@johpaz/hive-sdk 0.1.3 → 0.1.5

package/packages/core/src/agent/reflector.ts

@@ -0,0 +1,320 @@
+/**
+ * ACE Reflector — analyzes recent traces and produces insights.
+ *
+ * Runs in the background (never blocks the main agent loop).
+ * Triggered by the tracer after N new traces.
+ *
+ * Output → `reflections` collection → picked up by Curator.
+ *
+ * "Last processed trace" is tracked via a `cursors` collection doc
+ * (id="reflector:lastTrace") instead of the SQL json_each/MAX(CAST(...))
+ * trick the SQLite version used — HiveDB has no equivalent primitive.
+ */
+
+import { logger } from "../utils/logger"
+import { col, nextId } from "../storage/hive"
+import { getHiveDb } from "../storage/hivedb"
+import { loadConfig } from "../config/loader"
+import type { HiveDB, ToolStats } from "@johpaz/hive-db"
+import type { TraceDoc, ReflectionDoc, CursorDoc } from "../storage/collections"
+
+const log = logger.child("reflector")
+
+const MAX_TRACES_TO_ANALYZE = 30
+const MIN_TRACES_TO_RUN = 10
+const CURSOR_ID = "reflector:lastTrace"
+
+/** Main entry point — called from tracer.ts */
+export async function runReflector(): Promise<void> {
+  try {
+    log.info(`[reflector] Starting reflection cycle...`)
+
+    const cursorsCol = await col<CursorDoc>("cursors")
+    const cursorEntry = await cursorsCol.get(CURSOR_ID)
+    const lastProcessedId = cursorEntry?.doc.value ?? null
+    log.debug(`[reflector] Last processed trace ID: ${lastProcessedId ?? "(none)"}`)
+
+    const tracesCol = await col<TraceDoc>("traces")
+    let candidates = lastProcessedId
+      ? await tracesCol.scan({ start: lastProcessedId })
+      : await tracesCol.scan({})
+    if (lastProcessedId && candidates[0]?.id === lastProcessedId) candidates = candidates.slice(1)
+
+    const traceEntries = candidates.slice(0, MAX_TRACES_TO_ANALYZE)
+    const traces = traceEntries.map(e => e.doc)
+    log.debug(`[reflector] Fetched ${traces.length} traces from DB`)
+
+    if (traces.length < MIN_TRACES_TO_RUN) {
+      log.debug(`[reflector] Not enough traces (${traces.length}/${MIN_TRACES_TO_RUN}), skipping`)
+      return
+    }
+
+    log.info(`[reflector] Analyzing ${traces.length} traces...`)
+
+    // G9: when enabled, per-tool insights use whole-history stats from HiveDB's
+    // event log (toolStats) instead of counters built from just this batch, and
+    // causal threads add root-cause/learning-proposal insights on top.
+    const causalDb = loadConfig().causalLog?.enabled ? await getHiveDb() : null
+    const localInsights = await analyzeTracesLocally(traces, causalDb)
+    const causalInsights = await analyzeCausalThreads(traces, causalDb)
+    const insights = [...localInsights, ...causalInsights]
+
+    if (insights.length === 0) {
+      log.debug("[reflector] No insights generated")
+    } else {
+      const traceIds = JSON.stringify(traceEntries.map(e => e.id))
+      const reflectionsCol = await col<ReflectionDoc>("reflections")
+
+      for (const insight of insights) {
+        const id = await nextId("reflections")
+        await reflectionsCol.put(id, {
+          id,
+          trace_ids: traceIds,
+          insight_type: insight.type,
+          description: insight.description,
+          affected_tools: insight.affectedTools ? JSON.stringify(insight.affectedTools) : null,
+          affected_agents: insight.affectedAgents ? JSON.stringify(insight.affectedAgents) : null,
+          confidence: insight.confidence,
+          created_at: Date.now(),
+        })
+      }
+
+      log.info(`[reflector] Generated ${insights.length} insights`)
+    }
+
+    // Advance the cursor regardless of whether insights were generated —
+    // these traces have been considered either way.
+    const newCursor = traceEntries[traceEntries.length - 1].id
+    await cursorsCol.put(CURSOR_ID, { value: newCursor }, cursorEntry ? { expectedVersion: cursorEntry.version } : { expectedVersion: 0 })
+
+    // Trigger curator
+    const { runCurator } = await import("./curator")
+    await runCurator()
+
+    log.info(`[reflector] Reflection cycle completed successfully`)
+  } catch (err) {
+    log.error(`[reflector] Error during reflection:`, {
+      message: (err as Error).message,
