pi-mega-compact 0.6.1 → 0.6.3

package/extensions/mega-conflict-cmds.ts

@@ -11,6 +11,8 @@ import type { ExtensionAPI, ExtensionContext } from "@earendil-works/pi-coding-a
 import { detectConflicts, type ConflictReport } from "./conflict-scan.js";
 import { addMemory, listMemories, searchMemories, recallMemory, type MemoryRecord } from "../src/store/sqlite.js";
 import { resolveRepoRoot } from "./mega-config.js";
+import { defaultEmbedder } from "../src/embedder.js";
+import { upsertMemoryEmbedding } from "../src/store/memoryIndex.js";
 import { MegaRuntime } from "./mega-runtime.js";
 
 /** Run the conflict scan and format a human-readable report. */
@@ -82,6 +84,13 @@ export function registerConflictCommands(pi: ExtensionAPI, runtime: MegaRuntime)
         const tagMatches = [...text.matchAll(/#([\w-]+)/g)].map((m) => m[1]);
         const content = text.replace(/#[\w-]+/g, "").trim();
         const id = addMemory({ content, tags: tagMatches }, repo, runtime.currentStateDir);
+        // S24: mirror into the cross-repo memory index (fire-and-forget).
+        try {
+          const vec = defaultEmbedder().embed(content);
+          void upsertMemoryEmbedding(repo, id, content, vec);
+        } catch {
+          /* non-fatal */
+        }
         ctx.ui.notify(`[mega-memory] saved #${id} to ${repo.split(/[\\/]/).pop()}`);
         return;
       }
@@ -151,6 +160,12 @@ export function registerConflictCommands(pi: ExtensionAPI, runtime: MegaRuntime)
         const tagMatches = [...text.matchAll(/#([\w-]+)/g)].map((m) => m[1]);
         const content = text.replace(/#[\w-]+/g, "").trim();
         const id = addMemory({ content, tags: tagMatches }, repo, runtime.currentStateDir);
+        try {
+          const vec = defaultEmbedder().embed(content);
+          void upsertMemoryEmbedding(repo, id, content, vec);
+        } catch {
+          /* non-fatal */
+        }
         ctx.ui.notify(`[/m] saved #${id} to ${repo.split(/[\\/]/).pop()}`);
         return;
       }

package/extensions/mega-events.ts

@@ -19,8 +19,18 @@ import { driveNativeCompaction } from "./mega-compact-driver.js";
 import { computeLiveTrimCut, liveTrimSummaryMessage } from "./mega-trim.js";
 import { pressureFromPct, memoryReviewCadence, type MegaConfig } from "./mega-config.js";
 
