kc-beta 0.5.6 → 0.6.1

package/src/agent/engine.js

@@ -19,6 +19,9 @@ import { ReleaseTool } from "./tools/release.js";
 import { PhaseAdvanceTool } from "./tools/phase-advance.js";
 import { DocumentParseTool } from "./tools/document-parse.js";
 import { DocumentSearchTool } from "./tools/document-search.js";
+import { DocumentChunkTool } from "./tools/document-chunk.js";
+import { BundleSearchTool } from "./tools/bundle-search.js";
+import { DocumentClassifyTool } from "./tools/document-classify.js";
 import { WorkerLLMCallTool } from "./tools/worker-llm-call.js";
 import { WorkflowRunTool } from "./tools/workflow-run.js";
 import { RuleCatalogTool } from "./tools/rule-catalog.js";
@@ -38,6 +41,7 @@ import { SkillAuthoringPipeline } from "./pipelines/skill-authoring.js";
 import { SkillTestingPipeline } from "./pipelines/skill-testing.js";
 import { DistillationEngine as DistillationPipeline } from "./pipelines/distillation.js";
 import { ProductionQCPipeline } from "./pipelines/production-qc.js";
+import { FinalizationPipeline } from "./pipelines/finalization.js";
 import { EventLog } from "./event-log.js";
 import { ContextWindow } from "./context-window.js";
 import { SessionState } from "./session-state.js";
@@ -48,8 +52,10 @@ import { estimateTokens, estimateMessagesTokens } from "./token-counter.js";
 // or kc_max_tokens in the global config.
 const DEFAULT_KC_MAX_TOKENS = 65536;
 
-// Phases where worker LLM tools are available (DISTILL mode)
-const DISTILL_PHASES = new Set([Phase.DISTILLATION, Phase.PRODUCTION_QC]);
+// Phases where worker LLM tools are available (DISTILL mode).
+// E1: FINALIZATION inherits worker-LLM access so one-last-pass validation
+// runs + dashboard_render + workflow_run stay usable during packaging.
+const DISTILL_PHASES = new Set([Phase.DISTILLATION, Phase.PRODUCTION_QC, Phase.FINALIZATION]);
 
 // Linear phase order — used by auto-advance (Bug 4). Last phase has no successor.
 // Exported so the TUI's /phase slash command (src/cli/index.js) can call
@@ -60,6 +66,7 @@ export const NEXT_PHASE = {
   [Phase.SKILL_AUTHORING]: Phase.SKILL_TESTING,
   [Phase.SKILL_TESTING]: Phase.DISTILLATION,
   [Phase.DISTILLATION]: Phase.PRODUCTION_QC,
+  [Phase.PRODUCTION_QC]: Phase.FINALIZATION, // E1: new 7th phase
 };
 
 /**
@@ -158,10 +165,11 @@ export class AgentEngine {
     this.pipelines = {
       [Phase.BOOTSTRAP]: new ProjectInitializer(this.workspace),
       [Phase.EXTRACTION]: new RuleExtractionPipeline(this.workspace),
-      [Phase.SKILL_AUTHORING]: new SkillAuthoringPipeline(this.workspace),
+      [Phase.SKILL_AUTHORING]: new SkillAuthoringPipeline(this.workspace, this.taskManager),
       [Phase.SKILL_TESTING]: new SkillTestingPipeline(this.workspace),
       [Phase.DISTILLATION]: new DistillationPipeline(this.workspace),
       [Phase.PRODUCTION_QC]: new ProductionQCPipeline(this.workspace),
+      [Phase.FINALIZATION]: new FinalizationPipeline(this.workspace), // E1
     };
 
     // Skill discovery (Claude Code pattern: index in context, full content on demand)
@@ -181,6 +189,61 @@ export class AgentEngine {
     this._lastReady = Object.fromEntries(
       Object.keys(this.pipelines).map((p) => [p, false]),
     );
+
+    // B0.1: Heap sampler. Parent engines only — sub-agents share a process
+    // with the parent and would double-log. Writes a single JSONL line
+    // per minute to <workspace>/logs/heap.jsonl with the numbers needed
+    // to diagnose RSS creep (heapUsed/heapTotal/external/rss/arrayBuffers,
+    // plus active task count and history length). Always on, ~60 bytes
+    // per minute to disk.
+    this._heapSamplerStop = this._isSubagent ? null : this._startHeapSampler();
+  }
+
+  /**
+   * Start sampling process.memoryUsage() every 60 s into logs/heap.jsonl.
+   * Returns a stop fn. Timer is .unref()'d so it never keeps the process
+   * alive by itself. Failures are silently suppressed — this is a
+   * diagnostic, not a correctness feature.
+   */
+  _startHeapSampler() {
+    const logDir = path.join(this.workspace.cwd, "logs");
+    const logPath = path.join(logDir, "heap.jsonl");
+    const sample = () => {
+      try {
+        const mem = process.memoryUsage();
+        const row = {
+          t: new Date().toISOString(),
+          seq: this.eventLog?.currentSeq ?? 0,
+          phase: this.currentPhase,
+          rssMB: Math.round(mem.rss / 1024 / 1024),
+          heapUsedMB: Math.round(mem.heapUsed / 1024 / 1024),
+          heapTotalMB: Math.round(mem.heapTotal / 1024 / 1024),
+          externalMB: Math.round((mem.external || 0) / 1024 / 1024),
+          arrayBuffersMB: Math.round((mem.arrayBuffers || 0) / 1024 / 1024),
+          historyLen: this.history?.messages?.length ?? 0,
+          tasksPending: this.taskManager?.progress?.pending ?? 0,
+          tasksInProgress: this.taskManager?.progress?.inProgress ?? 0,
+        };
+        fs.mkdirSync(logDir, { recursive: true });
+        fs.appendFileSync(logPath, JSON.stringify(row) + "\n", "utf-8");
+      } catch { /* never fatal */ }
+    };
+    // Record one sample at startup so we have a baseline even on short runs.
+    sample();
+    const timer = setInterval(sample, 60_000);
+    timer.unref?.();
+    return () => {
+      try {
+        clearInterval(timer);
+        sample(); // one final sample on shutdown
+      } catch { /* ignore */ }
+    };
+  }
+
+  /** Stop background diagnostics. Call on graceful shutdown. */
+  stop() {
+    try { this._heapSamplerStop?.(); } catch { /* ignore */ }
+    this._heapSamplerStop = null;
   }
 
