@yeaft/webchat-agent 0.1.903 → 0.1.905

package/yeaft/session.js

@@ -17,7 +17,7 @@ import { initYeaftDir, DEFAULT_YEAFT_DIR, isWritable } from './init.js';
 import { loadConfig, loadMCPConfig } from './config.js';
 import { createTrace } from './debug-trace.js';
 import { createLLMAdapter } from './llm/adapter.js';
-import { ConversationStore } from './conversation/persist.js';
+import { ConversationStore, setDefaultRecentTurnsLimit } from './conversation/persist.js';
 import { SkillManager, createSkillManager } from './skills.js';
 import { MCPManager } from './mcp.js';
 import { createFullRegistry } from './tools/index.js';
@@ -156,6 +156,15 @@ export async function loadSession(options = {}) {
   // unref it (CLI / tests). Non-persisted — set per-session by caller.
   if (serverMode) config.serverMode = true;
 
+  // Propagate the (clamped) cold-start replay window to the conversation
+  // store. The default is 20 turns; a user wanting more recall after a
+  // fresh boot sets `yeaft.recentTurnsLimit` in ~/.yeaft/config.json.
+  // Called once per session boot — subsequent boots overwrite the
+  // module-level default safely (single-process model).
+  if (config?.yeaft?.recentTurnsLimit) {
+    setDefaultRecentTurnsLimit(config.yeaft.recentTurnsLimit);
+  }
+
   // ─── 2.1 Migration state check (task-334i) ────────────
   //        If the group-chat feature flag is on but migration has not
   //        completed, warn the user. Do NOT auto-run migration: that is

package/yeaft/snapshot-filter.js

@@ -0,0 +1,88 @@
+/**
+ * snapshot-filter.js — VP-isolated view of the in-memory group history.
+ *
+ * Mirrors the disk-replay rules in
+ * `agent/yeaft/conversation/persist.js#loadSessionHistoryForVp`:
+ *
+ *   - User rows (no `speakerVpId`): KEEP — every VP sees the prompt.
+ *   - This VP's own assistant rows + their paired tool rows: KEEP.
+ *   - OTHER VPs' assistant rows: KEEP TEXT ONLY (strip `toolCalls` AND
+ *     `thinkingBlocks` — thinking is VP-private per Anthropic's signed-
+ *     block contract; tool_use ids belong to that VP's own tool arc).
+ *   - OTHER VPs' tool result rows (role:'tool'): DROP — they pair with
+ *     stripped tool_use ids and would orphan on the LLM request.
+ *   - Rows with `_reflection` / `internal` / `systemOnly`: DROP — engine-
+ *     private; never enter another VP's context.
+ *
+ * Why this exists separately from the disk replay path:
+ *   `web-bridge.js` builds an in-memory `baseSnapshot` for every running
+ *   VP turn. Before this filter the snapshot was only `threadId`-scoped
+ *   and leaked other VPs' tool calls + thinking blocks into the next
+ *   turn's messages, producing orphan `tool_use` ids and Anthropic 422s.
+ *   The disk path uses `ConversationStore.loadSessionHistoryForVp` —
+ *   the same rules implemented here, but reading from `messages/*.md`
+ *   instead of the in-memory tape. We intentionally duplicate the rule
+ *   set rather than calling into persist.js so both call sites stay
+ *   independently audit-able.
+ *
+ * Pure function; does not mutate inputs.
+ *
+ * @param {object[]} snapshot — entries from getOrCreateSessionHistory(sessionId)
+ *                              (already threadId-filtered by the caller)
+ * @param {string} vpId — the VP we're about to send a turn for
+ * @returns {object[]}
+ */
+export function filterSnapshotForVp(snapshot, vpId) {
+  if (!Array.isArray(snapshot) || snapshot.length === 0) return [];
+  const out = [];
+  for (const m of snapshot) {
+    if (!m || typeof m !== 'object') continue;
+    // Reflection / engine-private rows are dropped regardless of vpId.
+    // Even the "no vpId, give me everything" code path must not leak
+    // these into a snapshot the caller will hand to an LLM — they were
+    // never meant to enter another VP's context, nor any VP's.
+    if (m._reflection || m.internal || m.systemOnly || m.systemOnlyMessage) continue;
+    if (!vpId) {
+      // No VP scope known: treat every row as "own". This is the
+      // fallback used by callers that don't (yet) have a vpId; the
+      // reflection filter above still applies.
+      out.push(m);
+      continue;
+    }
+    if (m.role === 'user') {
+      out.push(m);
+      continue;
+    }
+    if (m.role === 'assistant') {
+      if (!m.speakerVpId || m.speakerVpId === vpId) {
+        // Own assistant turn OR un-attributed (pre-rename) row — keep
+        // intact so this VP's tool arcs survive untouched.
+        out.push(m);
+      } else {
+        // Other VP's assistant turn — keep the visible text only.
+        // Stripping toolCalls is what makes the paired 'tool' rows
+        // below safe to drop without leaving orphan tool_use ids in
+        // the LLM payload. Stripping thinkingBlocks is required by
+        // Anthropic — signatures are VP-private and would fail
+        // server-side verification if echoed by a different VP.
+        const copy = { ...m };
+        delete copy.toolCalls;
+        delete copy.thinkingBlocks;
+        out.push(copy);
+      }
+      continue;
+    }
+    if (m.role === 'tool') {
+      // Tool results belong to the assistant turn that emitted the
+      // tool_use. We only keep ours; the other VPs' results paired
+      // with `toolCalls` we just stripped above, so they'd be
+      // orphans now anyway.
+      if (!m.speakerVpId || m.speakerVpId === vpId) out.push(m);
+      continue;
+    }
+    // Unknown role — keep as-is so future schema additions don't get
+    // silently dropped.
+    out.push(m);
+  }
+  return out;
+}