clementine-agent 1.1.8 → 1.1.10

package/dist/agent/assistant.js

@@ -1506,6 +1506,33 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
             }
             catch { /* non-fatal */ }
         }
+        // Conversational context — same signals the insight engine surfaces
+        // proactively (Phase 10), but injected directly into the agent's prompt
+        // so it can adjust its own approach. Scoped to chat sessions because
+        // cron/heartbeat don't have a "user feeling frustrated" axis to react to,
+        // and inflating their prompt doesn't help. Only injected when at least
+        // one signal fires — keeps the prompt clean during normal sessions.
+        if (!isAutonomous) {
+            try {
+                const { detectFrustrationSignals, detectRepeatedTopics } = require('./insight-engine.js');
+                const since24h = new Date(Date.now() - 24 * 60 * 60 * 1000).toISOString();
+                const since7d = new Date(Date.now() - 7 * 24 * 60 * 60 * 1000).toISOString();
+                const recent = this.getRecentActivity(since24h, 50);
+                const week = this.getRecentActivity(since7d, 200);
+                const frustration = detectFrustrationSignals(recent);
+                const topics = detectRepeatedTopics(week);
+                const allSignals = [...frustration, ...topics];
+                if (allSignals.length > 0) {
+                    const guidance = frustration.length > 0
+                        ? '\n\n**Adjust your approach:** When friction signals are present, lead with a clarifying question instead of assuming. Acknowledge the prior misunderstanding briefly without over-apologizing. Confirm understanding before acting.'
+                        : '\n\n**Use this context naturally:** Recurring topics may indicate an unresolved thread — if relevant, offer to close the loop or summarize current state. Do not force callbacks if not directly applicable.';
+                    volatileParts.push(`## Conversational Context\n\nSignals from recent sessions:\n` +
+                        allSignals.map(s => `- ${s}`).join('\n') +
+                        guidance);
+                }
+            }
+            catch { /* non-fatal — insight-engine optional */ }
+        }
         // Current context — date/time changes every minute, so it's volatile.
         const channel = deriveChannel({ sessionKey, isAutonomous, cronTier });
         const resolvedModel = resolveModel(model) ?? MODEL;
@@ -3786,6 +3813,12 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
     }
     async runCronJob(jobName, jobPrompt, tier = 1, maxTurns, model, workDir, timeoutMs, successCriteria, agentSlug) {
         setInteractionSource('autonomous');
+        // Tag every tool_use audit event with the cron job name + agent so
+        // analytics tool-usage can show "Bash×893 driven by market-leader-followup"
+        // instead of "driven by: unknown". Cleared on next setInteractionSource
+        // (cron/heartbeat boundary or interactive chat takeover).
+        const { setActiveQueryContext } = await import('./hooks.js');
+        setActiveQueryContext({ job: jobName, source: 'cron', agentSlug });
         const cronProfile = agentSlug && agentSlug !== 'clementine'
             ? this.profileManager.get(agentSlug)
             : null;
@@ -4274,6 +4307,10 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
             logger.info(`Unleashed task ${jobName}: starting phase ${phase}`);
             // Re-assert autonomous source — a chat message may have changed it between phases
             setInteractionSource('autonomous');
+            // Tag tool_use audit events with the unleashed job name (Phase 11).
+            // Re-asserted each phase since setInteractionSource clears the context.
+            const { setActiveQueryContext: _setActiveQueryContext } = await import('./hooks.js');
+            _setActiveQueryContext({ job: jobName, source: 'unleashed', agentSlug });
             // Unleashed phases run side-effect-heavy work; same logic as cron mode.
             const phaseGuard = new StallGuard('unleashed');
             const sdkOptions = this.buildOptions({

package/dist/agent/hooks.d.ts

@@ -62,6 +62,12 @@ export declare function getInteractionSource(): 'owner-dm' | 'owner-channel' | '
 export declare function getProfileTier(): number | null;
 export declare function getAuditLog(): string[];
 export declare function clearAuditLog(): void;
+export declare function setActiveQueryContext(ctx: {
+    job?: string | null;
+    source?: string | null;
+    agentSlug?: string | null;
+}): void;
+export declare function clearActiveQueryContext(): void;
 export declare function logToolUse(toolName: string, toolInput: Record<string, unknown>): void;
 export declare function getHeartbeatDisallowedTools(): string[];
 export declare const PRIVATE_URL_PATTERNS: RegExp[];