+/**
+ * DIAG accessor for the headless test harness: the most recently constructed
+ * MegaRuntime, so a test that loads the compiled extension via its default
+ * export can read diag counters (diagLiveTrimFires / diagBeforeCompactFires /
+ * diagBeforeCompactSupplied / diagAgentEndIdle) after firing synthetic events.
+ * No-op in production — nothing reads this outside tests.
+ */
+export let lastRuntime: MegaRuntime | undefined;
+
 /** Register all pi lifecycle event handlers. */
 export function registerEventHandlers(pi: ExtensionAPI, runtime: MegaRuntime, config: MegaConfig): void {
+  lastRuntime = runtime;
   // ---- Session lifecycle (state reset points) -------------------------------
   // Capture model/provider whenever it changes (drives real cost estimation).
   pi.on("model_select", async (_event, ctx) => {
@@ -61,6 +71,8 @@ export function registerEventHandlers(pi: ExtensionAPI, runtime: MegaRuntime, co
       try {
         const mr = await recallMemoriesAndInline({
           query, stateDir: runtime.getStateDir(), limit: 5,
+          crossRepo: config.crossRepoEnabled,
+          crossRepoCosine: config.crossRepoCosine,
         });
         if (!mr.empty) runtime.pendingMemoryRecallBlock = mr.block;
       } catch (err) {
@@ -86,7 +98,7 @@ export function registerEventHandlers(pi: ExtensionAPI, runtime: MegaRuntime, co
         }
         // S21: parallel memory recall. Trigram embedder is sub-ms; await is fine.
         try {
-          const mr = await recallMemoriesAndInline({ query, stateDir: runtime.getStateDir(), limit: 5 });
+          const mr = await recallMemoriesAndInline({ query, stateDir: runtime.getStateDir(), limit: 5, crossRepo: config.crossRepoEnabled, crossRepoCosine: config.crossRepoCosine });
           if (!mr.empty) runtime.pendingMemoryRecallBlock = mr.block;
         } catch (err) {
           runtime.logger.warn("memory-recall skipped", { err: String(err) });
@@ -143,6 +155,46 @@ export function registerEventHandlers(pi: ExtensionAPI, runtime: MegaRuntime, co
         const idle = ctx.isIdle?.() ?? true;
         const queued = ctx.hasPendingMessages?.() ?? false;
         const now = Date.now();
+        // DIAG (team-run relief): surface whether the agent is idle + over
+        // threshold at agent_end so we can see if a mid-run durable-trim trigger
+        // *should* have fired but didn't.
+        const overThreshold = (runtime.lastCtxTokens ?? 0) >= config.thresholdTokens;
+        runtime.diagAgentEndIdle++;
+        runtime.logger.info("agent-end-idle", {
+          sessionId: runtime.rt.sessionId,
+          idle,
+          queued,
+          overThreshold,
+          ctxPct: runtime.lastCtxPercent,
+          ctxTokens: runtime.lastCtxTokens,
+          thresholdTokens: config.thresholdTokens,
+          wouldNudge: idle && queued && now >= runtime.resumeNudgeUntil,
+        });
+        // S16+S24: MID-RUN DURABLE TRIM. During a long team run (sub-agents),
+        // pi's native durable compaction only fires from _checkCompaction at
+        // PARENT settle (agent-session.js:760/844), so the on-disk transcript +
+        // context meter balloon to ~150k and never relieve until the very end
+        // ("compacts but doesn't resume"). agent_end with activeAgents===0 is a
+        // SAFE, settled point: calling ctx.compact() here does NOT abort an
+        // in-flight turn (the S16 danger is only mid-turn). ctx.compact() runs
+        // pi's flow, which fires our session_before_compact handler to supply
+        // the durable trim (truncates the transcript from firstKeptEntryId).
+        // Guarded three ways: only when truly idle + over threshold, only when
+        // pi would actually compact (piCompactWouldNoop skips the user-facing
+        // no-op throw), and debounced (one durable trim per 2s) to avoid
+        // thrashing the transcript while sub-agents keep settling.
+        if (idle && overThreshold && now >= runtime.debounceUntil) {
+          if (!piCompactWouldNoop(ctx)) {
+            runtime.debounceUntil = now + 2000;
+            runtime.diagAgentEndDurable++;
+            runtime.logger.info("agent-end-durable-trigger", {
+              sessionId: runtime.rt.sessionId,
+              ctxTokens: runtime.lastCtxTokens,
+              thresholdTokens: config.thresholdTokens,
+            });
+            ctx.compact({ customInstructions: undefined }); // guardrails-allow PREVENT-PI-004: local ctx.compact() — no network; agent settled so no in-flight abort
+          }
+        }
         if (idle && queued && now >= runtime.resumeNudgeUntil) {
           runtime.resumeNudgeUntil = now + 30_000;
           pi.sendUserMessage("[mega-compact] continue from the compacted context above.");
@@ -215,21 +267,21 @@ export function registerEventHandlers(pi: ExtensionAPI, runtime: MegaRuntime, co
       Math.round((pct / 100) * (usage?.contextWindow ?? 0));
 
     // FAST GATE: token-based (tier threshold), not percentage-based.
-    if (currentTokens < config.thresholdTokens) return;
+    if (currentTokens < config.thresholdTokens) { runtime.diagCtxFastGate++; return; }
 
     const check = autoCompactCheck(currentTokens, config.thresholdTokens); // SERVER-STYLE CONFIRM (local)
-    if (!check.shouldCompact) return;
+    if (!check.shouldCompact) { runtime.diagCtxNoCompact++; return; }
 
     // Debounce so we don't fire on every context event past threshold.
     const now = Date.now();
-    if (now < runtime.debounceUntil) return;
+    if (now < runtime.debounceUntil) { runtime.diagCtxDebounce++; return; }
     runtime.debounceUntil = now + 2000;
 
     // Adaptive compression (Fix E): scale compression strength + keepFrom depth
     // with how close we are to the model context limit.
     const pressure = pressureFromPct(pct);
     const ran = runCompact(pi, runtime, config, ctx, messages, { compressionPressure: pressure });
-    if (ran.skipped) return;
+    if (ran.skipped) { runtime.diagCtxRunSkipped++; return; }
 