   /**
@@ -214,7 +277,10 @@ export class AgentEngine {
         new ArchiveFileTool(this.workspace),
         new ScheduleFetchTool(this.workspace),
         new ReleaseTool(this.workspace, { kcVersion: "0.5.2" }),
-        new PhaseAdvanceTool((to, reason, opts) => this._advancePhase(to, reason, opts)),
+        new PhaseAdvanceTool(
+          (to, reason, opts) => this._advancePhase(to, reason, opts),
+          () => this.currentPhase, // H1: tool reads phase BEFORE its own call
+        ),
         new DocumentParseTool(this.workspace, {
           mineruApiUrl: this.config.mineruApiUrl,
           mineruApiKey: this.config.mineruApiKey,
@@ -223,6 +289,12 @@ export class AgentEngine {
           ocrModel: vlmModel,
         }),
         new DocumentSearchTool(this.workspace),
+        // Group C — chunker/RAG infrastructure ported from AMC app. Core
+        // tools (not phase-gated): useful from BOOTSTRAP through FINALIZATION
+        // for any doc-heavy project, not just rule extraction.
+        new DocumentChunkTool(this.workspace),
+        new BundleSearchTool(this.workspace),
+        new DocumentClassifyTool(this.workspace, this.config),
         new RuleCatalogTool(this.workspace),
         new EvolutionCycleTool(this.workspace, this.cornerCases),
         new DashboardRenderTool(this.workspace),
@@ -239,7 +311,11 @@ export class AgentEngine {
       // Distillation+ only (DISTILL mode)
       distill: [
         workerLlm,
-        new WorkflowRunTool(this.workspace, this.versionManager, this.confidence),
+        new WorkflowRunTool(this.workspace, this.versionManager, this.confidence, {
+          // v0.6.1 A6: hook engine-emitted milestones so phase gates see workflow runs
+          recordMilestone: (phase, key, value) => this._recordMilestone(phase, key, value),
+          getCurrentPhase: () => this.currentPhase,
+        }),
         new TierDowngradeTool(this.workspace, workerLlm),
         new QCSampleTool(this.workspace),
       ],
@@ -313,7 +389,7 @@ export class AgentEngine {
   getContextStats() {
     const systemPrompt = this.context.build({
       agentMd: this._readAgentMd(),
-      skillIndex: this._skillLoader.formatForContext(),
+      skillIndex: this._skillLoader.formatForContext(this.currentPhase),
       pipelineState: this.pipelines[this.currentPhase]?.describeState?.() || null,
       workspaceState: this._buildWorkspaceState(),
     });
@@ -353,21 +429,37 @@ export class AgentEngine {
 
     // Heap-pressure diagnostic. The TUI has its own virtualization + tool-
     // output truncation (Bug 3 fixes), so Ink itself should never OOM. If we
-    // still see high heap usage, something else is leaking — log it once per
-    // pressure-crossing so operators can investigate without flooding logs.
+    // still see high heap usage, something else is leaking.
+    //
+    // A9: Original design logged once per pressure-crossing (edge-triggered),
+    // which went silent for 17h during E2E #3 while RSS climbed to 3.8GB.
+    // Now: still edge-trigger on entry (noisy otherwise), but ALSO re-emit
+    // every 15min while we're still above the threshold, so an operator
+    // watching logs after hour 4 still sees the signal. Drops to silent on
+    // recovery below 0.60.
     try {
       const mem = process.memoryUsage();
       const frac = mem.heapUsed / (mem.heapTotal || 1);
-      if (frac > 0.80 && !this._memPressureLogged) {
-        this._memPressureLogged = true;
-        this.eventLog.append("memory_pressure", {
-          heapUsedMB: Math.round(mem.heapUsed / 1024 / 1024),
-          heapTotalMB: Math.round(mem.heapTotal / 1024 / 1024),
-          rssMB: Math.round(mem.rss / 1024 / 1024),
-          historyLength: this.history.messages.length,
-        });
+      const now = Date.now();
+      const REPRESS_INTERVAL_MS = 15 * 60 * 1000;
+      if (frac > 0.80) {
+        const firstCrossing = !this._memPressureLogged;
+        const dueForRepress = this._memPressureLastEmittedAt &&
+          (now - this._memPressureLastEmittedAt) >= REPRESS_INTERVAL_MS;
+        if (firstCrossing || dueForRepress) {
+          this._memPressureLogged = true;
+          this._memPressureLastEmittedAt = now;
+          this.eventLog.append("memory_pressure", {
+            heapUsedMB: Math.round(mem.heapUsed / 1024 / 1024),
+            heapTotalMB: Math.round(mem.heapTotal / 1024 / 1024),
+            rssMB: Math.round(mem.rss / 1024 / 1024),
+            historyLength: this.history.messages.length,
+            kind: firstCrossing ? "crossing" : "sustained",
+          });
+        }
       } else if (frac < 0.60 && this._memPressureLogged) {
         this._memPressureLogged = false;  // re-arm for next crossing
+        this._memPressureLastEmittedAt = null;
       }
     } catch { /* process.memoryUsage failures are non-fatal */ }
   }
@@ -701,7 +793,7 @@ export class AgentEngine {
 
     const systemPrompt = this.context.build({
       agentMd: this._readAgentMd(),
-      skillIndex: this._skillLoader.formatForContext(),
+      skillIndex: this._skillLoader.formatForContext(this.currentPhase),
       pipelineState,
       workspaceState: this._buildWorkspaceState(),
     });
@@ -781,6 +873,30 @@ export class AgentEngine {
         });
 