package/dist/agent/hooks.js

@@ -141,6 +141,11 @@ export function setSendPolicyChecker(checker) {
 }
 export function setInteractionSource(source) {
     interactionSource = source;
+    // Clear any leftover query attribution context. Cron / unleashed paths
+    // immediately call setActiveQueryContext after this; interactive chat
+    // doesn't, so anything still set from a prior cron run gets reset.
+    activeJob = null;
+    activeSource = null;
 }
 export function getInteractionSource() {
     return interactionSource;
@@ -154,6 +159,25 @@ export function getAuditLog() {
 export function clearAuditLog() {
     auditLog.length = 0;
 }
+// Ambient job/source context so audit tool_use events carry attribution.
+// Set by the assistant before running a query; cleared after. Without this
+// the analytics view shows everything as "driven by: unknown". The
+// activeAgentSlug field is already declared above (line ~27) for the
+// existing send-policy infrastructure — we read but don't redeclare it.
+let activeJob = null;
+let activeSource = null;
+export function setActiveQueryContext(ctx) {
+    activeJob = ctx.job ?? null;
+    activeSource = ctx.source ?? null;
+    if (ctx.agentSlug !== undefined)
+        activeAgentSlug = ctx.agentSlug;
+}
+export function clearActiveQueryContext() {
+    activeJob = null;
+    activeSource = null;
+    // Don't clear activeAgentSlug — it's owned by the send-policy path,
+    // not by us. setInteractionSource resets it in the relevant transitions.
+}
 export function logToolUse(toolName, toolInput) {
     const timestamp = new Date().toLocaleTimeString('en-US', { hour12: false });
     const summary = summarizeToolCall(toolName, toolInput);
@@ -164,6 +188,9 @@ export function logToolUse(toolName, toolInput) {
         event_type: 'tool_use',
         tool_name: toolName,
         summary,
+        ...(activeJob ? { job: activeJob } : {}),
+        ...(activeSource ? { source: activeSource } : {}),
+        ...(activeAgentSlug ? { agent_slug: activeAgentSlug } : {}),
     });
 }
 // ── Heartbeat tool restrictions ─────────────────────────────────────

package/dist/analytics/tool-usage.d.ts

@@ -0,0 +1,61 @@
+/**
+ * Tool-usage analytics.
+ *
+ * Reads ~/.clementine/logs/audit.jsonl and aggregates tool_use events by
+ * family + name + source so a CLI report can answer:
+ *
+ *   - "What is the agent spending its tool calls on?"
+ *   - "Which integration (mcp__ family) is hottest?"
+ *   - "Which job/source is the biggest tool consumer?"
+ *
+ * Pure file read + in-memory aggregation — no daemon access required.
+ * Designed to run on multi-MB audit logs without buffering everything;
+ * we stream line-by-line.
+ */
+export interface ToolFamilyStats {
+    /** Family label — collapses mcp__ subnames into one bucket per server. */
+    family: string;
+    totalCalls: number;
+    /** Estimated cost attributed to this family (USD). Heuristic — see attributeCostsToToolUses. */
+    estimatedCostUsd: number;
+    /** Per-tool breakdown within the family, sorted by count desc. */
+    byTool: Array<{
+        tool: string;
+        count: number;
+    }>;
+    /** Per-source breakdown — which job/context drives this family. */
+    bySource: Array<{
+        source: string;
+        count: number;
+    }>;
+}
+export interface ToolUsageReport {
+    windowStart: string;
+    windowEnd: string;
+    totalToolCalls: number;
+    totalQueries: number;
+    families: ToolFamilyStats[];
+    /** Total cost (sum of query_complete events) over the window — context for tool counts. */
+    totalCostUsd: number;
+    /** Sum of cost attributed to tool calls (≤ totalCostUsd). The gap is the cost of
+     *  query_completes whose tool calls fell outside the window or weren't logged. */
+    attributedCostUsd: number;
+}
+/**
+ * Family normalization. Built-in SDK tools keep their name; MCP tools are
+ * grouped by server (mcp__<server>__<tool> → "mcp:<server>"). Anything
+ * else falls into "other".
+ */
+export declare function classifyToolFamily(toolName: string): string;
+/**
+ * Aggregate tool_use + query_complete events from audit.jsonl over the
+ * given window. Window bounds are ISO strings; entries outside are ignored.
+ *
+ * The function is forgiving: malformed lines are skipped, missing fields
+ * default to 'unknown'. Audit logs are append-only so we never need to
+ * worry about ordering.
+ */
+export declare function buildToolUsageReport(auditLogPath: string, windowStart: string, windowEnd: string): ToolUsageReport;
+/** Default audit log path — passed-through for CLI default + tests. */
+export declare function defaultAuditLogPath(baseDir: string): string;
+//# sourceMappingURL=tool-usage.d.ts.map

