lynkr 9.9.1 → 9.10.0

package/src/routing/intent-score.js

@@ -347,9 +347,21 @@ function _reconcile(anchorScore, anchorClass, classifierResult) {
     // Classifier says LOWER tier — trust it. Fixes over-routing.
     return { score: TIER_MIDPOINT[classifierTier], reconciled: 'down' };
   }
-  // Classifier says HIGHER tier — gate on confidence.
+  // Classifier says HIGHER tier — gate on confidence, and cap the jump at
+  // ONE band above the anchor. The classifier is a tiebreaker, not an
+  // oracle: qwen2.5:3b's confidence is degenerate (92% of eval answers say
+  // 0.9-1.0, incl. every wrong one), so an unbounded jump let a 3B verdict
+  // catapult "Who kills him ?" from anchor 25 to 88 → REASONING →
+  // subscription passthrough (live incident 2026-07-21). Escalating past
+  // the adjacent band now takes consecutive turns that keep re-scoring
+  // higher, which is exactly the persistence a genuinely hard conversation
+  // exhibits.
   if (classifierResult.confidence >= 0.8) {
-    return { score: TIER_MIDPOINT[classifierTier], reconciled: 'up' };
+    const cappedIdx = Math.min(classifierIdx, anchorIdx + 1);
+    return {
+      score: TIER_MIDPOINT[TIER_ORDER[cappedIdx]],
+      reconciled: cappedIdx < classifierIdx ? 'up_capped' : 'up',
+    };
   }
   return { score: anchorScore, reconciled: 'up_gated' };
 }
@@ -385,6 +397,12 @@ async function scoreIntent(payload, opts = {}) {
             classifierResult = await classifyDifficulty(text, {
               forceMatched: opts.forceMatched,
               riskLevel: opts.riskLevel,
+              // Condensed prior turns, threaded by the router's window loop.
+              // Without it a short follow-up ("Who kills him ?") is
+              // unclassifiable in isolation.
+              context: typeof payload?._conversationContext === 'string'
+                ? payload._conversationContext
+                : null,
             });
           } catch (err) {
             logger.debug({ err: err.message }, '[IntentScore] classifier failed — anchor only');

package/src/routing/session-affinity.js

@@ -114,12 +114,18 @@ function getPin(sessionId) {
  * Accepts a routing decision plus session-scoped stats used by the re-pin
  * heuristics.
  *
+ * `hasToolHistory` is sticky-true: once set to true it stays true for the
+ * pin's lifetime — a follow-up refresh that didn't carry tool blocks (e.g.
+ * a plain-text turn between tool exchanges) must not clear the flag, or the
+ * Signal-2 side-channel check would flip open again.
+ *
  * @param {string} sessionId
  * @param {{provider:string, model?:string|null, tier?:string|null}} decision
- * @param {{messageCount?:number|null, promptTokensEst?:number|null}} [stats]
+ * @param {{messageCount?:number|null, promptTokensEst?:number|null, hasToolHistory?:boolean}} [stats]
  */
 function setPin(sessionId, decision, stats = {}) {
   if (!sessionId || !decision?.provider) return;
+  const prev = pins.get(sessionId);
   const pin = {
     provider: decision.provider,
     model: decision.model ?? null,
@@ -127,6 +133,7 @@ function setPin(sessionId, decision, stats = {}) {
     score: typeof decision.score === 'number' ? decision.score : null,
     messageCount: stats.messageCount ?? null,
     promptTokensEst: stats.promptTokensEst ?? null,
+    hasToolHistory: !!(stats.hasToolHistory || prev?.hasToolHistory),
     ts: Date.now(),
   };
   // Refresh insertion order so active sessions aren't evicted.