         if (toolCallsAcc.size === 0) {
+          // A3: Empty-response guard. If the LLM returned no content AND no
+          // tool calls, count it. Two in a row almost always means the
+          // provider is silently failing (context exceeded, rate-limit
+          // corruption, auth expired) and continuing wastes tokens + time.
+          // Reset on any non-empty turn. Reason-tagged so /status can
+          // surface the running rate.
+          if (!collectedText || !collectedText.trim()) {
+            this._consecutiveEmptyResponses = (this._consecutiveEmptyResponses || 0) + 1;
+            this._totalEmptyResponses = (this._totalEmptyResponses || 0) + 1;
+            if (this._consecutiveEmptyResponses >= 2) {
+              const message =
+                `LLM returned empty response ${this._consecutiveEmptyResponses}× in a row — ` +
+                `likely context-length exceeded or provider-side silent failure. ` +
+                `Stopping this turn to prevent runaway API spend.`;
+              this.eventLog.append("error", { message, kind: "empty_response_streak" });
+              yield new AgentEvent({ type: "error", message });
+              this._consecutiveEmptyResponses = 0; // reset so next /run isn't blocked
+              return;
+            }
+          } else {
+            this._consecutiveEmptyResponses = 0;
+          }
+          this._totalTurns = (this._totalTurns || 0) + 1;
+
           // Bug 4 trigger (1): re-check phase criteria at end of every turn —
           // KC may have advanced state via conversation alone, without any
           // tool that the pipeline narrowly watches.
@@ -793,6 +909,10 @@ export class AgentEngine {
           return;
         }
 
+        // A3: A turn with tool_calls or content is not empty — reset streak.
+        this._consecutiveEmptyResponses = 0;
+        this._totalTurns = (this._totalTurns || 0) + 1;
+
         // Tool execution loop
         for (const tc of toolCallsAcc.values()) {
           let inputData = {};
@@ -803,6 +923,12 @@ export class AgentEngine {
           this.eventLog.append("tool_start", { name: tc.name, input: inputData });
           yield new AgentEvent({ type: "tool_start", name: tc.name, input: inputData });
 
+          // A1: Capture phase BEFORE tool execution. Some tools — notably
+          // phase_advance — mutate this.currentPhase via a callback without
+          // yielding any AgentEvent, so the TUI's status bar never gets the
+          // signal. We diff after execute() and emit a synthetic
+          // pipeline_event so subscribers can sync.
+          const beforePhase = this.currentPhase;
           const result = await this.toolRegistry.execute(tc.name, inputData);
 
           // Tool-call offloading: large outputs go to logs/tool_results/<traceId>.txt;
@@ -817,6 +943,29 @@ export class AgentEngine {
             isError: result.isError,
             traceId: offload?.traceId || null,
           });
+
+          // D3a: trace skill invocations. When the agent reads a SKILL.md via
+          // workspace_file (the canonical way KC "uses" a skill, since skills
+          // are progressively-disclosed markdown), emit a skill_invoked event.
+          // Makes "which skills did KC actually consult?" answerable in post-run
+          // analysis — before this, skills were opaque to the event log.
+          try {
+            if (
+              !result.isError &&
+              (tc.name === "workspace_file" || tc.name === "sandbox_exec")
+            ) {
+              const p = String(inputData?.path || inputData?.command || "");
+              const skillMatch = p.match(/(?:template\/)?skills\/[a-z-]+\/(?:meta-meta|meta|skill-creator)\/([a-zA-Z0-9_-]+)(?:\/SKILL\.md|\/)?|\bSKILL\.md\b/);
+              if (skillMatch) {
+                const skillName = skillMatch[1] || "(unknown)";
+                this.eventLog.append("skill_invoked", {
+                  skill: skillName,
+                  via_tool: tc.name,
+                  phase: this.currentPhase,
+                });
+              }
+            }
+          } catch { /* never let tracing break a tool call */ }
           yield new AgentEvent({
             type: "tool_result",
             name: tc.name,
@@ -837,6 +986,22 @@ export class AgentEngine {
           // user saw "CTX: 210% / stream terminated" with no recovery.
           this._maybeWindowAfterToolResult();
 
+          // A1: If the tool mutated the phase (e.g. phase_advance), emit the
+          // signal the TUI and pipelines need to re-sync state. Runs BEFORE
+          // pipeline.onToolResult so the fresh phase is active if the pipeline
+          // itself wants to react to the transition.
+          if (this.currentPhase !== beforePhase) {
+            yield new AgentEvent({
+              type: "pipeline_event",
+              data: {
+                type: "phase_changed",
+                from: beforePhase,
+                nextPhase: this.currentPhase,
+                reason: `via ${tc.name}`,
+              },
+            });
+          }
+
           // Pipeline controller: update state and re-register tools on phase change
           if (pipeline?.onToolResult) {
             const pEvent = pipeline.onToolResult(tc.name, inputData, result);
@@ -857,8 +1022,15 @@ export class AgentEngine {
         if (ev) yield ev;
 
       } catch (err) {
-        this.eventLog.append("error", { message: err.message });
-        yield new AgentEvent({ type: "error", message: err.message });
+        // A8: If the LLM client tagged the stream termination reason, pass
+        // it through. Upstream log consumers + the TUI can then distinguish
+        // "provider returned 429" from "socket died mid-token" from "SSE
+        // buffer exploded" — today they're all just "Error: ...".
+        const payload = { message: err.message };
+        if (err.streamTermination) payload.kind = err.streamTermination;
+        if (err.status) payload.status = err.status;
+        this.eventLog.append("error", payload);
+        yield new AgentEvent({ type: "error", message: err.message, ...payload });
         return;
       }
     }
@@ -889,22 +1061,219 @@ export class AgentEngine {
       return false;
     }
 