package/dist/analytics/tool-usage.js

@@ -0,0 +1,186 @@
+/**
+ * Tool-usage analytics.
+ *
+ * Reads ~/.clementine/logs/audit.jsonl and aggregates tool_use events by
+ * family + name + source so a CLI report can answer:
+ *
+ *   - "What is the agent spending its tool calls on?"
+ *   - "Which integration (mcp__ family) is hottest?"
+ *   - "Which job/source is the biggest tool consumer?"
+ *
+ * Pure file read + in-memory aggregation — no daemon access required.
+ * Designed to run on multi-MB audit logs without buffering everything;
+ * we stream line-by-line.
+ */
+import { existsSync, readFileSync } from 'node:fs';
+import path from 'node:path';
+/**
+ * Family normalization. Built-in SDK tools keep their name; MCP tools are
+ * grouped by server (mcp__<server>__<tool> → "mcp:<server>"). Anything
+ * else falls into "other".
+ */
+export function classifyToolFamily(toolName) {
+    if (!toolName)
+        return 'other';
+    // mcp__server-name__tool_name → mcp:server-name
+    const mcpMatch = toolName.match(/^mcp__([^_]+(?:[-_][^_]+)*)__/);
+    if (mcpMatch)
+        return `mcp:${mcpMatch[1]}`;
+    // Built-ins kept as their own families
+    const BUILTIN_FAMILIES = {
+        Bash: 'shell',
+        Read: 'fs-read',
+        Glob: 'fs-read',
+        Grep: 'fs-read',
+        Edit: 'fs-write',
+        Write: 'fs-write',
+        NotebookEdit: 'fs-write',
+        WebFetch: 'web',
+        WebSearch: 'web',
+        Agent: 'subagent',
+        Task: 'subagent',
+    };
+    return BUILTIN_FAMILIES[toolName] ?? toolName;
+}
+/**
+ * Time-proximity cost attribution. Audit events don't carry an explicit
+ * query_id linking tool_use to query_complete, so we group by a sliding
+ * window: tool_use events that occur AFTER the previous query_complete
+ * (or the window start) and AT-OR-BEFORE the next query_complete are
+ * attributed to that query. The query's cost is then divided evenly
+ * across the tool calls in its window.
+ *
+ * Caveats:
+ * - Concurrent queries (e.g. cron + chat in the same window) will mix.
+ *   Best-effort heuristic, not exact accounting.
+ * - Tool calls without a closing query_complete in the window get
+ *   attributed nothing — captured in the gap between totalCostUsd
+ *   and attributedCostUsd in the report.
+ * - The even-distribution assumption ignores per-call cost variance
+ *   (a single Bash that consumed 50k tokens vs a Read that consumed
+ *   200). For our purposes (aggregate "where is my budget going?")
+ *   this is good enough — actionable to within ~15% per family.
+ */
+function attributeCostsToToolUses(events) {
+    const perToolCost = new Map();
+    let pendingToolIndices = [];
+    for (const e of events) {
+        if (!e.isQueryComplete) {
+            if (e.toolEntryIndex !== undefined)
+                pendingToolIndices.push(e.toolEntryIndex);
+            continue;
+        }
+        // Query closed — distribute cost.
+        if (pendingToolIndices.length > 0 && typeof e.cost_usd === 'number' && Number.isFinite(e.cost_usd)) {
+            const perCall = e.cost_usd / pendingToolIndices.length;
+            for (const idx of pendingToolIndices) {
+                perToolCost.set(idx, (perToolCost.get(idx) ?? 0) + perCall);
+            }
+        }
+        pendingToolIndices = [];
+    }
+    return perToolCost;
+}
+/**
+ * Aggregate tool_use + query_complete events from audit.jsonl over the
+ * given window. Window bounds are ISO strings; entries outside are ignored.
+ *
+ * The function is forgiving: malformed lines are skipped, missing fields
+ * default to 'unknown'. Audit logs are append-only so we never need to
+ * worry about ordering.
+ */
+export function buildToolUsageReport(auditLogPath, windowStart, windowEnd) {
+    const startMs = Date.parse(windowStart);