     // LEGACY path (rollback): v0.4.28 ctx.compact() + the no-op gate. The
     // manual compact path aborts the in-flight turn — only used behind the flag.
@@ -262,7 +314,16 @@ export function registerEventHandlers(pi: ExtensionAPI, runtime: MegaRuntime, co
         summary: ran.result.summary,
         anchorUserMessages,
       });
-      if (cut === null) return; // unsafe / below anchor floor — no trim this call
+      if (cut === null) {
+        runtime.diagCtxCutNull++;
+        runtime.logger.info("live-trim-skip", {
+          sessionId: runtime.rt.sessionId,
+          compactedFrom: ran.result.compactedFrom,
+          viewLen: view.length,
+          anchorUserMessages,
+        });
+        return; // unsafe / below anchor floor — no trim this call
+      }
       const summaryMsg = liveTrimSummaryMessage({
         compactedFrom: ran.result.compactedFrom,
         summary: ran.result.summary,
@@ -276,8 +337,22 @@ export function registerEventHandlers(pi: ExtensionAPI, runtime: MegaRuntime, co
       } as unknown as AgentMessage;
       const recent = messages.slice(cut); // guardrails-allow PREVENT-PI-002: `cut` is the pre-sanitized `compactedFrom` produced by src/boundary.ts computeDropRange, so the preserved run begins on a toolPair-safe index.
       runtime.snapshot(ctx);
+      // DIAG (team-run relief): confirm the live trim actually fires + how big
+      // the window still is. The return is non-durable (per-LLM-call only), so
+      // this is the signal that the model is being fed a compacted view while
+      // the on-disk transcript + context meter keep growing.
+      runtime.diagLiveTrimFires++;
+      runtime.logger.info("live-trim", {
+        sessionId: runtime.rt.sessionId,
+        inputMsgs: messages.length,
+        outputMsgs: recent.length + 1,
+        compactedFrom: cut,
+        ctxPct: pct,
+        ctxTokens: usage?.tokens ?? null,
+      });
       return { messages: [summaryAgentMsg, ...recent] };
     } catch {
+      runtime.diagCtxThrown++;
       return; // non-fatal: no trim this call; the next context event retries
     }
   });
@@ -290,10 +365,25 @@ export function registerEventHandlers(pi: ExtensionAPI, runtime: MegaRuntime, co
   // there is no full-reload + additive recall inflation.
   pi.on("session_before_compact", async (event: SessionBeforeCompactEvent, ctx: ExtensionContext) => {
     runtime.resetRuntime(ctx.sessionManager.getSessionId());
+    // DIAG (team-run relief): this is the ONLY durable-trim entry point. Log
+    // every fire + whether we supplied a compaction (truncates transcript) or
+    // fell through to {} (pi runs its own). If this is sparse during a team
+    // run, the durable trim is firing too late (only at parent settle).
+    const prep = event.preparation;
+    runtime.diagBeforeCompactFires++;
+    runtime.logger.info("before-compact-entry", {
+      sessionId: runtime.rt.sessionId,
+      reason: event.reason,
+      hasPrep: !!prep,
+      msgsToSummarize: prep?.messagesToSummarize?.length ?? 0,
+      firstKeptEntryId: prep?.firstKeptEntryId ?? null,
+      activeAgents: runtime.activeAgents,
+    });
     if (!config.auto) return {}; // let pi run its own native compaction
     try {
       const result = driveNativeCompaction(event, runtime, config);
       if (result) {
+        runtime.diagBeforeCompactSupplied++;
         runtime.logger.info("native-compact", {
           sessionId: runtime.rt.sessionId,
           firstKeptEntryId: result.compaction.firstKeptEntryId,

package/extensions/mega-runtime.ts

@@ -143,6 +143,27 @@ export class MegaRuntime {
   lastCtxPercent: number | null = null;
   lastCtxWindow = 0;
 
