clementine-agent 1.1.6 → 1.1.8

package/dist/agent/assistant.d.ts

@@ -271,7 +271,7 @@ export declare class PersonalAssistant {
      * so follow-up conversation has context.
      */
     injectContext(sessionKey: string, userText: string, assistantText: string): void;
-    getRecentActivity(sinceIso: string): Array<{
+    getRecentActivity(sinceIso: string, maxEntries?: number): Array<{
         sessionKey: string;
         role: string;
         content: string;

package/dist/agent/assistant.js

@@ -4828,11 +4828,11 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
             }
         }
     }
-    getRecentActivity(sinceIso) {
+    getRecentActivity(sinceIso, maxEntries) {
         if (!this.memoryStore)
             return [];
         try {
-            return this.memoryStore.getRecentActivity(sinceIso);
+            return this.memoryStore.getRecentActivity(sinceIso, maxEntries);
         }
         catch {
             return [];

package/dist/agent/insight-engine.d.ts

@@ -47,13 +47,25 @@ export declare function maybeIncreaseCooldown(state: InsightState): void;
  * Returns structured event summaries that can be passed to an LLM for urgency rating.
  */
 export declare function gatherInsightSignals(gateway: {
-    getRecentActivity: (since: string) => Array<{
+    getRecentActivity: (since: string, maxEntries?: number) => Array<{
         sessionKey: string;
         role: string;
         content: string;
         createdAt: string;
     }>;
 }): string[];
+export declare function detectFrustrationSignals(activity: Array<{
+    sessionKey: string;
+    role: string;
+    content: string;
+    createdAt: string;
+}>): string[];
+export declare function detectRepeatedTopics(activity: Array<{
+    sessionKey: string;
+    role: string;
+    content: string;
+    createdAt: string;
+}>): string[];
 /**
  * Build a prompt for urgency rating (to be sent to a lightweight LLM).
  * Returns null if there are no signals worth evaluating.