package/src/routing/side-channel-detector.js

@@ -0,0 +1,103 @@
+/**
+ * Side-Channel Detector — first-user-message fingerprint drift
+ *
+ * Complements the pin-based side-channel signals in router.js:
+ *   1. Message-count regression   → needs pin
+ *   2. Tool-history mismatch      → needs pin with hasToolHistory
+ *   3. First-message fingerprint  → THIS module — works before any pin exists
+ *
+ * Purpose: on a session's first request, cache a fingerprint of messages[0].
+ * On later requests within the TTL, if messages[0] no longer matches, the
+ * payload has been replayed by a background utility caller (Claude Code
+ * wraps the original user text in `<session>...</session>` for title-gen,
+ * substitutes recap prompts, etc.).
+ *
+ * Trade-off: if a background call fires BEFORE the first real user turn
+ * lands, its wrapper prompt becomes the anchor. That's rare in practice —
+ * observed order in real traffic is user turn first, background call after.
+ * If it does happen, the real user's next request will be mislabelled once,
+ * then the anchor will "settle" once ttl expires.
+ *
+ * @module routing/side-channel-detector
+ */
+
+const crypto = require('crypto');
+const logger = require('../logger');
+
+const TTL_MS = 30 * 60 * 1000;   // 30 min — matches an interactive Claude Code session
+const MAX_ENTRIES = 2000;         // LRU cap (same order of magnitude as session-affinity)
+
+/** @type {Map<string, {hash:string, ts:number}>} */
+const anchors = new Map();
+
+function _evictIfNeeded() {
+  if (anchors.size <= MAX_ENTRIES) return;
+  const oldest = anchors.keys().next().value;
+  if (oldest !== undefined) anchors.delete(oldest);
+}
+
+function _firstUserText(messages) {
+  if (!Array.isArray(messages)) return null;
+  for (const m of messages) {
+    if (m?.role !== 'user') continue;
+    if (typeof m.content === 'string') return m.content;
+    if (Array.isArray(m.content)) {
+      const t = m.content.filter(b => b?.type === 'text').map(b => b.text || '').join('');
+      if (t) return t;
+    }
+  }
+  return null;
+}
+
+function _hash(str) {
+  return crypto.createHash('sha1').update(str).digest('hex').slice(0, 16);
+}
+
+/**
+ * Check whether the payload's first user message drifts from the cached
+ * anchor for this session. Returns true only when we have an anchor AND
+ * it disagrees; returns false when no anchor exists yet (bootstrap case)
+ * and stores this payload's hash as the new anchor.
+ *
+ * @param {string|null} sessionId
+ * @param {Array} messages
+ * @returns {boolean} true if drift detected (caller should treat as side channel)
+ */
+function check(sessionId, messages) {
+  if (!sessionId) return false;
+  const text = _firstUserText(messages);
+  if (!text) return false;
+  const hash = _hash(text);
+  const now = Date.now();
+
+  const cached = anchors.get(sessionId);
+  if (cached) {
+    // Refresh insertion order for LRU
+    anchors.delete(sessionId);
+    if (now - cached.ts > TTL_MS) {
+      // Expired — reset the anchor, no drift claim.
+      anchors.set(sessionId, { hash, ts: now });
+      _evictIfNeeded();
+      return false;
+    }
+    anchors.set(sessionId, cached);
+    if (cached.hash !== hash) {
+      logger.debug({ sessionId, cachedHash: cached.hash, newHash: hash },
+        '[SideChannel] First-message fingerprint drift');
+      return true;
+    }
+    return false;
+  }
+
+  // Bootstrap — first sighting of this session, store and pass.
+  anchors.set(sessionId, { hash, ts: now });
+  _evictIfNeeded();
+  return false;
+}
+
+/** Test/maintenance helper — clear the in-memory anchors. */
+function _clear() {
+  anchors.clear();
+}
+
+module.exports = { check, _clear };