+      stack: (err as Error).stack,
+    })
+  }
+}
+
+// ─── Local analysis (heuristic, no LLM call needed for basic patterns) ────────
+
+interface Insight {
+  type: "success_pattern" | "failure_pattern" | "optimization" | "ethics_violation" | "root_cause" | "learning_proposal"
+  description: string
+  affectedTools?: string[]
+  affectedAgents?: string[]
+  confidence: number
+}
+
+// ─── G9 causal-thread analysis (evaluateHarness) ──────────────────────────────
+
+interface DecisionNode {
+  seq: number
+  agent: string
+  description: string
+}
+
+interface CausalThreadShape {
+  decisions: DecisionNode[]
+  toolCalls: Array<{ seq: number; agent: string; tool: string }>
+  anomalies: unknown[]
+}
+
+interface HarnessFinding {
+  kind: "inefficientLoop" | "objectiveDrift" | "rootCause" | "insufficientEvidence"
+  seq?: number
+  description: string
+}
+
+interface HarnessProposal {
+  description: string
+  confidence: number
+}
+
+interface HarnessEvaluationShape {
+  processQuality: number
+  outputQuality: number
+  rootCause?: { seq: number; agent: string }
+  findings: HarnessFinding[]
+  proposals: HarnessProposal[]
+}
+
+/**
+ * Evaluate the causal thread of every distinct G9 stream touched by this
+ * batch of traces (one thread per agent-loop invocation — see
+ * agent-loop.ts's causalStreamId) with HiveDB's harness loop, and turn its
+ * root cause / findings / learning proposals into reflector insights.
+ *
+ * Runs alongside, not instead of, analyzeTracesLocally(): it needs a full
+ * causal thread per stream, so it operates per-run rather than per-batch.
+ */
+async function analyzeCausalThreads(traces: TraceDoc[], causalDb: HiveDB | null): Promise<Insight[]> {
+  if (!causalDb) return []
+
+  const insights: Insight[] = []
+  const streamIds = new Set(traces.map((t) => t.causal_stream_id).filter((s): s is string => !!s))
+
+  for (const streamId of streamIds) {
+    try {
+      const thread = (await causalDb.causalThread(streamId)) as CausalThreadShape
+      if (!thread.decisions?.length && !thread.toolCalls?.length) continue
+
+      const subset = traces.filter((t) => t.causal_stream_id === streamId)
+      const originalIntent = subset[0]?.input_summary ?? ""
+      const success = subset.every((t) => t.success)
+
+      const evaluation = (await causalDb.evaluateHarness({
+        causalThread: thread,
+        similarEpisodes: [],
+        originalIntent,
+        currentState: { success },
+        minConfidence: 0.5,
+      })) as HarnessEvaluationShape
+
+      // Description deliberately omits streamId/seq: curator.ts only reinforces
+      // an existing rule when its first-60-chars prefix matches exactly, so any
+      // per-run identifier in the text would make every occurrence of the same
+      // underlying root cause mint a brand new playbook rule instead of
+      // reinforcing one (confirmed via a local before/after canary run).
+      if (evaluation.rootCause) {
+        const decision = thread.decisions?.find((d) => d.seq === evaluation.rootCause!.seq)
+        insights.push({
+          type: "root_cause",
+          description: decision
+            ? `Root cause: decision "${decision.description}" (agent ${evaluation.rootCause.agent}) preceded a tool failure.`
+            : `Root cause: a decision by agent ${evaluation.rootCause.agent} preceded a tool failure.`,
+          affectedAgents: [evaluation.rootCause.agent],
+          confidence: 0.6,
+        })
+      }
+
+      // Only inefficientLoop findings here: harness.rs's find_root_cause() always
+      // emits evaluation.rootCause and an equivalent finding{kind:"rootCause"}
+      // together from the same resolution — including both would double-insight
+      // every single failure. The block above already covers it, with a richer
+      // description (the actual decision text, not just its seq).
+      for (const finding of evaluation.findings ?? []) {
+        if (finding.kind === "inefficientLoop") {
+          insights.push({
+            type: "root_cause",
+            description: finding.description,
+            confidence: 0.6,
+          })
+        }
+      }
+
+      for (const proposal of evaluation.proposals ?? []) {