package/dist/agent/insight-engine.js

@@ -189,7 +189,27 @@ export function gatherInsightSignals(gateway) {
     catch (err) {
         logger.debug({ err }, 'Failed to pull broken-jobs signals');
     }
-    // 6. Claim tracker — failed claims in the last N hours erode trust.
+    // 6. Conversational signals derived from recent transcripts.
+    //    Surfaces patterns IN the conversation itself, not just system events:
+    //    user frustration markers, repeating topics, etc. These are early
+    //    warning signs that the agent's responses may be off-track.
+    try {
+        const since24h = new Date(Date.now() - 24 * 60 * 60 * 1000).toISOString();
+        const since7d = new Date(Date.now() - 7 * 24 * 60 * 60 * 1000).toISOString();
+        // 24h frustration scan — 50 entries plenty to count corrections in a day.
+        const recent = gateway.getRecentActivity(since24h, 50);
+        for (const s of detectFrustrationSignals(recent))
+            signals.push(s);
+        // 7d repeat-topic scan — pull more entries since topics span sessions.
+        // Cap at 200 to keep keyword extraction cheap.
+        const week = gateway.getRecentActivity(since7d, 200);
+        for (const s of detectRepeatedTopics(week))
+            signals.push(s);
+    }
+    catch (err) {
+        logger.debug({ err }, 'Failed to pull conversational signals');
+    }
+    // 7. Claim tracker — failed claims in the last N hours erode trust.
     //    Surface them so the owner sees "Clementine said she'd do X; she
     //    didn't" instead of silently swallowing the miss.
     try {
@@ -214,6 +234,95 @@ export function gatherInsightSignals(gateway) {
     }
     return signals;
 }
+// ── Conversational signal detectors ─────────────────────────────────
+//
+// Pure functions over recent transcript activity. Exported so the insight
+// dashboard / debug commands can run them independently of the full
+// gatherInsightSignals path.
+/**
+ * Markers that suggest the user is correcting or frustrated with the
+ * agent's last response. Tuned to start-of-message tokens since
+ * mid-message "no" or "actually" is often just normal narrative.
+ */
+const CORRECTION_PATTERNS = [
+    /^(no|nope|not\b)/i,
+    /^(actually|wait)\b/i,
+    /^(that['’]?s| that is) (wrong|not|incorrect|backwards|opposite)/i,
+    /^I (meant|said|wanted|asked)\b/i,
+    /^you (didn['’]?t|misunderstood|got it wrong|missed)/i,
+    /^(stop|cancel|undo|nevermind|never mind)\b/i,
+];
+export function detectFrustrationSignals(activity) {
+    const signals = [];
+    let count = 0;
+    const sessionsAffected = new Set();
+    for (const entry of activity) {
+        if (entry.role !== 'user')
+            continue;
+        const trimmed = entry.content.trim();
+        for (const re of CORRECTION_PATTERNS) {
+            if (re.test(trimmed)) {
+                count++;
+                sessionsAffected.add(entry.sessionKey);
+                break;
+            }
+        }
+    }
+    if (count >= 3) {
+        signals.push(`Conversation friction: ${count} user correction(s) across ${sessionsAffected.size} session(s) in the last 24h — recent agent responses may be off-track`);
+    }
+    return signals;
+}
+/**
+ * Words too generic to count as a topic — would otherwise dominate the
+ * "recurring topic" signal with noise like "thanks", "okay", "please".
+ */
+const TOPIC_STOPWORDS = new Set([
+    'about', 'after', 'again', 'against', 'because', 'before', 'being', 'between',
+    'could', 'doing', 'don’t', 'down', 'during', 'each', 'from', 'further',
+    'going', 'gonna', 'have', 'having', 'here', 'into', 'just', 'know', 'like',
+    'maybe', 'might', 'more', 'most', 'much', 'need', 'okay', 'only', 'other',
+    'over', 'please', 'really', 'said', 'same', 'some', 'still', 'such', 'than',
+    'that', 'them', 'then', 'there', 'these', 'they', 'thing', 'think', 'this',
+    'those', 'through', 'thanks', 'time', 'told', 'under', 'until', 'using', 'very',
+    'want', 'wanted', 'wants', 'were', 'what', 'when', 'where', 'which', 'while',
+    'will', 'with', 'would', 'your', 'yours', 'yeah', 'yes',
+    'tonight', 'today', 'tomorrow', 'morning', 'evening', 'session', 'work',
+    'doing', 'made', 'make', 'making', 'sure', 'right', 'wrong', 'good', 'bad',
+    'much', 'many', 'lots',
+]);
+export function detectRepeatedTopics(activity) {
+    // Build a (keyword → set of session IDs) map. A keyword that shows up in
+    // 3+ DISTINCT sessions across the window is "recurring" — could be an
+    // unresolved thread, a project the user is grinding on, or a question
+    // they've asked multiple ways.
+    const sessionsForKeyword = new Map();
+    for (const entry of activity) {
+        if (entry.role !== 'user')
+            continue;
+        const text = entry.content.toLowerCase();
+        // Word extraction: 5+ chars, alpha-only (no numbers/punctuation).
+        const matches = text.match(/[a-z][a-z’]{4,15}/g) ?? [];
+        const seenInThisMessage = new Set();
+        for (const w of matches) {
+            if (TOPIC_STOPWORDS.has(w))
+                continue;
+            if (seenInThisMessage.has(w))
+                continue; // dedupe within a single message
+            seenInThisMessage.add(w);
+            if (!sessionsForKeyword.has(w))
+                sessionsForKeyword.set(w, new Set());
+            sessionsForKeyword.get(w).add(entry.sessionKey);
+        }
+    }
+    // Rank by session-spread; surface the top 2 to avoid flooding insight
+    // notifications with too many topic mentions.
+    const ranked = [...sessionsForKeyword.entries()]
+        .filter(([, sessions]) => sessions.size >= 3)
+        .sort((a, b) => b[1].size - a[1].size)
+        .slice(0, 2);
+    return ranked.map(([keyword, sessions]) => `Recurring topic "${keyword}" came up across ${sessions.size} sessions this week — possible ongoing thread`);
+}
 /**
  * Build a prompt for urgency rating (to be sent to a lightweight LLM).
  * Returns null if there are no signals worth evaluating.