-    const phaseSummary = `[${this.currentPhase.toUpperCase()} → ${nextPhase.toUpperCase()}]: ${reason}${force && nextPhase !== expected ? " (forced)" : ""}`;
+    // v0.6.1 B1: build engine-appended hard-counts block + heuristic mismatch
+    // detection so the LLM-narrated reason can be cross-checked against
+    // ground-truth telemetry. Phase summaries become diagnostic, not just
+    // narrative.
+    const engineCounts = this._buildEngineCountsBlock(this.currentPhase);
+    const mismatchPrefix = this._detectSummaryMismatch(reason, this.currentPhase) ? "⚠️ POSSIBLE MISMATCH: " : "";
+    const phaseSummary =
+      `[${this.currentPhase.toUpperCase()} → ${nextPhase.toUpperCase()}]: ${mismatchPrefix}${reason}` +
+      (force && nextPhase !== expected ? " (forced)" : "") +
+      (engineCounts ? `\n  (engine) ${engineCounts}` : "");
     this._phaseSummaries.push(phaseSummary);
     this.eventLog.append("phase_transition", {
       from: this.currentPhase,
       to: nextPhase,
       reason,
+      engineCounts: engineCounts || null,
+      possibleMismatch: !!mismatchPrefix,
       forced: force && nextPhase !== expected,
     });
+    const fromPhase = this.currentPhase;
     this.currentPhase = nextPhase;
     this._registerToolsForPhase(this.currentPhase);
     this.workspace.setPhase(this.currentPhase);
     this._createTasksForPhase(this.currentPhase);
     this.saveState();
+
+    // B8: Soft signal — surface any sub-agents left running from the prior
+    // phase so the main agent's next turn can decide whether to kill them.
+    // NOT automated: phase_advance can fire from _maybeAutoAdvance on a
+    // criteria-flip, and auto-killing would couple lifecycle with blast
+    // radius. This just informs.
+    try {
+      const agentTool = this._buildTools?.core?.find((t) => t?.name === "agent_tool");
+      const runningIds = agentTool?.getRunningTaskIds?.() || [];
+      if (runningIds.length > 0) {
+        this.eventLog.append("stale_subagents", {
+          from_phase: fromPhase,
+          to_phase: nextPhase,
+          running_task_ids: runningIds,
+          hint: "These sub-agents were dispatched during the prior phase. Consider operation=poll to check status, or operation=kill to abort if stale.",
+        });
+      }
+    } catch { /* never let signal emission break phase advance */ }
+
     return true;
   }
 