+    const endMs = Date.parse(windowEnd);
+    // family → { totalCalls, totalCost, perTool, perSource }
+    const families = new Map();
+    let totalToolCalls = 0;
+    let totalQueries = 0;
+    let totalCost = 0;
+    if (!existsSync(auditLogPath)) {
+        return {
+            windowStart, windowEnd, totalToolCalls: 0, totalQueries: 0,
+            families: [], totalCostUsd: 0, attributedCostUsd: 0,
+        };
+    }
+    const raw = readFileSync(auditLogPath, 'utf-8');
+    const toolEntries = [];
+    const sequence = [];
+    for (const line of raw.split('\n')) {
+        if (!line)
+            continue;
+        let entry;
+        try {
+            entry = JSON.parse(line);
+        }
+        catch {
+            continue;
+        }
+        if (!entry.ts)
+            continue;
+        const tsMs = Date.parse(entry.ts);
+        if (Number.isNaN(tsMs))
+            continue;
+        if (tsMs < startMs || tsMs > endMs)
+            continue;
+        if (entry.event_type === 'tool_use' && entry.tool_name) {
+            const family = classifyToolFamily(entry.tool_name);
+            const source = entry.job || entry.source || 'unknown';
+            toolEntries.push({ family, source, toolName: entry.tool_name });
+            sequence.push({ ts: tsMs, isQueryComplete: false, toolEntryIndex: toolEntries.length - 1 });
+            totalToolCalls++;
+        }
+        else if (entry.event_type === 'query_complete') {
+            totalQueries++;
+            const cost = typeof entry.cost_usd === 'number' && Number.isFinite(entry.cost_usd) ? entry.cost_usd : 0;
+            totalCost += cost;
+            sequence.push({ ts: tsMs, isQueryComplete: true, cost_usd: cost });
+        }
+    }
+    // Second pass: attribute each query's cost across its preceding tool_use events.
+    const perToolCost = attributeCostsToToolUses(sequence);
+    let attributedCost = 0;
+    // Third pass: bucket toolEntries into family stats, summing attributed cost.
+    for (let i = 0; i < toolEntries.length; i++) {
+        const t = toolEntries[i];
+        const cost = perToolCost.get(i) ?? 0;
+        attributedCost += cost;
+        let bucket = families.get(t.family);
+        if (!bucket) {
+            bucket = { totalCalls: 0, totalCost: 0, perTool: new Map(), perSource: new Map() };
+            families.set(t.family, bucket);
+        }
+        bucket.totalCalls++;
+        bucket.totalCost += cost;
+        bucket.perTool.set(t.toolName, (bucket.perTool.get(t.toolName) ?? 0) + 1);
+        bucket.perSource.set(t.source, (bucket.perSource.get(t.source) ?? 0) + 1);
+    }
+    const familyStats = [...families.entries()]
+        .map(([family, b]) => ({
+        family,
+        totalCalls: b.totalCalls,
+        estimatedCostUsd: Number(b.totalCost.toFixed(4)),
+        byTool: [...b.perTool.entries()]
+            .map(([tool, count]) => ({ tool, count }))
+            .sort((a, c) => c.count - a.count),
+        bySource: [...b.perSource.entries()]
+            .map(([source, count]) => ({ source, count }))
+            .sort((a, c) => c.count - a.count),
+    }))
+        // Sort by cost first (the actionable signal); fall back to call count.
+        .sort((a, b) => b.estimatedCostUsd - a.estimatedCostUsd || b.totalCalls - a.totalCalls);
+    return {
+        windowStart,
+        windowEnd,
+        totalToolCalls,
+        totalQueries,
+        families: familyStats,
+        totalCostUsd: Number(totalCost.toFixed(4)),
+        attributedCostUsd: Number(attributedCost.toFixed(4)),
+    };
+}
+/** Default audit log path — passed-through for CLI default + tests. */
+export function defaultAuditLogPath(baseDir) {
+    return path.join(baseDir, 'logs', 'audit.jsonl');
+}
+//# sourceMappingURL=tool-usage.js.map