package/dist/gateway/fix-applier.js

@@ -358,6 +358,18 @@ function applyAdvisorRuleFix(jobName, autoApply, opts) {
     }
     appendAudit({ kind: 'advisor-rule', jobName, file: targetPath, ruleId: autoApply.ruleId, diff });
     logger.info({ jobName, ruleId: autoApply.ruleId, file: targetPath }, 'Applied advisor-rule fix');
+    // Phase 8.1 — record this autoApply for verification. The next
+    // AUTOAPPLY_VERDICT_WINDOW non-skipped runs decide whether the rule
+    // helped; if not, fix-verification auto-reverts (deletes the file).
+    // Lazy import to avoid circular dependency (fix-verification imports
+    // failure-monitor which transitively touches the cron path).
+    import('./fix-verification.js').then(({ recordAutoApplyForVerification }) => {
+        recordAutoApplyForVerification(jobName, {
+            kind: 'advisor-rule',
+            file: targetPath,
+            ruleId: autoApply.ruleId,
+        });
+    }).catch(err => logger.warn({ err, jobName }, 'Failed to record autoApply for verification'));
     return {
         ok: true,
         message: `Wrote advisor rule ${autoApply.ruleId} (hot-reloads on next eval)`,
@@ -408,6 +420,15 @@ function applyPromptOverrideFix(jobName, autoApply, opts) {
         diff,
     });
     logger.info({ jobName, scope: autoApply.scope, scopeKey: autoApply.scopeKey, file: targetPath }, 'Applied prompt-override fix');
+    // Phase 8.1 — same multi-run verification flow as advisor-rule.
+    import('./fix-verification.js').then(({ recordAutoApplyForVerification }) => {
+        recordAutoApplyForVerification(jobName, {
+            kind: 'prompt-override',
+            file: targetPath,
+            scope: autoApply.scope,
+            scopeKey: autoApply.scopeKey,
+        });
+    }).catch(err => logger.warn({ err, jobName }, 'Failed to record autoApply for verification'));
     return {
         ok: true,
         message: `Wrote prompt override ${autoApply.scope}${autoApply.scopeKey ? `:${autoApply.scopeKey}` : ''}`,