+        insights.push({
+          type: "learning_proposal",
+          description: proposal.description,
+          confidence: proposal.confidence,
+        })
+      }
+    } catch (err) {
+      log.warn(`[reflector] Causal-thread analysis failed for stream ${streamId}: ${(err as Error).message}`)
+    }
+  }
+
+  return insights
+}
+
+async function analyzeTracesLocally(traces: TraceDoc[], causalDb: HiveDB | null): Promise<Insight[]> {
+  const insights: Insight[] = []
+
+  // G9: whole-history stats per tool touched by this batch (undefined when
+  // disabled, or when the tool has no events in the log yet).
+  const statsByTool = new Map<string, ToolStats>()
+  if (causalDb) {
+    const distinctTools = new Set(traces.map((t) => t.tool_used).filter((t): t is string => !!t))
+    for (const tool of distinctTools) {
+      try {
+        const stats = await causalDb.toolStats(tool)
+        if (stats) statsByTool.set(tool, stats)
+      } catch (err) {
+        log.warn(`[reflector] toolStats(${tool}) failed: ${(err as Error).message}`)
+      }
+    }
+  }
+
+  // ── Failure patterns ─────────────────────────────────────────────────────
+  const failures = traces.filter((t) => !t.success)
+  if (failures.length > 3) {
+    // Group by tool
+    const toolFailures: Record<string, number> = {}
+    for (const f of failures) {
+      if (f.tool_used) {
+        toolFailures[f.tool_used] = (toolFailures[f.tool_used] || 0) + 1
+      }
+    }
+    for (const [tool, batchCount] of Object.entries(toolFailures)) {
+      const count = statsByTool.get(tool)?.errors ?? batchCount
+      if (count >= 3) {
+        insights.push({
+          type: "failure_pattern",
+          description: `Tool '${tool}' failed ${count} times recently. Consider verifying its configuration or avoiding it for this type of task.`,
+          affectedTools: [tool],
+          confidence: Math.min(0.9, count / 10),
+        })
+      }
+    }
+  }
+
+  // ── Slow tools ───────────────────────────────────────────────────────────
+  const slowThresholdMs = 5000
+  const slowTools: Record<string, number[]> = {}
+  for (const t of traces) {
+    if (t.tool_used && (t.duration_ms ?? 0) > slowThresholdMs) {
+      if (!slowTools[t.tool_used]) slowTools[t.tool_used] = []
+      slowTools[t.tool_used].push(t.duration_ms!)
+    }
+  }
+  for (const [tool, durations] of Object.entries(slowTools)) {
+    if (durations.length >= 3) {
+      const stats = statsByTool.get(tool)
+      const avg = stats && stats.invocations > 0
+        ? Math.round(stats.totalLatencyMs / stats.invocations)
+        : Math.round(durations.reduce((a, b) => a + b, 0) / durations.length)
+      insights.push({
+        type: "optimization",
+        description: `Tool '${tool}' is consistently slow (avg ${avg}ms). Cache results when possible or use a faster alternative.`,
+        affectedTools: [tool],
+        confidence: 0.6,
+      })
+    }
+  }
+
+  // ── High success pattern ─────────────────────────────────────────────────
+  const successByTool: Record<string, { ok: number; total: number }> = {}
+  for (const t of traces) {
+    if (!t.tool_used) continue
+    if (!successByTool[t.tool_used]) successByTool[t.tool_used] = { ok: 0, total: 0 }
+    successByTool[t.tool_used].total++
+    if (t.success) successByTool[t.tool_used].ok++
+  }
+  for (const [tool, stats] of Object.entries(successByTool)) {
+    if (stats.total >= 5 && stats.ok / stats.total >= 0.9) {
+      insights.push({
+        type: "success_pattern",
+        description: `Tool '${tool}' has a high success rate (${stats.ok}/${stats.total}). Prefer it for related tasks.`,
+        affectedTools: [tool],
+        confidence: stats.ok / stats.total,
+      })
+    }
+  }
+
+  // ── High token usage ─────────────────────────────────────────────────────
+  const highTokenTraces = traces.filter((t) => (t.tokens_used ?? 0) > 4000)
+  if (highTokenTraces.length > 3) {
+    insights.push({
+      type: "optimization",
+      description: `${highTokenTraces.length} recent calls used >4000 tokens. Be more concise and use tool results as summaries, not raw dumps.`,
+      confidence: 0.7,
+    })
+  }
+
+  return insights
+}