+  /**
+   * DIAG counters for the "team run doesn't relieve context" investigation.
+   * Plain integers, incremented at the three compaction decision points. They
+   * let a headless test drive the real event handlers and assert the firing
+   * cadence without scraping log files. Inert in production (the live-trim and
+   * before-compact probes also emit logger.info, but these counters are always
+   * updated and cost nothing).
+   */
+  diagLiveTrimFires = 0;     // context handler returned a trimmed view
+  diagBeforeCompactFires = 0; // session_before_compact handler entered
+  diagBeforeCompactSupplied = 0; // session_before_compact supplied our trim
+  diagAgentEndIdle = 0;       // agent_end with activeAgents===0
+  diagAgentEndDurable = 0;    // agent_end fired ctx.compact() (mid-run durable trim)
+  // Per-skip-path counters for the team-run diagnosis.
+  diagCtxFastGate = 0;        // returned at token fast-gate (below threshold)
+  diagCtxNoCompact = 0;       // autoCompactCheck().shouldCompact === false
+  diagCtxDebounce = 0;        // debounceUntil not yet elapsed
+  diagCtxRunSkipped = 0;      // runCompact() returned skipped
+  diagCtxCutNull = 0;         // computeLiveTrimCut returned null (anchor/boundary)
+  diagCtxThrown = 0;          // live-trim try threw (caught)
+
   /**
    * Live 0–1 pressure: how full the context window is relative to the compaction
    * threshold. Computed from the most recent context event the runtime already