package/dist/gateway/fix-verification.d.ts

@@ -13,6 +13,29 @@ interface PendingVerification {
     recordedAt: string;
     preFailureCount: number;
     preLastError: string | null;
+    /** Used by Phase 8.1 — when set, the verifier is also responsible for
+     * deleting this artifact if the fix doesn't help. Existing CRON.md edits
+     * leave this unset (they're hand-edits, not auto-applies, so we never
+     * revert them automatically). */
+    autoApply?: AutoApplyTracker;
+    /** Run-by-run history accumulated since the fix was applied. Single-run
+     * verdicts (the original CRON.md flow) only need the first entry; multi-
+     * run autoApply verifications need the accumulated sample. */
+    postRunOutcomes?: Array<'ok' | 'error' | 'retried'>;
+}
+/**
+ * Tracks an autoApply that's currently being verified. When the verdict
+ * window closes negatively, revertFix() uses these fields to undo.
+ */
+export interface AutoApplyTracker {
+    kind: 'advisor-rule' | 'prompt-override';
+    /** Absolute path of the file the apply wrote. */
+    file: string;
+    /** advisor-rule only: the rule's id, used by the loader's hot-reload. */
+    ruleId?: string;
+    /** prompt-override only: scope label for the verdict message. */
+    scope?: 'global' | 'agent' | 'job';
+    scopeKey?: string;
 }
 /**
  * Compare an old and new jobs list and record verifications for any job that:
@@ -25,15 +48,32 @@ interface PendingVerification {
  * pending verification" rather than waiting for a run that will never come.
  */
 export declare function recordEditsForFailingJobs(oldJobs: CronJobDefinition[], newJobs: CronJobDefinition[]): void;
+/**
+ * Phase 8.1 — record a pending verification for an autoApply (advisor-rule
+ * or prompt-override) so the verifier can roll the fix back if the next
+ * AUTOAPPLY_VERDICT_WINDOW runs don't show improvement.
+ *
+ * Called from fix-applier.applyFix on success. Idempotent: if a previous
+ * verification for the same job is still pending, the new tracker overwrites
+ * it (the most-recent fix is the one we're verifying).
+ */
+export declare function recordAutoApplyForVerification(jobName: string, tracker: AutoApplyTracker): void;
 /**
  * After a cron run completes, check whether we were waiting on a fix
- * verification for this job. If so, send the owner a verdict and clear it.
+ * verification for this job. Two flows:
+ *
+ *   1. Hand-edit (CRON.md) — verdict on the FIRST non-skipped run. Original
+ *      Phase 7 behavior, preserved.
+ *   2. AutoApply (advisor-rule / prompt-override) — accumulate up to
+ *      AUTOAPPLY_VERDICT_WINDOW outcomes, then decide. If 0 successes,
+ *      revert the file. Either way, DM the verdict.
  *
- * Skipped runs (circuit breaker, pre-check exit, etc.) don't carry signal
- * and shouldn't count as a verdict either way.
+ * Skipped runs don't carry signal and don't advance the window in either flow.
  */
 export declare function checkAndDeliverVerification(entry: CronRunEntry, send: (text: string) => Promise<unknown>): Promise<void>;
 /** Read-only accessor for dashboards or debugging. */
 export declare function listPendingVerifications(): PendingVerification[];
+/** Test helper — clear all state. */
+export declare function _resetVerificationState(): void;
 export {};
 //# sourceMappingURL=fix-verification.d.ts.map