package/src/server.js

@@ -18,52 +18,13 @@ const { getCircuitBreakerRegistry } = require("./clients/circuit-breaker");
 const metrics = require("./metrics");
 const logger = require("./logger");
 const { initialiseMcp } = require("./mcp");
-const { registerStubTools } = require("./tools/stubs");
-const { registerWorkspaceTools } = require("./tools/workspace");
-const { registerExecutionTools } = require("./tools/execution");
-const { registerWebTools } = require("./tools/web");
-const { registerIndexerTools } = require("./tools/indexer");
-const { registerEditTools } = require("./tools/edits");
-const { registerGitTools } = require("./tools/git");
-const { registerTaskTools } = require("./tools/tasks");
-const { registerTestTools } = require("./tools/tests");
-const { registerMcpTools } = require("./tools/mcp");
-const { registerAgentTaskTool } = require("./tools/agent-task");
-const { registerDecomposeTool } = require("./tools/decompose");
 const { initConfigWatcher, getConfigWatcher } = require("./config/watcher");
 const { initializeHeadroom, shutdownHeadroom, getHeadroomManager } = require("./headroom");
 const { getWorkerPool, isWorkerPoolReady } = require("./workers/pool");
-const lazyLoader = require("./tools/lazy-loader");
-const { setLazyLoader } = require("./tools");
 const { waitForOllama } = require("./clients/ollama-startup");
 
 initialiseMcp();
 
-setLazyLoader(lazyLoader);
-
-const LAZY_TOOLS_ENABLED = process.env.LAZY_TOOLS_ENABLED !== "false";
-
-if (LAZY_TOOLS_ENABLED) {
-  // Only load core tools at startup (stubs, workspace, execution)
-  lazyLoader.loadCoreTools();
-  logger.info({ mode: "lazy" }, "Lazy tool loading enabled - other tools will load on demand");
-} else {
-  // Backwards compatibility: load all tools at startup
-  registerStubTools();
-  registerWorkspaceTools();
-  registerExecutionTools();
-  registerWebTools();
-  registerIndexerTools();
-  registerEditTools();
-  registerGitTools();
-  registerTaskTools();
-  registerTestTools();
-  registerMcpTools();
-  registerAgentTaskTool();
-  registerDecomposeTool();
-  logger.info({ mode: "eager" }, "All tools loaded at startup");
-}
-
 function createApp() {
   const app = express();
   const path = require('path');
@@ -171,13 +132,6 @@ function createApp() {
     res.json({ enabled: true, ...cache.getStats() });
   });
 
-  app.get("/metrics/lazy-tools", (req, res) => {
-    res.json({
-      enabled: LAZY_TOOLS_ENABLED,
-      ...lazyLoader.getLoaderStats(),
-    });
-  });
-
   app.use(router);
 
   app.use('/dashboard', require('./dashboard/router'));