package/extensions/mega-teamrun.test.ts

@@ -0,0 +1,164 @@
+/**
+ * mega-teamrun.test.ts — regression test for the "auto-compact runs but context
+ * never relieves during a team run (sub-agents)" bug.
+ *
+ * Loads the REAL compiled extension (extensions/mega-compact.js) through a
+ * faithful mock pi (mirrors mega-compact.test.ts's harness) and drives the
+ * exact event sequence a long team run produces:
+ *
+ *   agent_start -> context (over threshold) xN -> agent_end   (repeat x3)
+ *
+ * Asserts the TWO fixes:
+ *   1. live trim FIRES per-call (computeLiveTrimCut no longer returns null on
+ *      the anchor floor — was `cutNull`, liveTrimFires===0 before the fix).
+ *   2. the DURABLE trim fires at agent_end while idle + over threshold
+ *      (mid-run durable trigger), not only at parent settle.
+ *
+ * The mock ctx.compact() drives session_before_compact so we observe the
+ * durable truncation. Counters come from MegaRuntime.diag* (set behind the
+ * real handler code, inert in production).
+ *
+ * MEGACOMPACT_PGLITE_DISABLED keeps the run fast (no WASM index init).
+ */
+
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { mkdtempSync, rmSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { createRequire } from "node:module";
+import { closeVectorIndex } from "../src/store/vectorIndex.js";
+import type { AgentMessage } from "@earendil-works/pi-agent-core";
+
+const require = createRequire(import.meta.url);
+const baseTmp = mkdtempSync(join(tmpdir(), "mc-team-"));
+process.env.MEGACOMPACT_INDEX_DIR = join(baseTmp, "index");
+process.env.MEGACOMPACT_PGLITE_DISABLED = "true"; // fast: skip WASM index
+let counter = 0;
+
+function harness() {
+  const stateDir = join(baseTmp, `run-${counter++}`);
+  process.env.MEGACOMPACT_STATE_DIR = stateDir;
+  process.env.MEGACOMPACT_DEBUG = "true";
+  process.env.MEGACOMPACT_THRESHOLD_TOKENS = "50";
+  process.env.MEGACOMPACT_FAST_GATE_PCT = "1";
+  process.env.MEGACOMPACT_ANCHOR_USER_MESSAGES = "1";
+  process.env.MEGACOMPACT_DURABLE_TRIM_FLOOR = "0"; // piCompactWouldNoop must not skip
+  process.env.MEGACOMPACT_MEMORY_AUTO_REVIEW = "false";
+  process.env.MEGACOMPACT_RAPTOR_ENABLED = "false";
+  delete process.env.MEGACOMPACT_LEGACY_DURABLE_TRIM;
+
+  const handlers: Record<string, Function> = {};
+  const compactCalls: any[] = [];
+
+  function msg(role: string, text: string, toolName?: string): AgentMessage {
+    if (role === "assistant" && toolName) {
+      return { role: "assistant", content: [{ type: "toolCall", name: toolName, id: "c1", arguments: {} }], api: "anthropic-messages", provider: "anthropic", model: "m", usage: { inputTokens: 1, outputTokens: 1, cacheReadTokens: 0, cacheWriteTokens: 0 }, stopReason: "tool_use", timestamp: 0 } as unknown as AgentMessage;
+    }
+    if (role === "toolResult" && toolName) {
+      return { role: "toolResult", content: [{ type: "text", text }], toolCallId: "c1", toolName, isError: false, timestamp: 0 } as unknown as AgentMessage;
+    }
+    return { role: "user", content: text, timestamp: 0 } as unknown as AgentMessage;
+  }
+
+  const session: AgentMessage[] = [];
+  for (let i = 0; i < 14; i++) {
+    session.push(msg("user", `actually we decided to use approach ${i} for module ${i}`));
+    session.push(msg("assistant", `edited module ${i}`, "Edit"));
+    session.push(msg("toolResult", `edited module ${i}`, "Edit"));
+  }
+
+  const toEntry = (m: AgentMessage, i: number): any => ({ type: "message", id: `e${i}`, parentId: null, timestamp: String(i), message: m });
+  const sessionManager = {
+    getSessionId: () => "sess_team_001",
+    getEntries: () => session.map(toEntry),
+    getBranch: () => session.map(toEntry),
+  };
+
+  function makeCtx(over: Partial<any> = {}) {
+    return {
+      ui: { setStatus: () => {}, notify: () => {}, select: () => {}, confirm: async () => true, input: async () => "", setWidget: () => {} },
+      mode: "tui" as any, hasUI: true, cwd: stateDir, sessionManager,
+      modelRegistry: {} as any, model: undefined, isIdle: () => true, isProjectTrusted: () => true,
+      signal: undefined, abort: () => {}, hasPendingMessages: () => false, shutdown: () => {},
+      getContextUsage: () => ({ tokens: 200000, contextWindow: 200000, percent: 100 }),
+      // Mock ctx.compact() runs pi's flow and fires session_before_compact.
+      compact: (opts?: any) => {
+        compactCalls.push(opts);
+        if (handlers["session_before_compact"]) {
+          return handlers["session_before_compact"](
+            { type: "session_before_compact", reason: "threshold", willRetry: false, signal: undefined, preparation: { firstKeptEntryId: "e2", messagesToSummarize: session.slice(0, 2), tokensBefore: 500 } } as any,
+            makeCtx(),
+          );
+        }
+        return undefined;
+      },
+      getSystemPrompt: () => "system base",
+      ...over,
+    } as any;
+  }
+
+  const pi = {
+    on: (ev: string, h: Function) => { handlers[ev] = h; },
+    registerCommand: () => {}, registerTool: () => {}, registerShortcut: () => {},
+    registerFlag: () => {}, getFlag: () => undefined, registerMessageRenderer: () => {},
+    registerEntryRenderer: () => {}, sendMessage: () => {}, sendUserMessage: () => {},
+    appendEntry: () => {}, setSessionName: () => {}, getSessionName: () => undefined,
+    setLabel: () => {}, exec: async () => ({ stdout: "", stderr: "", code: 0 }),
+    getActiveTools: () => [], getAllTools: () => [], setActiveTools: () => {},
+    getCommands: () => [], setModel: async () => false, getThinkingLevel: () => "off" as any,
+    setThinkingLevel: () => {},
+  } as any;
+
+  const mod = require("./mega-compact.js") as { default: (p: any) => void };
+  mod.default(pi);
+  const { lastRuntime } = require("./mega-events.js") as { lastRuntime: any };
+
+  const fire = (ev: string, event: any, ctx: any) => handlers[ev](event, ctx);
+  return {
+    stateDir, handlers, compactCalls, fire, ctx: makeCtx, session,
+    runtime: lastRuntime, // MegaRuntime with diag* counters
+    // Advance the debounce so agent_end (same instant) can trigger durable trim.
+    clearDebounce: () => { if (lastRuntime) lastRuntime.debounceUntil = 0; },
+  };
+}
+
+test("team run: live trim fires AND durable trim fires per sub-agent (relieves context)", async () => {
+  const h = harness();
+  const ctx = h.ctx();
+  for (let a = 0; a < 3; a++) {
+    await h.fire("agent_start", { type: "agent_start", messages: [] }, ctx);
+    for (let i = 0; i < 4; i++) {
+      await h.fire("context", { type: "context", messages: h.session }, ctx);
+    }
+    // Real team runs settle seconds after the last context event; mimic that
+    // so the 2s debounce has elapsed and the durable trigger can fire.
+    await new Promise((r) => setTimeout(r, 2100));
+    h.clearDebounce();
+    await h.fire("agent_end", { type: "agent_end", messages: [] }, ctx);
+  }
+  const rt = h.runtime;
+  // FIX 1: live trim must actually fire (was 0 — computeLiveTrimCut returned null).
+  assert.ok(rt.diagLiveTrimFires > 0, "live trim fires during the team run (anchor-floor fix)");
+  assert.equal(rt.diagCtxCutNull, 0, "no live-trim cut skipped on anchor floor");
+  // FIX 2: durable trim must fire at each agent_end (was 0 — only at parent settle).
+  assert.equal(rt.diagAgentEndDurable, 3, "mid-run durable trigger fired at each agent_end");
+  assert.equal(rt.diagBeforeCompactSupplied, 3, "our durable trim supplied 3x (context relieved)");
+  assert.ok(h.compactCalls.length >= 3, "ctx.compact() invoked for durable trim between sub-agents");
+});
+
+test("control: session_before_compact supplies a durable compaction (parent settles)", async () => {
+  const h = harness();
+  const res = await h.fire(
+    "session_before_compact",
+    { type: "session_before_compact", reason: "threshold", willRetry: false, signal: undefined, preparation: { firstKeptEntryId: "e2", messagesToSummarize: h.session.slice(0, 4), tokensBefore: 500 } } as any,
+    h.ctx(),
+  );
+  assert.ok(res?.compaction, "compaction result returned to pi");
+  assert.equal(res.compaction.firstKeptEntryId, "e2", "reuses pi's boundary (PREVENT-PI-002)");
+});
+
+test("cleanup", async () => {
+  await closeVectorIndex();
+  rmSync(baseTmp, { recursive: true, force: true });
+});