package/dist/gateway/fix-verification.js

@@ -15,6 +15,13 @@ import { BASE_DIR } from '../config.js';
 import { computeBrokenJobs } from './failure-monitor.js';
 const logger = pino({ name: 'clementine.fix-verification' });
 const STATE_FILE = path.join(BASE_DIR, 'cron', 'fix-verifications.json');
+/**
+ * Number of post-fix runs we accumulate before deciding an autoApply
+ * verdict. Single sample is too noisy; ten is too patient. Three is
+ * a tight window: 0/3 successes after a "fix" is overwhelming evidence
+ * the fix didn't help.
+ */
+const AUTOAPPLY_VERDICT_WINDOW = 3;
 function loadState() {
     try {
         if (!existsSync(STATE_FILE))
@@ -109,12 +116,62 @@ export function recordEditsForFailingJobs(oldJobs, newJobs) {
     if (mutated)
         saveState(state);
 }
+/**
+ * Phase 8.1 — record a pending verification for an autoApply (advisor-rule
+ * or prompt-override) so the verifier can roll the fix back if the next
+ * AUTOAPPLY_VERDICT_WINDOW runs don't show improvement.
+ *
+ * Called from fix-applier.applyFix on success. Idempotent: if a previous
+ * verification for the same job is still pending, the new tracker overwrites
+ * it (the most-recent fix is the one we're verifying).
+ */
+export function recordAutoApplyForVerification(jobName, tracker) {
+    const state = loadState();
+    const broken = computeBrokenJobs();
+    const b = broken.find(x => x.jobName === jobName);
+    state.pending[jobName] = {
+        jobName,
+        recordedAt: new Date().toISOString(),
+        preFailureCount: b?.errorCount48h ?? 0,
+        preLastError: b?.lastErrors[0] ?? null,
+        autoApply: tracker,
+        postRunOutcomes: [],
+    };
+    saveState(state);
+    logger.info({ job: jobName, kind: tracker.kind, file: tracker.file }, 'Recorded autoApply for verification — will track next runs');
+}
+/**
+ * Undo an autoApply by deleting the file the apply wrote. Best-effort:
+ * a missing file is not an error (might have been hand-deleted). Returns
+ * true if a file was actually removed.
+ */
+function revertAutoApply(tracker) {
+    try {
+        if (existsSync(tracker.file)) {
+            // Use unlinkSync from fs — kept dynamic to avoid a top-of-file import
+            // we don't otherwise need.
+            const { unlinkSync } = require('node:fs');
+            unlinkSync(tracker.file);
+            logger.warn({ file: tracker.file, kind: tracker.kind }, 'Reverted autoApply — fix did not help');
+            return true;
+        }
+    }
+    catch (err) {
+        logger.warn({ err, file: tracker.file }, 'Failed to delete autoApply file during revert');
+    }
+    return false;
+}
 /**
  * After a cron run completes, check whether we were waiting on a fix
- * verification for this job. If so, send the owner a verdict and clear it.
+ * verification for this job. Two flows:
+ *
+ *   1. Hand-edit (CRON.md) — verdict on the FIRST non-skipped run. Original
+ *      Phase 7 behavior, preserved.
+ *   2. AutoApply (advisor-rule / prompt-override) — accumulate up to
+ *      AUTOAPPLY_VERDICT_WINDOW outcomes, then decide. If 0 successes,
+ *      revert the file. Either way, DM the verdict.
  *
- * Skipped runs (circuit breaker, pre-check exit, etc.) don't carry signal
- * and shouldn't count as a verdict either way.
+ * Skipped runs don't carry signal and don't advance the window in either flow.
  */
 export async function checkAndDeliverVerification(entry, send) {
     if (entry.status === 'skipped')
@@ -123,15 +180,60 @@ export async function checkAndDeliverVerification(entry, send) {
     const pending = state.pending[entry.jobName];
     if (!pending)
         return;
+    // Hand-edit flow — single-run verdict, unchanged.
+    if (!pending.autoApply) {
+        delete state.pending[entry.jobName];
+        saveState(state);
+        const ok = entry.status === 'ok';
+        const verdict = ok ? '✅ succeeded' : '⚠️ still failing';
+        const ageMin = Math.max(1, Math.round((Date.now() - Date.parse(pending.recordedAt)) / 60000));
+        const detail = ok ? '' : `\nError: ${(entry.error ?? 'unknown').split('\n')[0].slice(0, 200)}`;
+        const msg = `**[Fix verification]** \`${entry.jobName}\` ${verdict} on its first run after edit (${ageMin}m later).${detail}`;
+        try {
+            await send(msg);
+        }
+        catch (err) {
+            logger.warn({ err, job: entry.jobName }, 'Failed to send fix verification DM');
+        }
+        return;
+    }
+    // AutoApply flow — accumulate the sample first.
+    const outcomes = pending.postRunOutcomes ?? [];
+    outcomes.push(entry.status);
+    pending.postRunOutcomes = outcomes;
+    if (outcomes.length < AUTOAPPLY_VERDICT_WINDOW) {
+        // Not enough sample yet — persist accumulated state, wait for more runs.
+        saveState(state);
+        return;
+    }
+    // Decision time.
     delete state.pending[entry.jobName];
     saveState(state);
-    const ok = entry.status === 'ok';
-    const verdict = ok ? '✅ succeeded' : '⚠️ still failing';
+    const successes = outcomes.filter(o => o === 'ok').length;
     const ageMin = Math.max(1, Math.round((Date.now() - Date.parse(pending.recordedAt)) / 60000));
-    const detail = ok
-        ? ''
-        : `\nError: ${(entry.error ?? 'unknown').split('\n')[0].slice(0, 200)}`;
-    const msg = `**[Fix verification]** \`${entry.jobName}\` ${verdict} on its first run after edit (${ageMin}m later).${detail}`;
+    const tracker = pending.autoApply;
+    const scopeLabel = tracker.scope
+        ? `${tracker.kind}:${tracker.scope}${tracker.scopeKey ? `:${tracker.scopeKey}` : ''}`
+        : `${tracker.kind}${tracker.ruleId ? `:${tracker.ruleId}` : ''}`;
+    if (successes === 0) {
+        // Fix didn't help — revert and notify.
+        const reverted = revertAutoApply(tracker);
+        const msg = `**[Fix verification — REVERTED]** \`${entry.jobName}\`: ` +
+            `auto-applied ${scopeLabel} did not help (0/${outcomes.length} runs succeeded over ${ageMin}m). ` +
+            (reverted ? `Reverted ${path.basename(tracker.file)}.` : `Tried to revert but file was already gone.`);
+        try {
+            await send(msg);
+        }
+        catch (err) {
+            logger.warn({ err, job: entry.jobName }, 'Failed to send fix-revert DM');
+        }
+        logger.warn({ job: entry.jobName, scopeLabel, reverted }, 'Auto-reverted ineffective autoApply');
+        return;
+    }
+    const verdict = successes === outcomes.length
+        ? `✅ verified — ${successes}/${outcomes.length} runs succeeded`
+        : `⚠️ partial — ${successes}/${outcomes.length} runs succeeded`;
+    const msg = `**[Fix verification]** \`${entry.jobName}\`: auto-applied ${scopeLabel} ${verdict} over ${ageMin}m.`;
     try {
         await send(msg);
     }
@@ -143,4 +245,8 @@ export async function checkAndDeliverVerification(entry, send) {
 export function listPendingVerifications() {
     return Object.values(loadState().pending);
 }
+/** Test helper — clear all state. */
+export function _resetVerificationState() {
+    saveState({ pending: {} });
+}
 //# sourceMappingURL=fix-verification.js.map

