pi-mega-compact 0.6.1 → 0.6.2

package/dist/extensions/mega-conflict-cmds.js

@@ -9,6 +9,8 @@
 import { detectConflicts } from "./conflict-scan.js";
 import { addMemory, listMemories, searchMemories, recallMemory } from "../src/store/sqlite.js";
 import { resolveRepoRoot } from "./mega-config.js";
+import { defaultEmbedder } from "../src/embedder.js";
+import { upsertMemoryEmbedding } from "../src/store/memoryIndex.js";
 /** Run the conflict scan and format a human-readable report. */
 export function validateExtensions() {
     const report = detectConflicts();
@@ -74,6 +76,14 @@ export function registerConflictCommands(pi, runtime) {
                 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 ${repo.split(/[\\/]/).pop()}`);
                 return;
             }
@@ -142,6 +152,13 @@ export function registerConflictCommands(pi, runtime) {
                 const tagMatches = [...text.matchAll(/#([\w-]+)/g)].map((m) => m[1]);
                 const content = text.replace(/#[\w-]+/g, "").trim();
                 const id = addMemory({ content, tags: tagMatches }, repo, runtime.currentStateDir);
+                try {
+                    const vec = defaultEmbedder().embed(content);
+                    void upsertMemoryEmbedding(repo, id, content, vec);
+                }
+                catch {
+                    /* non-fatal */
+                }
                 ctx.ui.notify(`[/m] saved #${id} to ${repo.split(/[\\/]/).pop()}`);
                 return;
             }

package/dist/extensions/mega-events.js

@@ -15,8 +15,17 @@ import { recallMemoriesAndInline } from "../src/recall.js";
 import { driveNativeCompaction } from "./mega-compact-driver.js";
 import { computeLiveTrimCut, liveTrimSummaryMessage } from "./mega-trim.js";
 import { pressureFromPct, memoryReviewCadence } 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;
 /** Register all pi lifecycle event handlers. */
 export function registerEventHandlers(pi, runtime, config) {
+    lastRuntime = runtime;
     // ---- Session lifecycle (state reset points) -------------------------------
     // Capture model/provider whenever it changes (drives real cost estimation).
     pi.on("model_select", async (_event, ctx) => {
@@ -56,6 +65,8 @@ export function registerEventHandlers(pi, runtime, config) {
             try {
                 const mr = await recallMemoriesAndInline({
                     query, stateDir: runtime.getStateDir(), limit: 5,
+                    crossRepo: config.crossRepoEnabled,
+                    crossRepoCosine: config.crossRepoCosine,
                 });
                 if (!mr.empty)
                     runtime.pendingMemoryRecallBlock = mr.block;
@@ -82,7 +93,7 @@ export function registerEventHandlers(pi, runtime, config) {
                 }
                 // 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;
                 }
@@ -139,6 +150,46 @@ export function registerEventHandlers(pi, runtime, config) {
                 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.");
@@ -207,22 +258,30 @@ export function registerEventHandlers(pi, runtime, config) {
         const currentTokens = usage?.tokens ?? estimateSessionTokens(view) ??
             Math.round((pct / 100) * (usage?.contextWindow ?? 0));
         // FAST GATE: token-based (tier threshold), not percentage-based.
-        if (currentTokens < config.thresholdTokens)
+        if (currentTokens < config.thresholdTokens) {
+            runtime.diagCtxFastGate++;
             return;
+        }
         const check = autoCompactCheck(currentTokens, config.thresholdTokens); // SERVER-STYLE CONFIRM (local)
-        if (!check.shouldCompact)
+        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)
+        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)
+        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.
         // Read live from env (in addition to the load-time config) so the flag can be
@@ -254,8 +313,16 @@ export function registerEventHandlers(pi, runtime, config) {
                 summary: ran.result.summary,
                 anchorUserMessages,
             });
-            if (cut === null)
+            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,
@@ -269,9 +336,23 @@ export function registerEventHandlers(pi, runtime, config) {
             };
             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
         }
     });
@@ -283,11 +364,26 @@ export function registerEventHandlers(pi, runtime, config) {
     // there is no full-reload + additive recall inflation.
     pi.on("session_before_compact", async (event, ctx) => {
         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/dist/extensions/mega-runtime.js

@@ -120,6 +120,26 @@ export class MegaRuntime {
     lastCtxTokens = null;
     lastCtxPercent = 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/dist/extensions/mega-teamrun.test.js

@@ -0,0 +1,143 @@
+/**
+ * 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";
+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 = {};
+    const compactCalls = [];
+    function msg(role, text, toolName) {
+        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 };
+        }
+        if (role === "toolResult" && toolName) {
+            return { role: "toolResult", content: [{ type: "text", text }], toolCallId: "c1", toolName, isError: false, timestamp: 0 };
+        }
+        return { role: "user", content: text, timestamp: 0 };
+    }
+    const session = [];
+    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 ${i}`, "Edit"));
+    }
+    const toEntry = (m, i) => ({ type: "message", id: `e${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 = {}) {
+        return {
+            ui: { setStatus: () => { }, notify: () => { }, select: () => { }, confirm: async () => true, input: async () => "", setWidget: () => { } },
+            mode: "tui", hasUI: true, cwd: stateDir, sessionManager,
+            modelRegistry: {}, 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) => {
+                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 } }, makeCtx());
+                }
+                return undefined;
+            },
+            getSystemPrompt: () => "system base",
+            ...over,
+        };
+    }
+    const pi = {
+        on: (ev, h) => { 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",
+        setThinkingLevel: () => { },
+    };
+    const mod = require("./mega-compact.js");
+    mod.default(pi);
+    const { lastRuntime } = require("./mega-events.js");
+    const fire = (ev, event, ctx) => 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 } }, 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/dist/extensions/mega-trim.js

@@ -22,8 +22,39 @@ export function computeLiveTrimCut(view, opts) {
         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;
 }
 /** The formatted compacted-region summary as a user-role engine message. */

package/dist/src/memoryOps.js

@@ -1,9 +1,44 @@
 import { addMemory, listMemories, replaceMemory, removeMemory, } 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) {
+    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, content) {
     const norm = content.trim().toLowerCase();
     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, memoryId, content) {
+    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
@@ -19,7 +54,7 @@ export async function applyMemoryOps(ops, stateDir) {
             // 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,
                 tags: [],
@@ -27,6 +62,8 @@ export async function applyMemoryOps(ops, stateDir) {
                 target: op.memory.target,
                 sourceTurn: op.memory.sourceTurn,
             }, 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);
@@ -37,16 +74,19 @@ export async function applyMemoryOps(ops, stateDir) {
                     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,
                     tags: [],
                     category: op.memory.category,
                     sourceTurn: op.memory.sourceTurn,
                 }, repo, stateDir);
+                indexMemoryWrite(stateDir, id, op.memory.content);
             }
         }
         else {

package/dist/src/memoryRecall.js

@@ -58,3 +58,51 @@ export async function recallMemories(query, stateDir, opts = {}) {
     }
     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, stateDir, opts = {}) {
+    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 = [];
+    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,
+            },
+            score: h.score,
+            repoId: h.repoId,
+        });
+    }
+    out.sort((a, b) => b.score - a.score);
+    return out;
+}

package/dist/src/memoryRecall.test.js

@@ -90,3 +90,55 @@ 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;
+    }
+});