     writeRequest(child, 1, "initialize", {
-      clientInfo: { name: "kairo", title: "Kairo", version: "0.23.0" },
+      clientInfo: { name: "kairo", title: "Kairo", version: "0.23.2" },
       capabilities: {}
     });
   });

package/src/global/observability/codex-usage.js

@@ -151,7 +151,7 @@ export async function readCodexUsage({
     child.once?.("close", () => { if (!finished) finish(unknown("codex app-server closed before rate limits")); });
 
     writeRequest(child, 1, "initialize", {
-      clientInfo: { name: "kairo", title: "Kairo", version: "0.23.0" },
+      clientInfo: { name: "kairo", title: "Kairo", version: "0.23.2" },
       capabilities: {}
     });
   });