package/dist/gateway/router.d.ts

@@ -170,7 +170,7 @@ export declare class Gateway {
      * Get recent transcript activity across all sessions.
      * Used by heartbeat to know what happened since the last check.
      */
-    getRecentActivity(sinceIso: string): Array<{
+    getRecentActivity(sinceIso: string, maxEntries?: number): Array<{
         sessionKey: string;
         role: string;
         content: string;

package/dist/gateway/router.js

@@ -1447,8 +1447,8 @@ export class Gateway {
      * Get recent transcript activity across all sessions.
      * Used by heartbeat to know what happened since the last check.
      */
-    getRecentActivity(sinceIso) {
-        return this.assistant.getRecentActivity(sinceIso);
+    getRecentActivity(sinceIso, maxEntries) {
+        return this.assistant.getRecentActivity(sinceIso, maxEntries);
     }
     /**
      * Search memory (FTS5) for context relevant to a query.

package/dist/memory/store.js

@@ -1094,6 +1094,7 @@ export class MemoryStore {
         }
         const rows = this.conn.prepare(sql).all(...params);
         const scored = [];
+        const nowMs = Date.now();
         for (const row of rows) {
             try {
                 const vec = embeddingsModule.deserializeEmbedding(row.embedding);
@@ -1109,6 +1110,16 @@ export class MemoryStore {
                 // Soft isolation: apply boost (only when not strict)
                 if (!strict && agentSlug && row.agent_slug === agentSlug)
                     score *= 1.4;
+                // Temporal decay — same policy as FTS scoring (Phase 9d). Without
+                // this, vector and FTS rankings disagree on freshness: FTS prefers
+                // recent at equal relevance but vector treats all timestamps
+                // equally, so MMR rerank surfaces stale matches when vector wins.
+                // Same 30-day half-life, same 0.4 floor — see store.ts FTS path
+                // for design rationale.
+                if (row.updated_at) {
+                    const daysOld = Math.max(0, (nowMs - new Date(row.updated_at).getTime()) / 86_400_000);
+                    score *= Math.max(0.4, temporalDecay(daysOld, 30));
+                }
                 scored.push({
                     sourceFile: row.source_file,
                     section: row.section,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "clementine-agent",
-  "version": "1.1.6",
+  "version": "1.1.8",
   "description": "Clementine — Personal AI Assistant (TypeScript)",
   "type": "module",
   "main": "dist/index.js",