package/src/agents/context-manager.js

@@ -1,236 +0,0 @@
-const fs = require("fs");
-const path = require("path");
-const logger = require("../logger");
-
-class ContextManager {
-  constructor() {
-    this.transcriptsDir = path.join(process.cwd(), "data", "agent-transcripts");
-    this.ensureTranscriptsDir();
-  }
-
-  ensureTranscriptsDir() {
-    if (!fs.existsSync(this.transcriptsDir)) {
-      fs.mkdirSync(this.transcriptsDir, { recursive: true });
-    }
-  }
-
-  /**
-   * Create fresh context for subagent
-   * @param {Object} agentDef - Agent definition
-   * @param {string} taskPrompt - Task from main agent
-   * @param {Object} mainContext - Minimal context from main agent
-   * @returns {Object} - Fresh context for subagent
-   */
-  createSubagentContext(agentDef, taskPrompt, mainContext = {}) {
-    const agentId = this.generateAgentId();
-    const transcriptPath = path.join(this.transcriptsDir, `agent-${agentId}.jsonl`);
-
-    // Initialize transcript file
-    fs.writeFileSync(transcriptPath, "");
-
-    // Build minimal context (NOT full main agent history)
-    const messages = [];
-
-    // System prompt from agent definition
-    messages.push({
-      role: "system",
-      content: agentDef.systemPrompt
-    });
-
-    // Optional: Add minimal context from main agent
-    if (mainContext.relevant_context) {
-      messages.push({
-        role: "system",
-        content: `Context from main agent:\n${mainContext.relevant_context}`
-      });
-    }
-
-    // Task prompt
-    messages.push({
-      role: "user",
-      content: taskPrompt
-    });
-
-    const context = {
-      agentId,
-      agentName: agentDef.name,
-      transcriptPath,
-      messages,
-      steps: 0,
-      maxSteps: agentDef.maxSteps,
-      model: agentDef.model,
-      allowedTools: agentDef.allowedTools,
-      startTime: Date.now(),
-
-      // Original task prompt — consumed by the Reflector to infer task type.
-      taskPrompt,
-
-      // In-memory transcript mirror of the JSONL file — consumed by the
-      // Reflector for tool-usage / error-recovery analysis. The JSONL file
-      // remains the durable record; this array is the hot-path copy.
-      transcript: [],
-
-      // Token tracking
-      inputTokens: 0,
-      outputTokens: 0,
-
-      // State
-      terminated: false,
-      result: null
-    };
-
-    this.writeTranscriptEntry(transcriptPath, {
-      type: "agent_start",
-      agentId,
-      agentName: agentDef.name,
-      taskPrompt,
-      timestamp: Date.now()
-    });
-
-    logger.info({ agentId, agentName: agentDef.name }, "Created fresh subagent context");
-
-    return context;
-  }
-
-  /**
-   * Add message to subagent context
-   */
-  addMessage(context, message) {
-    context.messages.push(message);
-
-    this.writeTranscriptEntry(context.transcriptPath, {
-      type: "message",
-      agentId: context.agentId,
-      message,
-      timestamp: Date.now()
-    });
-  }
-
-  /**
-   * Record tool execution in transcript
-   */
-  recordToolCall(context, toolName, input, output, error = null) {
-    const entry = {
-      type: "tool_call",
-      agentId: context.agentId,
-      step: context.steps,
-      toolName,
-      input,
-      output: error ? null : output,
-      error: error ? error.message : null,
-      timestamp: Date.now()
-    };
-
-    // Keep the in-memory transcript in sync so the Reflector (which reads
-    // context.transcript) sees tool calls without re-parsing the JSONL file.
-    if (Array.isArray(context.transcript)) {
-      context.transcript.push(entry);
-    }
-
-    this.writeTranscriptEntry(context.transcriptPath, entry);
-  }
-
-  /**
-   * Complete subagent execution
-   */
-  completeExecution(context, result) {
-    context.terminated = true;
-    context.result = result;
-
-    this.writeTranscriptEntry(context.transcriptPath, {
-      type: "agent_complete",
-      agentId: context.agentId,
-      result,
-      stats: {
-        steps: context.steps,
-        durationMs: Date.now() - context.startTime,
-        inputTokens: context.inputTokens,
-        outputTokens: context.outputTokens
-      },
-      timestamp: Date.now()
-    });
-
-    logger.info({
-      agentId: context.agentId,
-      steps: context.steps,
-      durationMs: Date.now() - context.startTime
-    }, "Subagent execution completed");
-  }
-
-  /**
-   * Fail subagent execution
-   */
-  failExecution(context, error) {
-    context.terminated = true;
-
-    this.writeTranscriptEntry(context.transcriptPath, {
-      type: "agent_failed",
-      agentId: context.agentId,
-      error: error.message,
-      stats: {
-        steps: context.steps,
-        durationMs: Date.now() - context.startTime
-      },
-      timestamp: Date.now()
-    });
-
-    logger.error({
-      agentId: context.agentId,
-      error: error.message
-    }, "Subagent execution failed");
-  }
-
-  /**
-   * Write entry to transcript file (JSONL format)
-   */
-  writeTranscriptEntry(transcriptPath, entry) {
-    try {
-      fs.appendFileSync(transcriptPath, JSON.stringify(entry) + "\n");
-    } catch (error) {
-      logger.warn({ error: error.message }, "Failed to write transcript entry");
-    }
-  }
-
-  /**
-   * Read transcript for debugging
-   */
-  readTranscript(agentId) {
-    const transcriptPath = path.join(this.transcriptsDir, `agent-${agentId}.jsonl`);
-
-    if (!fs.existsSync(transcriptPath)) {
-      return null;
-    }
-
-    const lines = fs.readFileSync(transcriptPath, "utf8").split("\n").filter(l => l.trim());
-    return lines.map(line => JSON.parse(line));
-  }
-
-  /**
-   * Generate unique agent ID
-   */
-  generateAgentId() {
-    return `${Date.now()}_${Math.random().toString(36).slice(2, 10)}`;
-  }
-
-  /**
-   * Clean old transcripts (older than 7 days)
-   */
-  cleanOldTranscripts() {
-    const maxAge = 7 * 24 * 60 * 60 * 1000; // 7 days
-    const now = Date.now();
-
-    const files = fs.readdirSync(this.transcriptsDir);
-
-    for (const file of files) {
-      const filePath = path.join(this.transcriptsDir, file);
-      const stats = fs.statSync(filePath);
-
-      if (now - stats.mtimeMs > maxAge) {
-        fs.unlinkSync(filePath);
-        logger.debug({ file }, "Cleaned old transcript");
-      }
-    }
-  }
-}
-
-module.exports = ContextManager;