package/extensions/mega-trim.ts

@@ -46,7 +46,34 @@ export function computeLiveTrimCut(view: EngineMessage[], opts: BuildLiveTrimVie
   if (cut <= 0) return null; // nothing safe to cut — keep everything this call
   const recent = view.slice(cut);
   const userCount = recent.filter((m) => m.role === "user").length;
-  if (userCount < opts.anchorUserMessages) return null;
+  // ANCHOR FLOOR (PREVENT-PI-001): the recent window must keep at least
+  // `anchorUserMessages` user messages. The original compactedFrom can land on a
+  // run that starts with fewer than that (e.g. the preserved region begins on a
+  // tool pair, or the session's tail is tool-heavy). Instead of bailing out and
+  // skipping the live trim entirely this call (which left the model fed a
+  // 150k-context window during long team runs), walk `cut` backward until the
+  // preserved run contains enough user messages — bounded by the boundary-safe
+  // constraint so we never split a tool pair. Falls back to null only when the
+  // whole view can't satisfy the floor (tiny sessions) — the next context event
+  // retries.
+  if (userCount < opts.anchorUserMessages) {
+    let c = cut;
+    while (c > 1) {
+      c--;
+      if (!isBoundarySafe(view, c)) continue;
+      const recentNow = view.slice(c);
+      const usersNow = recentNow.filter((m) => m.role === "user").length;
+      if (usersNow >= opts.anchorUserMessages) { cut = c; break; }
+    }
+    if (cut > 1) {
+      const finalRecent = view.slice(cut);
+      if (finalRecent.filter((m) => m.role === "user").length < opts.anchorUserMessages) {
+        return null; // cannot satisfy the floor without dropping too much — retry next call
+      }
+    } else {
+      return null;
+    }
+  }
   return cut;
 }
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-mega-compact",
-  "version": "0.6.1",
+  "version": "0.6.3",
   "description": "Layered, local, vector-backed context compressor for pi — supersede/collapse/cluster compaction with deduped inline recall.",
   "type": "module",
   "license": "BSD-2-Clause",

package/src/memoryOps.ts

@@ -12,6 +12,24 @@ import {
   removeMemory,
   type MemoryRecord,
 } from "./store/sqlite.js";
+import { defaultEmbedder } from "./embedder.js";
+import { upsertMemoryEmbedding } from "./store/memoryIndex.js";
+import { execSync } from "node:child_process"; // guardrails-allow PREVENT-PI-004: read-only `git rev-parse` to scope the memory index per-repo
+
+/** Resolve the current repo's git root (mirrors extensions/mega-config.ts but
+ *  kept local so src/ stays pi-agnostic — no extension-layer import). */
+function resolveRepoRootLocal(cwd: string): string | undefined {
+  try {
+    const out = execSync("git rev-parse --show-toplevel", {
+      cwd,
+      encoding: "utf-8",
+      stdio: ["ignore", "pipe", "ignore"],
+    }).trim();
+    return out || undefined;
+  } catch {
+    return undefined;
+  }
+}
 