package/packages/core/src/agent/routing-intent.ts

@@ -0,0 +1,22 @@
+function normalizeIntent(value: string): string {
+  return value
+    .normalize("NFD")
+    .replace(/\p{Diacritic}/gu, "")
+    .toLowerCase();
+}
+
+/**
+ * Calendar operations are external data mutations/lookups, not future Hive
+ * executions. Keep them out of the native cron tool and skill selectors so
+ * the coordinator can discover the configured calendar MCP server instead.
+ */
+export function isCalendarOperation(value: string): boolean {
+  const text = normalizeIntent(value);
+  const namesCalendar = /\b(calendario|calendar|agenda)\b/.test(text);
+  const calendarAction = /\b(crear?|agend\w*|anad\w*|agreg\w*|consult\w*|list\w*|modific\w*|mov\w*|cancel\w*|elimin\w*|invit\w*|reserv\w*|disponibilidad|availability)\b/.test(text);
+  if (namesCalendar && calendarAction) return true;
+
+  const directlySchedulesMeeting = /\b(crea(?:r)?|agenda(?:r)?|programa(?:r)?|reserva(?:r)?)\s+(?:un|una|la|el|nueva|nuevo)?\s*(reunion|cita|meeting)\b/.test(text);
+  const managesAttendees = /\b(invit\w*|asistente\w*|attendee\w*)\b.*\b(reunion|cita|meeting)\b|\b(reunion|cita|meeting)\b.*\b(invit\w*|asistente\w*|attendee\w*)\b/.test(text);
+  return directlySchedulesMeeting || managesAttendees;
+}

package/packages/core/src/{harness → agent}/run-epoch.ts

@@ -6,10 +6,11 @@
  * were unchanged.
  */
 
+import rootPackage from "../../../../package.json";
+
 export interface RunEpoch {
   provider: string;
   model: string;
-  /** Caller-supplied version identifier (e.g. the host app's package version) — the harness doesn't assume its own version is what matters. */
   app_version: string;
   tool_catalog_hash: string;
 }
@@ -22,11 +23,11 @@ function hashToolNames(names: string[]): string {
   return hash.toString(16);
 }
 
-export function buildRunEpoch(opts: { provider: string; model: string; appVersion: string; toolNames: string[] }): RunEpoch {
+export function buildRunEpoch(opts: { provider: string; model: string; toolNames: string[] }): RunEpoch {
   return {
     provider: opts.provider,
     model: opts.model,
-    app_version: opts.appVersion,
+    app_version: (rootPackage as { version: string }).version,
     tool_catalog_hash: hashToolNames(opts.toolNames),
   };
 }