package/src/agents/decomposition/dispatcher.js

@@ -1,185 +0,0 @@
-/**
- * Subtask dispatcher (Phase 3).
- *
- * Executes a validated plan respecting its dependency DAG:
- *   - subtasks are grouped into topological "levels" (Kahn's algorithm)
- *   - subtasks in the same level have no dependency on each other → run in
- *     parallel via the existing ParallelCoordinator (spawnParallel)
- *   - a subtask receives ONLY its own prompt plus the compressed results of the
- *     subtasks it depends on (context isolation — the token win)
- *
- * The spawn functions are injectable for testing.
- */
-
-const logger = require("../../logger");
-
-// Cap how much of a dependency's result we forward, to preserve the
-// context-isolation savings (subagents already return summaries; this bounds
-// pathological cases).
-const MAX_CONTEXT_CHARS = 2000;
-
-/**
- * Group subtasks into dependency levels. Returns array of arrays of ids.
- * Throws if the graph is unresolvable (should not happen — planner validated).
- */
-function topologicalLevels(subtasks) {
-  const byId = new Map(subtasks.map((s) => [s.id, s]));
-  const indegree = new Map(subtasks.map((s) => [s.id, 0]));
-  const dependents = new Map(subtasks.map((s) => [s.id, []]));
-
-  for (const s of subtasks) {
-    for (const dep of s.dependsOn) {
-      if (!byId.has(dep)) continue;
-      indegree.set(s.id, indegree.get(s.id) + 1);
-      dependents.get(dep).push(s.id);
-    }
-  }
-
-  const levels = [];
-  let frontier = subtasks.filter((s) => indegree.get(s.id) === 0).map((s) => s.id);
-  const resolved = new Set();
-
-  while (frontier.length > 0) {
-    levels.push(frontier);
-    const next = [];
-    for (const id of frontier) {
-      resolved.add(id);
-      for (const child of dependents.get(id)) {
-        indegree.set(child, indegree.get(child) - 1);
-        if (indegree.get(child) === 0) next.push(child);
-      }
-    }
-    frontier = next;
-  }
-
-  if (resolved.size !== subtasks.length) {
-    throw new Error("Unresolvable subtask graph (cycle or dangling dependency)");
-  }
-  return levels;
-}
-
-function compressResult(text) {
-  if (typeof text !== "string") return String(text ?? "");
-  if (text.length <= MAX_CONTEXT_CHARS) return text;
-  return text.slice(0, MAX_CONTEXT_CHARS) + "\n…[truncated]";
-}
-
-function buildContextForSubtask(subtask, resultsById) {
-  if (!subtask.dependsOn || subtask.dependsOn.length === 0) return null;
-  const parts = [];
-  for (const dep of subtask.dependsOn) {
-    const r = resultsById.get(dep);
-    if (r && r.success && r.result) {
-      parts.push(`Result of subtask ${dep}:\n${compressResult(r.result)}`);
-    } else if (r) {
-      parts.push(`Subtask ${dep} did not complete successfully.`);
-    }
-  }
-  return parts.length > 0 ? parts.join("\n\n") : null;
-}
-
-/**
- * Dispatch a validated plan.
- * @param {Object} plan - { subtasks: [...] }
- * @param {Object} [options]
- * @param {string} [options.sessionId]
- * @param {string} [options.cwd]
- * @param {Function} [options.spawnParallel] - (agentTypes[], prompts[], opts) => results[]