+  /**
+   * v0.6.1 A6: Single chokepoint for engine-emitted milestone updates.
+   * Tools call this on successful execution to bump pipeline counters that
+   * the phase-gate hardening (A2-A5) depends on. Without engine emission,
+   * gates fall back to filesystem scans which can miss work that didn't
+   * follow canonical output paths (E2E #4: `unified_qc.py` wrote to
+   * `output/results/`, production-qc only scanned `output/qc/`).
+   *
+   * The mutation routes through the pipeline's existing internal state, so
+   * exportState/importState round-trips work unchanged and the gate sees a
+   * unified view of (filesystem-scanned + engine-emitted) signals.
+   *
+   * Three modes inferred from value shape:
+   *  - increment counter: pipeline[key] is number, value is number → add
+   *  - set in dict-by-id:  pipeline[key] is object, value is { id, value? } → assign
+   *  - dedupe-add to array: pipeline[key] is array, value is string → push if absent
+   *
+   * @param {string} phase - Pipeline name (e.g., "distillation")
+   * @param {string} key - Field on the pipeline (e.g., "workflowsTested")
+   * @param {*} value - Shape varies by target type (see modes above)
+   * @returns {boolean} true if a write happened
+   */
+  _recordMilestone(phase, key, value) {
+    const pipeline = this.pipelines?.[phase];
+    if (!pipeline) return false;
+    const target = pipeline[key];
+    // increment counter
+    if (typeof target === "number" && typeof value === "number") {
+      pipeline[key] = target + value;
+      return true;
+    }
+    // set on dict-by-id
+    if (target && typeof target === "object" && !Array.isArray(target)
+        && value && typeof value === "object" && "id" in value) {
+      target[value.id] = "value" in value ? value.value : true;
+      return true;
+    }
+    // dedupe-add to array
+    if (Array.isArray(target) && typeof value === "string") {
+      if (!target.includes(value)) target.push(value);
+      return true;
+    }
+    return false;
+  }
+
+  /**
+   * v0.6.1 B1: build a one-line "engine counts" block summarizing the
+   * pipeline's ground-truth telemetry at the moment of phase advance.
+   * Different phases surface different metrics; we keep this short so the
+   * appended summary line stays readable.
+   *
+   * @param {string} fromPhase - The phase being LEFT (we summarize its work)
+   * @returns {string} block text, or "" if pipeline has nothing to report
+   */
+  _buildEngineCountsBlock(fromPhase) {
+    const pipeline = this.pipelines?.[fromPhase];
+    if (!pipeline) return "";
+    const parts = [];
+    try {
+      switch (fromPhase) {
+        case "extraction": {
+          const total = pipeline._catalogRuleCount?.() ?? pipeline.rulesExtracted?.length ?? 0;
+          parts.push(`rulesExtracted: ${pipeline.rulesExtracted?.length ?? 0}`);
+          parts.push(`rulesWithChunkRefs: ${pipeline.rulesWithChunkRefs?.length ?? 0}/${total}`);
+          parts.push(`rulesWithTests: ${pipeline.rulesWithTests?.length ?? 0}`);
+          parts.push(`coverageAudited: ${pipeline.coverageAudited ? "yes" : "no"}`);
+          break;
+        }
+        case "skill_authoring": {
+          const totalRules = pipeline.totalRules?.length ?? 0;
+          const covered = pipeline.ruleIdsCovered?.size ?? 0;
+          parts.push(`rulesCovered: ${covered}/${totalRules}`);
+          parts.push(`skillDirsAuthored: ${pipeline.skillsAuthored?.length ?? 0}`);
+          if (this.taskManager) {
+            const t = this.taskManager.countByPhase("skill_authoring");
+            const d = this.taskManager.countByPhase("skill_authoring", "completed");
+            const f = this.taskManager.countByPhase("skill_authoring", "failed");
+            parts.push(`tasksCompleted: ${d}/${t}${f > 0 ? ` (+${f} failed)` : ""}`);
+          }
+          break;
+        }
+        case "skill_testing": {
+          const total = pipeline.skillsToTest?.length ?? 0;
+          const tested = Object.keys(pipeline.skillsTested || {}).length;
+          const passing = pipeline.skillsPassing?.length ?? 0;
+          parts.push(`skillsTested: ${tested}/${total}`);
+          parts.push(`skillsPassing: ${passing}`);
+          parts.push(`iterations: ${pipeline.iterationCount ?? 0}`);
+          break;
+        }
+        case "distillation": {
+          const total = pipeline.skillsToDistill?.length ?? 0;
+          const created = Object.keys(pipeline.workflowsCreated || {}).length;
+          const tested = Object.keys(pipeline.workflowsTested || {}).length;
+          const passing = pipeline.workflowsPassing?.length ?? 0;
+          parts.push(`workflowsCreated: ${created}/${total}`);
+          parts.push(`workflowsTested: ${tested}/${total}`);
+          parts.push(`workflowsPassing: ${passing}/${total}`);
+          break;
+        }
+        case "production_qc": {
+          parts.push(`batchesProcessed: ${pipeline.batchesProcessed ?? 0}`);
+          parts.push(`documentsReviewed: ${pipeline.documentsReviewed ?? 0}`);
+          parts.push(`monitoring: ${pipeline.monitoringPhase ?? "?"}`);
+          break;
+        }
+        // bootstrap / finalization: no specific counters, fall through
+      }
+    } catch { /* never let summary build break phase advance */ }
+    return parts.join(", ");
+  }
+
+  /**
+   * v0.6.1 B1: heuristic mismatch detection. Conservative regex over the
+   * LLM's free-form reason for percentages and counts, compared against
+   * engine truth. INFORMATIONAL only — never blocks the transition. False
+   * positives are acceptable (the warning is a hint to the human reviewer,
+   * not a hard signal). False negatives are also acceptable (this catches
+   * the loud, numerical claims; subtle ones still slip through).
+   *
+   * Returns true if the agent's reason mentions a count or percentage that
+   * doesn't match engine state.
+   */
+  _detectSummaryMismatch(reason, fromPhase) {
+    if (!reason || typeof reason !== "string") return false;
+    const pipeline = this.pipelines?.[fromPhase];
+    if (!pipeline) return false;
+    try {
+      // Match "N/M" fractions and standalone counts
+      const fractionMatches = [...reason.matchAll(/(\d+)\s*\/\s*(\d+)/g)];
+      // Match "N rules / skills / workflows / tasks"
+      const countMatches = [...reason.matchAll(/(\d+)\s*(rules?|skills?|workflows?|tasks?|条规则|个技能)/gi)];
+      // Match accuracy claims like "95%", "0.95"
+      const pctMatches = [...reason.matchAll(/(\d+(?:\.\d+)?)\s*%/g)];
+
+      // Phase-specific cross-checks (cheap conservative comparisons)
+      if (fromPhase === "skill_authoring" && this.taskManager) {
+        const completed = this.taskManager.countByPhase("skill_authoring", "completed");
+        const total = this.taskManager.countByPhase("skill_authoring");
+        for (const m of fractionMatches) {
+          const claimedDone = parseInt(m[1], 10);
+          const claimedTotal = parseInt(m[2], 10);
+          if (claimedTotal === total && claimedDone > completed + 5) return true;
+        }
+      }
+      if (fromPhase === "skill_testing") {
+        const tested = Object.keys(pipeline.skillsTested || {}).length;
+        const passing = pipeline.skillsPassing?.length ?? 0;
+        for (const m of pctMatches) {
+          const claimed = parseFloat(m[1]);
+          // If claimed > 50% but engine sees 0 tested, that's suspicious
+          if (claimed >= 50 && tested === 0 && passing === 0) return true;
+        }
+      }
+      if (fromPhase === "production_qc") {
+        const batches = pipeline.batchesProcessed ?? 0;
+        // Any "complete" or large-count claim while batches==0 is suspicious
+        if (batches === 0) {
+          if (countMatches.some((m) => parseInt(m[1], 10) > 10)) return true;
+          if (pctMatches.some((m) => parseFloat(m[1]) > 50)) return true;
+        }
+      }
+    } catch { /* informational only — never block */ }
+    return false;
+  }
+
   /**
    * Bug 4 trigger (1) auto-detect, edge-triggered (Bug 5): only fires on a
    * fresh false → true flip in `exitCriteriaMet()`. Sessions resumed in an
@@ -972,6 +1341,16 @@ export class AgentEngine {
   /**
    * Create per-rule tasks when entering a new phase.
    * Reads the rule catalog and creates one task per rule for the given phase.
+   *
+   * D6: For skill_authoring / skill_testing, filter rules via the bundle
+   * classification cache (`cache/bundles/<hash>.classification.json`,
+   * written by document_classify). Rules whose `applicable_product_types`
+   * or `report_types` don't overlap with the bundle's classification get
+   * SKIPPED at task-creation time — we don't mutate catalog.json to mark
+   * them not_applicable, we just keep them out of the task queue. The
+   * finalization phase (Group E) will report them in the coverage
+   * artifact as "not applicable to this bundle." Conservative default:
+   * if no classification exists, include all rules (pre-B9 behavior).
    */
   _createTasksForPhase(phase) {
     if (!this.taskManager) return; // Sub-agents don't manage tasks
@@ -980,28 +1359,258 @@ export class AgentEngine {
 
     try {
       const catalog = JSON.parse(fs.readFileSync(catalogPath, "utf-8"));
-      const rules = normalizeRuleCatalog(catalog);
-      if (rules.length > 0) {
-        this.taskManager.createRuleTasks(rules, phase);
+      let rules = normalizeRuleCatalog(catalog);
+      if (rules.length === 0) return;
+
+      // D6: applicability pre-filter (skill phases only — bootstrap/extraction
+      // have no task creation here per A6).
+      if (phase === "skill_authoring" || phase === "skill_testing") {
+        const classification = this._loadBundleClassification();
+        if (classification) {
+          const before = rules.length;
+          rules = rules.filter((r) => this._ruleAppliesToBundle(r, classification));
+          if (rules.length < before) {
+            this.eventLog.append("applicability_prefilter", {
+              phase,
+              classification: {
+                product_type: classification.product_type,
+                report_type: classification.report_type,
+                source: classification.source,
+              },
+              rules_before: before,
+              rules_after: rules.length,
+              skipped: before - rules.length,
+            });
+          }
+        }
       }
+      this.taskManager.createRuleTasks(rules, phase);
     } catch { /* skip if catalog can't be read */ }
   }
 