 /** Find a memory row whose content exactly matches (case-insensitive). */
 function findByContent(memories: MemoryRecord[], content: string): MemoryRecord | undefined {
@@ -19,6 +37,23 @@ function findByContent(memories: MemoryRecord[], content: string): MemoryRecord
   return memories.find((m) => m.content.trim().toLowerCase() === norm);
 }
 
+/**
+ * Fire-and-forget mirror of a memory write into the cross-repo PGlite index
+ * (S24 optional memory-RAG mirror). Best-effort + non-fatal: never blocks the
+ * SQLite write and degrades to the same-repo scan if the index is disabled or
+ * fails. `repoId` is the resolved git root so the memory is findable from other
+ * repos; falls back to the state dir when outside git.
+ */
+function indexMemoryWrite(stateDir: string, memoryId: number, content: string): void {
+  const repoId = resolveRepoRootLocal(stateDir) ?? stateDir;
+  try {
+    const vec = defaultEmbedder().embed(content);
+    void upsertMemoryEmbedding(repoId, memoryId, content, vec);
+  } catch {
+    /* non-fatal — embedding/index failure must never break the SQLite write */
+  }
+}
+
 /**
  * Apply add/replace/remove ops to the memories table. Replaces are matched by
  * existing content; removes by content. Idempotent: an add that already exists
@@ -32,7 +67,7 @@ export async function applyMemoryOps(ops: MemoryOp[], stateDir: string): Promise
     if (op.op === "add") {
       // Skip if an identical memory already exists.
       if (findByContent(existing, op.memory.content)) continue;
-      addMemory(
+      const id = addMemory(
         {
           kind: op.memory.category,
           content: op.memory.content,
@@ -44,6 +79,8 @@ export async function applyMemoryOps(ops: MemoryOp[], stateDir: string): Promise
         repo,
         stateDir,
       );
+      // S24: mirror into the cross-repo index (fire-and-forget; non-fatal).
+      indexMemoryWrite(stateDir, id, op.memory.content);
     } else if (op.op === "replace") {
       const match = findByContent(existing, op.targetContent);
       if (match) {
@@ -53,9 +90,11 @@ export async function applyMemoryOps(ops: MemoryOp[], stateDir: string): Promise
           category: op.memory.category,
           sourceTurn: op.memory.sourceTurn,
         }, stateDir);
+        // S24: re-mirror under the same memory id (fire-and-forget; non-fatal).
+        indexMemoryWrite(stateDir, match.id, op.memory.content);
       } else {
         // Target missing (e.g. earlier in-conversation contradiction) → add.
-        addMemory(
+        const id = addMemory(
           {
             kind: op.memory.category,
             content: op.memory.content,
@@ -66,6 +105,7 @@ export async function applyMemoryOps(ops: MemoryOp[], stateDir: string): Promise
           repo,
           stateDir,
         );
+        indexMemoryWrite(stateDir, id, op.memory.content);
       }
     } else {
       const match = findByContent(existing, op.content);

