niahere 0.4.5 → 0.4.6

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "niahere",
-  "version": "0.4.5",
+  "version": "0.4.6",
   "description": "A personal AI assistant daemon — chat, scheduled jobs, persona system, extensible via skills.",
   "type": "module",
   "scripts": {

package/src/agent/backends/codex.ts

@@ -6,6 +6,7 @@ import type { Attachment } from "../../types/attachment";
 import type { McpSourceContext } from "../../mcp";
 import { CodexNormalizer } from "./codex-normalize";
 import { mintRun, revokeRun } from "../mcp-endpoint";
+import { isCliProviderDownError } from "../../utils/retry";
 
 /**
  * Resolve the codex binary's absolute path. The daemon runs under launchd with a
@@ -188,7 +189,7 @@ class CodexSession implements AgentSession {
           type: "error",
           message: stderr.trim() || `codex exited ${exit}`,
           retryable: false,
-          providerDown: false,
+          providerDown: isCliProviderDownError(stderr),
         };
       }
     } finally {

package/src/core/scheduler.ts

@@ -3,9 +3,25 @@ import { runJob } from "./runner";
 import { getConfig } from "../utils/config";
 import { log } from "../utils/log";
 import { computeInitialNextRun, computeNextRun } from "../utils/schedule";
+import type { JobResult } from "../types";
 
 export { computeInitialNextRun, computeNextRun };
 
+/**
+ * Log a finished job at a level matching its outcome. `runJob` resolves with
+ * `status: "error"` instead of throwing, so the caller's `.catch` only ever sees
+ * an infrastructure fault — a job that ran and failed must be branched on here
+ * or it is indistinguishable from success in the log.
+ */
+export function logJobOutcome(result: JobResult): void {
+  const fields = { job: result.job, status: result.status, duration: result.duration_ms };
+  if (result.status === "error") {
+    log.error({ ...fields, error: result.error, terminal_reason: result.terminal_reason }, "scheduler: job failed");
+    return;
+  }
+  log.info(fields, "scheduler: job completed");
+}
+
 function isWithinActiveHours(): boolean {
   const config = getConfig();
   const { start, end } = config.activeHours;
@@ -57,11 +73,9 @@ async function tick(): Promise<void> {
     runningJobs.add(job.name);
 
     runJob(job)
-      .then((result) => {
-        log.info({ job: job.name, status: result.status, duration: result.duration_ms }, "scheduler: job completed");
-      })
+      .then(logJobOutcome)
       .catch((err) => {
-        log.error({ err, job: job.name }, "scheduler: job failed");
+        log.error({ err, job: job.name }, "scheduler: job crashed");
       })
       .finally(() => {
         runningJobs.delete(job.name);

package/src/utils/retry.ts

@@ -31,6 +31,34 @@ export function isProviderDownError(error: string | null | undefined): boolean {
   return !trimmed || trimmed.toLowerCase() === "unknown error";
 }
 
+const CLI_PROVIDER_DOWN_PATTERNS = [
+  /authenticat/i,
+  /unauthoriz/i,
+  /\b(401|403)\b/,
+  /not logged in/i,
+  /\blogin\b/i,
+  /\bcredentials?\b/i,
+  /\bapi key\b/i,
+  /(session|token) expired/i,
+  /\b(econnrefused|enotfound|etimedout|econnreset)\b/i,
+  /connection (refused|reset|timed out)/i,
+  /network (error|unreachable)/i,
+  /failed to refresh available models/i,
+];
+
+/**
+ * Provider-down classification for CLI-subprocess backends (codex), whose
+ * failures surface as free-form stderr rather than the Claude SDK's structured
+ * error. Matches the ways a CLI reports "I could not reach or authenticate with
+ * the provider" — as opposed to a task that ran and genuinely failed, which must
+ * stay a real error so the chain does not burn a second backend replaying it.
+ */
+export function isCliProviderDownError(stderr: string | null | undefined): boolean {
+  const trimmed = stderr?.trim();
+  if (!trimmed) return true;
+  return CLI_PROVIDER_DOWN_PATTERNS.some((p) => p.test(trimmed));
+}
+
 /** Sleep for ms milliseconds. */
 export function sleep(ms: number): Promise<void> {
   return new Promise((r) => setTimeout(r, ms));