+  /**
+   * D6: Load the most recent bundle classification cache, if one exists.
+   * Written by the `document_classify` tool. Returns null if no cache or
+   * unreadable — callers must treat null as "all rules apply."
+   */
+  _loadBundleClassification() {
+    const cacheDir = path.join(this.workspace.cwd, "cache", "bundles");
+    if (!fs.existsSync(cacheDir)) return null;
+    let entries;
+    try { entries = fs.readdirSync(cacheDir); }
+    catch { return null; }
+    const files = entries
+      .filter((n) => n.endsWith(".classification.json"))
+      .map((n) => {
+        const p = path.join(cacheDir, n);
+        try { return { path: p, mtime: fs.statSync(p).mtimeMs }; }
+        catch { return null; }
+      })
+      .filter(Boolean)
+      .sort((a, b) => b.mtime - a.mtime);
+    if (files.length === 0) return null;
+    try { return JSON.parse(fs.readFileSync(files[0].path, "utf-8")); }
+    catch { return null; }
+  }
+
+  /**
+   * D6: Rule-applicability check mirroring the AMC app's `applies_to`.
+   * Conservative: returns true when we don't have enough info to
+   * confidently skip (missing fields on rule, or classification with
+   * empty product/report).
+   */
+  _ruleAppliesToBundle(rule, classification) {
+    const docProduct = classification?.product_type || "";
+    const docReport = classification?.report_type || "";
+    const ruleProducts = rule.applicable_product_types || rule.applicable_sections || [];
+    const ruleReports = rule.report_types || [];
+
+    const allProducts = ruleProducts.length === 0 ||
+      ruleProducts.some((x) => x === "全部" || x === "all" || x === "");
+    const allReports = ruleReports.length === 0 ||
+      ruleReports.some((x) => x === "全部" || x === "all" || x === "");
+    if (allProducts && allReports) return true;
+
+    const productOk = allProducts || (
+      docProduct && ruleProducts.some((rp) => rp.includes(docProduct) || docProduct.includes(rp))
+    );
+    const reportOk = allReports || (
+      docReport && ruleReports.some((rr) => rr.includes(docReport) || docReport.includes(rr))
+    );
+
+    // Unknown classification → don't prefilter, let the agent judge.
+    if (!docProduct && !docReport) return true;
+    return productOk && reportOk;
+  }
+
+  /**
+   * D1: Enrich a skill_authoring / skill_testing task prompt with the
+   * rule's source context — reads `source_chunk_ids` back-refs from
+   * catalog.json (populated by extraction) and fetches chunk text from
+   * the most recent BundleTree cache. Falls back to the minimal prompt
+   * when catalog / cache aren't available.
+   *
+   * Previously the task prompt was ONE line — "Continue with next task:
+   * ${title}" — leaving the skill-author agent to re-read the rule and
+   * re-find its evidence per task. Auto-attach saves the LLM turn
+   * needed for document_search on every task, and ensures the author
+   * sees the exact regulation text the extractor used to justify the
+   * rule.
+   *
+   * @param {{id: string, title: string, ruleId?: string, phase: string}} task
+   * @returns {string}
+   */
+  _buildEnrichedTaskPrompt(task) {
+    const fallback = `Continue with next task: ${task.title}` +
+      (task.ruleId ? ` (rule: ${task.ruleId})` : "");
+
+    // Only enrich for rule-anchored phases
+    if (task.phase !== "skill_authoring" && task.phase !== "skill_testing") {
+      return fallback;
+    }
+    if (!task.ruleId) return fallback;
+
+    // Find the rule in catalog.json
+    const catalogPath = path.join(this.workspace.cwd, "rules", "catalog.json");
+    if (!fs.existsSync(catalogPath)) return fallback;
+    let rules;
+    try {
+      rules = normalizeRuleCatalog(JSON.parse(fs.readFileSync(catalogPath, "utf-8")));
+    } catch { return fallback; }
+    const rule = rules.find((r) => r.id === task.ruleId);
+    if (!rule) return fallback;
+
+    // Assemble the enriched brief. Every section is optional — when a
+    // back-ref or cache is missing, just skip that section rather than
+    // failing back to the minimal prompt.
+    const lines = [];
+    lines.push(`# Task: ${task.title}`);
+    lines.push("");
+    lines.push(`## Rule ${rule.id}`);
+    if (rule.source_ref) lines.push(`Source: ${rule.source_ref}`);
+    if (rule.severity) lines.push(`Severity: ${rule.severity}`);
+    if (rule.description) lines.push(`\n${rule.description}`);
+    if (rule.falsifiability_statement) lines.push(`\n**Falsifiability**: ${rule.falsifiability_statement}`);
+    if (rule.test_case_stub) lines.push(`**Test stub**: ${rule.test_case_stub}`);
+
+    // D1: if rule has source_chunk_ids AND a BundleTree cache exists,
+    // pull chunk text inline so the author doesn't need to call
+    // bundle_search manually. Bounded to ~3000 tokens total to avoid
+    // blowing the author's context budget.
+    const chunkIds = Array.isArray(rule.source_chunk_ids) ? rule.source_chunk_ids : [];
+    if (chunkIds.length > 0) {
+      const chunks = this._loadChunksFromBundleCache(chunkIds);
+      if (chunks.length > 0) {
+        lines.push("");
+        lines.push("## Source context");
+        let totalChars = 0;
+        const MAX_CHARS = 7500; // ~3000 CJK tokens
+        for (const ch of chunks) {
+          const header = `### ${ch.title || ch.chunk_id} · ${ch.source_file} p.${(ch.page_range || [1, 1]).join("-")}`;
+          const body = (ch.content || "").trim();
+          const block = `${header}\n${body}\n`;
+          if (totalChars + block.length > MAX_CHARS) {
+            lines.push(`\n[…${chunks.length - chunks.indexOf(ch)} more source chunks truncated; use bundle_search to retrieve them…]`);
+            break;
+          }
+          lines.push("");
+          lines.push(block);
+          totalChars += block.length;
+        }
+      }
+    }
+
+    // Sibling rules (same source_ref prefix) — helps the author see the
+    // surrounding catalog and avoid re-implementing cross-referenced logic.
+    const siblings = this._findSiblingRuleIds(rule, rules);
+    if (siblings.length > 0) {
+      lines.push("");
+      lines.push(`## Sibling rules (same regulation section)`);
+      lines.push(siblings.map((id) => `- ${id}`).join("\n"));
+    }
+
+    lines.push("");
+    lines.push("Write the skill to `rule_skills/<rule_id>/SKILL.md` + detect script. Prefer 1 rule = 1 skill dir (use `check_rNNN_rMMM.py` naming ONLY when rules share evidence and fail together).");
+
+    return lines.join("\n");
+  }
+
+  /** D1: Load chunk text from the most recent BundleTree cache. */
+  _loadChunksFromBundleCache(chunkIds) {
+    const cacheDir = path.join(this.workspace.cwd, "cache", "bundles");
+    if (!fs.existsSync(cacheDir)) return [];
+    let entries;
+    try { entries = fs.readdirSync(cacheDir); }
+    catch { return []; }
+    const candidates = entries
+      .filter((n) => n.endsWith(".json") && !n.endsWith(".classification.json"))
+      .map((n) => {
+        const p = path.join(cacheDir, n);
+        try { return { path: p, mtime: fs.statSync(p).mtimeMs }; }
+        catch { return null; }
+      })
+      .filter(Boolean)
+      .sort((a, b) => b.mtime - a.mtime);
+    if (candidates.length === 0) return [];
+    let tree;
+    try { tree = JSON.parse(fs.readFileSync(candidates[0].path, "utf-8")); }
+    catch { return []; }
+    const out = [];
+    for (const cid of chunkIds) {
+      const ch = tree.chunks?.[cid];
+      if (ch) out.push(ch);
+    }
+    return out;
+  }
+
+  /** D1: Rules that share the same regulation article (naive: source_ref prefix). */
+  _findSiblingRuleIds(rule, allRules) {
+    if (!rule.source_ref) return [];
+    const prefix = rule.source_ref.split(/[第条款项]/)[0].trim();
+    if (!prefix) return [];
+    return allRules
+      .filter((r) => r.id !== rule.id && (r.source_ref || "").startsWith(prefix))
+      .slice(0, 8)
+      .map((r) => r.id);
+  }
+
   /**
    * Ralph-loop: run a turn, then auto-continue through pending tasks.
    * Compacts context aggressively between tasks to prevent context blowup.
    * If no tasks exist, behaves identically to runTurn().
    *
    * @param {string} userMessage
+   * @param {{parallelism?: number}} [opts] — B1: optional parallel mode.
+   *   N > 1 dispatches tasks through N concurrent subagents (using the
+   *   agent_tool infrastructure from B8). Clamped to `effectiveParallelism`
+   *   from config.js — which silently downgrades to 1 unless
+   *   KC_PARALLELISM_VERIFIED=1 is set AND heap.jsonl shows flat RSS
+   *   (B0.6 guard; prevents accidental $100+ runaway runs).
    * @yields {AgentEvent}
    */
-  async *runTaskLoop(userMessage) {
+  async *runTaskLoop(userMessage, opts = {}) {
     // Sub-agents don't run task loops — they execute one task and exit
     if (!this.taskManager) {
       yield* this.runTurn(userMessage);
       return;
     }
 