package/dist/cli/index.js

@@ -1270,6 +1270,59 @@ async function cmdConfigKeychainFixAcl(opts) {
     }
     console.log();
 }
+// ── Analytics ────────────────────────────────────────────────────────
+async function cmdAnalyticsToolUsage(opts) {
+    const { buildToolUsageReport, defaultAuditLogPath } = await import('../analytics/tool-usage.js');
+    const hours = Math.max(1, parseInt(opts.hours ?? '24', 10) || 24);
+    const limit = Math.max(1, parseInt(opts.limit ?? '10', 10) || 10);
+    const end = new Date();
+    const start = new Date(end.getTime() - hours * 60 * 60 * 1000);
+    const report = buildToolUsageReport(defaultAuditLogPath(BASE_DIR), start.toISOString(), end.toISOString());
+    if (opts.json) {
+        console.log(JSON.stringify(report, null, 2));
+        return;
+    }
+    const DIM = '\x1b[0;90m';
+    const BOLD = '\x1b[1m';
+    const CYAN = '\x1b[0;36m';
+    const GREEN = '\x1b[0;32m';
+    const YELLOW = '\x1b[0;33m';
+    const RESET = '\x1b[0m';
+    console.log();
+    console.log(`  ${BOLD}Window:${RESET} last ${hours}h ${DIM}(${start.toISOString()} → ${end.toISOString()})${RESET}`);
+    console.log(`  ${BOLD}Total tool calls:${RESET} ${report.totalToolCalls.toLocaleString()}`);
+    console.log(`  ${BOLD}Total queries:${RESET}    ${report.totalQueries.toLocaleString()}`);
+    console.log(`  ${BOLD}Total cost:${RESET}       ${GREEN}$${report.totalCostUsd.toFixed(4)}${RESET} ${DIM}(attributed to tools: $${report.attributedCostUsd.toFixed(4)})${RESET}`);
+    console.log();
+    if (report.families.length === 0) {
+        console.log(`  ${DIM}No tool_use events in window.${RESET}`);
+        console.log();
+        return;
+    }
+    const top = report.families.slice(0, limit);
+    const maxCost = Math.max(...top.map(f => f.estimatedCostUsd), 0.0001);
+    const familyWidth = Math.max(...top.map(f => f.family.length), 12);
+    console.log(`  ${BOLD}Top ${top.length} tool families ${DIM}(ranked by attributed cost)${RESET}`);
+    for (const f of top) {
+        const pct = report.attributedCostUsd > 0
+            ? ((f.estimatedCostUsd / report.attributedCostUsd) * 100).toFixed(1)
+            : '0.0';
+        const barLen = Math.round((f.estimatedCostUsd / maxCost) * 24);
+        const bar = '█'.repeat(barLen).padEnd(24);
+        console.log(`    ${CYAN}${f.family.padEnd(familyWidth)}${RESET}  ` +
+            `${GREEN}$${f.estimatedCostUsd.toFixed(2).padStart(7)}${RESET} ` +
+            `${pct.padStart(5)}%  ` +
+            `${DIM}${String(f.totalCalls).padStart(5)} calls${RESET}  ` +
+            `${YELLOW}${bar}${RESET}`);
+        const topTools = f.byTool.slice(0, 2).map(t => `${t.tool}×${t.count}`).join(', ');
+        const topSource = f.bySource[0];
+        console.log(`      ${DIM}top tools: ${topTools}${RESET}`);
+        if (topSource && topSource.source !== 'unknown') {
+            console.log(`      ${DIM}driven by: ${topSource.source} (${topSource.count} calls)${RESET}`);
+        }
+    }
+    console.log();
+}
 // ── Advisor commands ────────────────────────────────────────────────
 const ADVISOR_MODES = ['off', 'shadow', 'primary'];
 function readAdvisorMode() {
@@ -1817,6 +1870,18 @@ advisorCmd
     .command('rules')
     .description('List loaded advisor rules')
     .action(cmdAdvisorRules);
+const analyticsCmd = program
+    .command('analytics')
+    .description('Production telemetry: tool usage, cost breakdowns');
+analyticsCmd
+    .command('tool-usage')
+    .description('Show which tool families are firing most over a time window')
+    .option('-h, --hours <n>', 'Window size in hours (default 24)', '24')
+    .option('--json', 'Emit machine-readable JSON')
+    .option('-l, --limit <n>', 'Show top N families (default 10)', '10')
+    .action(async (opts) => {
+    await cmdAnalyticsToolUsage(opts);
+});
 const dashCmd = program
     .command('dashboard')
     .description('Launch local command center')

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "clementine-agent",
-  "version": "1.1.8",
+  "version": "1.1.10",
   "description": "Clementine — Personal AI Assistant (TypeScript)",
   "type": "module",
   "main": "dist/index.js",