clementine-agent 1.0.15 → 1.0.17

package/dist/agent/self-improve.js

@@ -220,6 +220,26 @@ export class SelfImproveLoop {
                 // Check plateau
                 if (consecutiveLow >= this.config.plateauLimit) {
                     logger.info({ consecutiveLow }, 'Plateau detected — stopping');
+                    // Record the plateau in the experiment log so it's not silently
+                    // invisible. Helps the dashboard and failure monitor distinguish
+                    // "exhausted diverse hypotheses" from "crashed mid-run".
+                    const plateauExperiment = {
+                        id: randomBytes(4).toString('hex'),
+                        iteration: i,
+                        startedAt: new Date(loopStart).toISOString(),
+                        finishedAt: new Date().toISOString(),
+                        durationMs: Date.now() - loopStart,
+                        area: 'soul',
+                        target: 'n/a',
+                        hypothesis: 'No new hypothesis — diversity constraint exhausted',
+                        proposedChange: '',
+                        baselineScore: 0,
+                        score: 0,
+                        accepted: false,
+                        approvalStatus: 'denied',
+                        reason: 'Plateau: no novel improvement area remaining',
+                    };
+                    this.appendExperimentLog(plateauExperiment);
                     break;
                 }
                 state.currentIteration = i;

package/dist/gateway/failure-monitor.js

@@ -175,7 +175,20 @@ export function computeBrokenJobs(now = Date.now()) {
         const lastRunMs = Date.parse(lastEntry.startedAt);
         // Always consult the breaker state — a stuck breaker is the primary
         // signal for "job has been silently broken for days".
-        const cb = lastCircuitBreakerEvent(jobName);
+        let cb = lastCircuitBreakerEvent(jobName);
+        // Clear a "stuck" breaker flag if we see an ok run AFTER the last
+        // breaker engagement. The scheduler only logs a circuit-recovery
+        // event when consecutiveErrors >= 5 at recovery time — but a
+        // successful manual/probe run resets consecutiveErrors to 0 first,
+        // so the recovery branch never fires and the advisor log keeps the
+        // breaker appearing engaged forever. Fix: use run-log truth instead.
+        if (cb.engagedAt) {
+            const engagedMs = Date.parse(cb.engagedAt);
+            const hasOkSinceBreaker = entries.some(e => e.status === 'ok' && Date.parse(e.startedAt) > engagedMs);
+            if (hasOkSinceBreaker) {
+                cb = { engagedAt: null, lastOpinion: cb.lastOpinion };
+            }
+        }
         if (!cb.engagedAt && Number.isFinite(lastRunMs) && lastRunMs < dormantCutoffMs) {
             continue;
         }
@@ -278,24 +291,29 @@ function detectSelfImproveBreakage(now) {
         }
         catch { /* non-fatal */ }
     }
-    const lastRunMs = state.lastRunAt ? Date.parse(state.lastRunAt) : 0;
-    const lookback48h = now - 48 * 60 * 60 * 1000;
     const staleLookback = now - 7 * 24 * 60 * 60 * 1000; // 7 days
     const recentExperiments = experiments.filter(e => {
         const ts = e.startedAt ? Date.parse(e.startedAt) : 0;
         return Number.isFinite(ts) && ts >= staleLookback;
     });
     const recentErrors = recentExperiments.filter(e => e.approvalStatus === 'denied' && (e.reason?.startsWith('Error') ?? false));
-    // Three break modes:
-    //  a. state.infraError is set (loop detected unfixable infra issue)
-    //  b. all 3+ most recent experiments within lookback are errors
-    //  c. loop ran recently but no new experiments appeared (silent early-exit)
+    // Break modes we care about:
+    //  a. state.infraError is set — loop detected unfixable infra issue
+    //  b. state.status === 'failed' — run threw, didn't complete normally
+    //  c. all 3+ most recent experiments are errors — persistent iteration failures
+    //
+    // Deliberately NOT flagging "silent early exit" (lastRunAt recent but no new
+    // experiments) when state.status === 'completed'. That's the expected
+    // plateau state: the hypothesizer returns null for every iteration because
+    // the diversity constraint has blocked every previously-targeted area, the
+    // loop skips, plateau triggers, loop exits cleanly. Not broken — saturated.
+    // Forcing alarm on a saturated-but-healthy loop would make the monitor
+    // unusable long-term.
     const hasInfraError = !!state.infraError;
+    const runFailed = state.status === 'failed';
     const allRecentErrored = recentExperiments.length >= 3
         && recentExperiments.every(e => e.approvalStatus === 'denied');
-    const silentEarlyExit = lastRunMs > lookback48h
-        && recentExperiments.length === 0;
-    if (!hasInfraError && !allRecentErrored && !silentEarlyExit)
+    if (!hasInfraError && !runFailed && !allRecentErrored)
         return null;
     const lastErrors = [];
     for (let i = experiments.length - 1; i >= 0 && lastErrors.length < 3; i--) {
@@ -304,21 +322,12 @@ function detectSelfImproveBreakage(now) {
             continue;
         lastErrors.push(err.slice(0, 400));
     }
-    // If we don't have an explicit infraError but the last recorded error
-    // looks schema-related, surface it — this captures the state where all
-    // iterations died with the same API 400 but state.infraError never got
-    // persisted (happens when MAX_INFRA_ERRORS isn't crossed within a run).
-    const lastLoggedError = experiments.length > 0 ? (experiments[experiments.length - 1].error ?? '') : '';
-    const inferredInfraSchema = /input_schema|tools\.\d+\.custom/i.test(lastLoggedError);
     let opinion;
     if (hasInfraError) {
         opinion = `infra: ${state.infraError.category} — ${state.infraError.diagnostic.slice(0, 200)}`;
     }
-    else if (silentEarlyExit && inferredInfraSchema) {
-        opinion = 'loop ran but produced no experiments — last logged error was an MCP tool schema validation (API 400). Check external MCP servers (claude_desktop_config.json, Claude Code settings) for a recently-updated package exposing a malformed input_schema.';
-    }
-    else if (silentEarlyExit) {
-        opinion = 'loop ran but produced no experiments — likely crashing before iteration (check metrics gathering or hypothesis generation)';
+    else if (runFailed) {
+        opinion = 'loop exited with status=failed — check daemon log for the thrown error';
     }
     else {
         opinion = `${recentErrors.length}/${recentExperiments.length} recent iterations errored`;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "clementine-agent",
-  "version": "1.0.15",
+  "version": "1.0.17",
   "description": "Clementine — Personal AI Assistant (TypeScript)",
   "type": "module",
   "main": "dist/index.js",