- * @returns {Promise<{results: Array, levels: Array, stats: Object}>}
- */
-async function dispatchPlan(plan, options = {}) {
-  const spawnParallel = options.spawnParallel || require("../index").spawnParallel;
-  const subtasks = plan.subtasks;
-  const byId = new Map(subtasks.map((s) => [s.id, s]));
-  const levels = topologicalLevels(subtasks);
-  const resultsById = new Map();
-
-  let totalInputTokens = 0;
-  let totalOutputTokens = 0;
-  let totalSubagents = 0;
-
-  for (let li = 0; li < levels.length; li++) {
-    const levelIds = levels[li];
-    const levelSubtasks = levelIds.map((id) => byId.get(id));
-
-    const agentTypes = levelSubtasks.map((s) => s.agentType);
-    const prompts = levelSubtasks.map((s) => s.prompt);
-    const perTaskContext = levelSubtasks.map((s) => buildContextForSubtask(s, resultsById));
-
-    logger.info(
-      { level: li, count: levelIds.length, ids: levelIds },
-      "[Decomposition] Dispatching subtask level"
-    );
-
-    // spawnParallel shares one options object; pass per-task context by spawning
-    // the level as individual parallel calls when contexts differ.
-    const levelResults = await runLevel(
-      spawnParallel,
-      agentTypes,
-      prompts,
-      perTaskContext,
-      options
-    );
-
-    levelResults.forEach((res, idx) => {
-      const st = levelSubtasks[idx];
-      totalSubagents += 1;
-      totalInputTokens += res?.stats?.inputTokens || 0;
-      totalOutputTokens += res?.stats?.outputTokens || 0;
-      resultsById.set(st.id, {
-        id: st.id,
-        agentType: st.agentType,
-        success: !!res?.success,
-        result: res?.success ? res.result : null,
-        error: res?.success ? null : res?.error || "unknown error",
-        stats: res?.stats || {},
-      });
-    });
-  }
-
-  const results = subtasks.map((s) => resultsById.get(s.id));
-  return {
-    results,
-    levels,
-    stats: {
-      subagents: totalSubagents,
-      inputTokens: totalInputTokens,
-      outputTokens: totalOutputTokens,
-    },
-  };
-}
-
-/**
- * Run one level. When subtasks in the level have differing injected contexts we
- * spawn them as separate parallel calls (each with its own mainContext), then
- * await all. When none need context, a single spawnParallel batch is used.
- */
-async function runLevel(spawnParallel, agentTypes, prompts, perTaskContext, options) {
-  const anyContext = perTaskContext.some((c) => c);
-
-  if (!anyContext) {
-    return spawnParallel(agentTypes, prompts, {
-      sessionId: options.sessionId,
-      cwd: options.cwd,
-    });
-  }
-
-  // Mixed/with-context: one spawnParallel call per subtask so each gets its own
-  // mainContext, executed concurrently.
-  const calls = agentTypes.map((type, i) =>
-    spawnParallel([type], [prompts[i]], {
-      sessionId: options.sessionId,
-      cwd: options.cwd,
-      mainContext: perTaskContext[i] ? { relevant_context: perTaskContext[i] } : undefined,
-    }).then((arr) => arr[0])
-  );
-  return Promise.all(calls);
-}
-
-module.exports = {
-  dispatchPlan,
-  topologicalLevels,
-  buildContextForSubtask,
-  compressResult,
-  MAX_CONTEXT_CHARS,
-};