+    // B1: resolve effective parallelism. Caller opts override config.
+    const requested = Number.isFinite(opts.parallelism)
+      ? Math.max(1, Math.min(8, opts.parallelism))
+      : (this.config.effectiveParallelism?.() ?? 1);
+
+    if (requested > 1) {
+      yield* this._runTaskLoopParallel(userMessage, requested);
+      return;
+    }
+
+    yield* this._runTaskLoopSerial(userMessage);
+  }
+
+  /** B1: original serial ralph-loop path — one task at a time, shared
+   *  conversation history. Unchanged from pre-v0.6.0 behavior. */
+  async *_runTaskLoopSerial(userMessage) {
     // Run the initial turn (user's request)
     yield* this.runTurn(userMessage);
 
@@ -1015,8 +1624,11 @@ export class AgentEngine {
         await this.compact({ recentCount: 8 });
       }
 
-      const task = this.taskManager.getNextPending();
-      this.taskManager.updateTask(task.id, { status: "in_progress" });
+      // B2: atomic claim — for serial we could use getNextPending, but
+      // using claimNextPending gives us consistent state fields (worker
+      // label, startedAt) whether in serial or parallel mode.
+      const task = this.taskManager.claimNextPending("serial");
+      if (!task) break;
 
       // Yield task progress event for TUI
       yield new AgentEvent({
@@ -1030,14 +1642,15 @@ export class AgentEngine {
         },
       });
 
-      // Synthesize a task-focused prompt
-      const taskPrompt = `Continue with next task: ${task.title}` +
-        (task.ruleId ? ` (rule: ${task.ruleId})` : "");
+      // D1: synthesize a task-focused prompt, enriched with rule source
+      // context (rule NL + source_ref + chunk text + sibling ids) when
+      // the catalog + BundleTree cache are available. Falls back to the
+      // minimal "Continue with next task" line otherwise.
+      const taskPrompt = this._buildEnrichedTaskPrompt(task);
 
       yield* this.runTurn(taskPrompt);
 
-      this.taskManager.updateTask(task.id, { status: "completed" });
-      this.taskManager.save();
+      this.taskManager.markDone(task.id);
       this.saveState();
 
       yield new AgentEvent({
@@ -1074,6 +1687,190 @@ export class AgentEngine {
     }
   }
 