package/src/memoryRecall.test.ts

@@ -98,3 +98,61 @@ test("recallMemories: fresher reference beats older at equal similarity", async
 test("cleanup memrec", () => {
   rmSync(baseTmp, { recursive: true, force: true });
 });
+
+// ---- S24: cross-repo memory recall (PGlite mirror) ---------------------------
+test("recallMemoriesAndInline: surfaces a memory saved in ANOTHER repo via cross-repo index", async () => {
+  // Isolate the global PGlite index to a temp dir shared by both "repos".
+  process.env.MEGACOMPACT_INDEX_DIR = join(baseTmp, "xrepo-index");
+  const repoA = join(baseTmp, "repo-a");
+  const repoB = join(baseTmp, "repo-b");
+  try {
+    // repoA owns a decision about the store backend.
+    const { applyMemoryOps } = await import("./memoryOps.js");
+    await applyMemoryOps(
+      [{ op: "add", memory: { content: "we standardized on node:sqlite for the store backend", category: "decision", sourceTurn: 0 } }],
+      repoA,
+    );
+    // repoB is a fresh session with NO local memory about the store backend.
+    const { recallMemoriesAndInline } = await import("./recall.js");
+    const res = await recallMemoriesAndInline({
+      query: "what store backend do we use?",
+      stateDir: repoB,
+      limit: 5,
+      crossRepo: true,
+      crossRepoCosine: 0.3,
+    });
+    assert.ok(!res.empty, "cross-repo recall found the other repo's memory");
+    assert.ok(/node:sqlite/.test(res.block), "the node:sqlite decision was recalled from repo A");
+    assert.ok(res.report.some((r) => /from /.test(r)), "report labels the memory as cross-repo");
+  } finally {
+    const { closeMemoryIndex } = await import("./store/memoryIndex.js");
+    await closeMemoryIndex();
+    delete process.env.MEGACOMPACT_INDEX_DIR;
+  }
+});
+
+test("recallMemoriesAndInline: cross-repo disabled when MEGACOMPACT_PGLITE_DISABLED", async () => {
+  process.env.MEGACOMPACT_INDEX_DIR = join(baseTmp, "xrepo-index-off");
+  process.env.MEGACOMPACT_PGLITE_DISABLED = "true";
+  const repoA = join(baseTmp, "repo-a2");
+  const repoB = join(baseTmp, "repo-b2");
+  try {
+    const { applyMemoryOps } = await import("./memoryOps.js");
+    await applyMemoryOps(
+      [{ op: "add", memory: { content: "we standardized on node:sqlite for the store backend", category: "decision", sourceTurn: 0 } }],
+      repoA,
+    );
+    const { recallMemoriesAndInline } = await import("./recall.js");
+    const res = await recallMemoriesAndInline({
+      query: "what store backend do we use?",
+      stateDir: repoB,
+      limit: 5,
+      crossRepo: true,
+    });
+    // Index disabled → no cross-repo hit; repoB has no local memory → empty.
+    assert.equal(res.empty, true, "cross-repo recall degrades to empty when disabled");
+  } finally {
+    delete process.env.MEGACOMPACT_PGLITE_DISABLED;
+    delete process.env.MEGACOMPACT_INDEX_DIR;
+  }
+});

package/src/memoryRecall.ts

@@ -81,3 +81,55 @@ export async function recallMemories(
   }
   return top;
 }
+
+/**
+ * Cross-repo memory recall (S24): augments the same-repo `recallMemories` with
+ * HNSW NN over the global PGlite `memory_index` (other repos' memories). Content
+ * is read inline from the index hit (the recall process can't open other repos'
+ * SQLite dirs), so no other-repo db access is required. Returns hits sorted by
+ * descending cosine, above `crossRepoCosine`. De-duped by content against
+ * `sameRepoContent` so we never surface a memory the same-repo scan already has.
+ * Non-fatal: any index failure returns []. Best-effort + PREVENT-PI-004 (local
+ * WASM only).
+ */
+export async function recallMemoriesCrossRepo(
+  query: string,
+  stateDir: string,
+  opts: RecallMemoriesOptions & { crossRepoCosine?: number; limit?: number } = {},
+): Promise<Array<{ memory: MemoryRecord; score: number; repoId: string }>> {
+  const embedder = opts.embedder ?? defaultEmbedder();
+  const queryVec = embedder.embed(query);
+  const { searchMemoriesAsync } = await import("./store/memoryIndex.js");
+  const k = opts.limit ?? 5;
+  const floor = opts.crossRepoCosine ?? 0.3;
+  const hits = await searchMemoriesAsync(queryVec, { k });
+  if (!hits.length) return [];
+  // Mark same-repo content as already-covered so we don't duplicate it.
+  const sameRepo = new Set(
+    listMemories(opts.repo ?? null, 1000, stateDir).map((m) => m.content.trim().toLowerCase()),
+  );
+  const out: Array<{ memory: MemoryRecord; score: number; repoId: string }> = [];
+  for (const h of hits) {
+    if (h.score < floor) continue;
+    if (sameRepo.has(h.content.trim().toLowerCase())) continue;
+    out.push({
+      memory: {
+        id: h.memoryId,
+        repo: h.repoId,
+        kind: "note",
+        content: h.content,
+        tags: [],
+        createdAt: 0,
+        lastRecalledAt: null,
+        category: null,
+        target: null,
+        lastReferenced: null,
+        sourceTurn: null,
+      } as MemoryRecord,
+      score: h.score,
+      repoId: h.repoId,
+    });
+  }
+  out.sort((a, b) => b.score - a.score);
+  return out;
+}