+  /**
+   * B1: Parallel ralph-loop — N concurrent subagents each executing one
+   * task at a time, claimed atomically from TaskManager.
+   *
+   * Implementation: leverages B8's agent_tool infrastructure. Each worker
+   * slot is a sub-engine with its own heap-isolated history; workspace
+   * writes are serialized through B9's file locks. The main engine acts
+   * as dispatcher — it claims tasks and spawns subagents, then waits.
+   *
+   * Chosen over in-process history-forking because: (a) sub-engines are
+   * already heap-isolated (good under B0's RSS-safety regime); (b)
+   * kill authority from B8 applies uniformly; (c) no runTurn refactor
+   * needed — the engine's conversation-state assumptions stay intact.
+   * Trade-off: each task pays a cold-start cost (re-read AGENT.md,
+   * skill index, pipeline state). For 100+ task sessions this is
+   * amortized against the 2-4× wall-clock speedup.
+   */
+  async *_runTaskLoopParallel(userMessage, parallelism) {
+    // Initial turn: main agent reads user request, creates tasks.
+    yield* this.runTurn(userMessage);
+
+    const agentTool = this._buildTools.core.find((t) => t?.name === "agent_tool");
+    if (!agentTool) {
+      // Shouldn't happen (agent_tool is core), but fall back safely.
+      yield new AgentEvent({
+        type: "error",
+        message: "agent_tool not registered; parallel mode requires it. Falling back to serial.",
+      });
+      yield* this._runTaskLoopSerial("");
+      return;
+    }
+
+    // Event queue so concurrent workers can yield progress through a
+    // single async-generator consumer. push-style with a notifier.
+    const eventQueue = [];
+    let notify = null;
+    const enq = (ev) => {
+      eventQueue.push(ev);
+      if (notify) { const n = notify; notify = null; n(); }
+    };
+
+    // In-flight: subagent task_id → { task, promise }
+    const inFlight = new Map();
+
+    const dispatch = async () => {
+      while (inFlight.size < parallelism) {
+        const task = this.taskManager.claimNextPending(`pool${inFlight.size}`);
+        if (!task) return;
+
+        const workerLabel = `pool${[...inFlight.keys()].length}`;
+        const subId = `pool_${task.id}`.replace(/[^A-Za-z0-9_-]/g, "_").slice(0, 60);
+
+        // D1: build the enriched brief with source context. Parallel workers
+        // are subagents — each with zero conversation history, so the brief
+        // must carry everything they need. Even more important to have
+        // source context inline vs. expecting them to call document_search.
+        const enriched = this._buildEnrichedTaskPrompt(task);
+        const brief =
+          enriched +
+          `\n\nNOTE (parallel worker): write outputs via workspace_file or ` +
+          `rule_catalog — do NOT write to shared coordination files ` +
+          `(rules/catalog.json, rules/manifest.json) via sandbox_exec; they're ` +
+          `lock-protected and bypassing the lock will race with other workers.`;
+
+        enq(new AgentEvent({
+          type: "task_progress",
+          data: {
+            taskId: task.id, title: task.title, ruleId: task.ruleId,
+            status: "in_progress", worker: workerLabel,
+            progress: this.taskManager.progress,
+          },
+        }));
+
+        // Spawn via the tool's public API. agent_tool writes status.txt,
+        // abort controller, etc. We read _runningTasks to get a promise
+        // handle we can await.
+        const spawnRes = await agentTool.execute({
+          operation: "spawn",
+          task_description: brief,
+          task_id: subId,
+        });
+
+        if (spawnRes.isError) {
+          this.taskManager.markFailed(task.id, `spawn failed: ${spawnRes.content}`);
+          enq(new AgentEvent({
+            type: "task_progress",
+            data: { taskId: task.id, status: "failed", worker: workerLabel },
+          }));
+          continue;
+        }
+
+        const entry = agentTool._runningTasks.get(subId);
+        if (!entry) {
+          // Sub-agent completed synchronously (no events) — mark done.
+          this.taskManager.markDone(task.id);
+          enq(new AgentEvent({
+            type: "task_progress",
+            data: { taskId: task.id, status: "completed", worker: workerLabel },
+          }));
+          continue;
+        }
+
+        const trackedPromise = entry.promise.then(
+          () => ({ taskId: task.id, subId, ok: true }),
+          (e) => ({ taskId: task.id, subId, ok: false, error: e?.message || String(e) }),
+        );
+        inFlight.set(subId, { task, workerLabel, promise: trackedPromise });
+      }
+    };
+
+    // Prime the pool
+    await dispatch();
+
+    // Drain events + replenish until queue is empty and all in-flight done.
+    while (inFlight.size > 0 || eventQueue.length > 0) {
+      // Drain all queued events first
+      while (eventQueue.length > 0) yield eventQueue.shift();
+
+      if (inFlight.size === 0) break;
+
+      // Wait for either the next event OR a worker to complete
+      const workerCompletion = Promise.race([...inFlight.values()].map((v) => v.promise));
+      const eventArrival = new Promise((resolve) => { notify = () => resolve("event"); });
+      const winner = await Promise.race([
+        workerCompletion.then((done) => ({ kind: "worker", done })),
+        eventArrival.then(() => ({ kind: "event" })),
+      ]);
+
+      if (winner.kind === "worker") {
+        const { taskId, subId, ok, error } = winner.done;
+        const entry = inFlight.get(subId);
+        inFlight.delete(subId);
+
+        if (ok) {
+          this.taskManager.markDone(taskId);
+          enq(new AgentEvent({
+            type: "task_progress",
+            data: {
+              taskId, status: "completed",
+              worker: entry?.workerLabel,
+              progress: this.taskManager.progress,
+            },
+          }));
+        } else {
+          this.taskManager.markFailed(taskId, error);
+          enq(new AgentEvent({
+            type: "task_progress",
+            data: {
+              taskId, status: "failed",
+              worker: entry?.workerLabel,
+              error,
+              progress: this.taskManager.progress,
+            },
+          }));
+        }
+
+        // Refill the pool. If no pending tasks left, in-flight drains naturally.
+        await dispatch();
+      }
+      // event winner: loop re-iterates and drains eventQueue
+    }
+
+    this.saveState();
+
+    // After all workers done, check for phase auto-advance (same as serial path).
+    if (this._allCurrentPhaseTasksComplete()) {
+      const pipeline = this.pipelines[this.currentPhase];
+      let exitMet = false;
+      try { exitMet = !!pipeline?.exitCriteriaMet?.(); } catch { exitMet = false; }
+      if (exitMet) {
+        const next = NEXT_PHASE[this.currentPhase];
+        if (next) {
+          const advanced = this._advancePhase(next, "all parallel tasks completed + exit criteria met");
+          if (advanced) {
+            yield new AgentEvent({
+              type: "pipeline_event",
+              data: { type: "phase_ready", nextPhase: next, message: "all phase tasks done; exit criteria met" },
+            });
+          }
+        }
+      }
+    }
+  }
+
   /**
    * True when every task tagged with the current phase is in a terminal state
    * (completed | failed | skipped) and at least one such task exists. Used by