pi-mega-compact 0.20.85 → 0.20.87

package/dist/extensions/mega-pipeline/compact.js

@@ -1,345 +1,14 @@
 /**
- * compact.ts — full compaction pipeline (Trident) + pi no-op prediction.
- *
- * `runCompact` runs the full Trident pipeline (fast-gate aside) and persists a
- * checkpoint. `piCompactWouldNoop` predicts whether pi's `ctx.compact()` would
- * throw a no-op error. Both mutate the shared MegaRuntime (token accounting,
- * ticker, status, events) and are driven by the event + command handlers in
- * mega-events.ts / mega-commands.ts.
- */
-import { sessionEntryToContextMessages } from "@earendil-works/pi-coding-agent";
-import { compactSession } from "../../src/engine.js";
-import { normalizeSessionId } from "../../src/store.js";
-import { repoKey } from "../../src/store/repoKey.js";
-import { estimateBlockTokens } from "../../src/tokens.js";
-import { touchSession, logDaily, incCompactCount, incCacheHitTokens } from "../../src/store/sqlite.js";
-import { consolidateMemories } from "../../src/memory.js";
-import { C, MARKER_TYPE, } from "../mega-runtime.js";
-import { resolveRepoRoot, preserveRecentForPressure } from "../mega-config.js";
-import { runRaptor } from "../../src/dedup/raptor/index.js";
-import { isRaptorTreeFresh } from "../../src/dedup/raptor/buildHistory.js";
-import { loadDedupConfig } from "../../src/config/dedup.js";
-import { upsertEmbedding as indexUpsertEmbedding } from "../../src/store/vectorIndex.js";
-import { runMemoryReview } from "./memory-review.js";
-import { vectorList } from "../../src/vectorStore.js";
-/** Run the full compaction pipeline and persist a checkpoint. Returns the result. */
-export function runCompact(pi, runtime, config, ctx, messages, opts = {}) {
-    runtime.bindRepo(ctx.cwd);
-    const sid = normalizeSessionId(ctx.sessionManager.getSessionId());
-    runtime.resetRuntime(sid);
-    runtime.rt.sessionId = sid;
-    const view = runtime.engineView(messages);
-    // keepFrom deepens with context pressure (Fix E): under high pressure we
-    // compact more of the session, down to the preserveRecentMin floor.
-    const preserve = preserveRecentForPressure(opts.compressionPressure ?? 0, config.preserveRecent, config.preserveRecentMin);
-    const keepFrom = opts.keepFrom ?? Math.max(0, view.length - preserve);
-    // For very small sessions (fewer messages than preserveRecent), allow
-    // compacting everything except the last message — the user explicitly
-    // requested compaction, so don't refuse it just because the session is short.
-    if (keepFrom <= 0) {
-        if (view.length <= 1)
-            return { skipped: true };
-        // Use the fallback: compact everything except the last message
-        const fallbackKeepFrom = view.length - 1;
-        return doCompact(view, fallbackKeepFrom, opts, sid, config, pi, ctx, runtime);
-    }
-    return doCompact(view, keepFrom, opts, sid, config, pi, ctx, runtime);
-}
-function doCompact(view, keepFrom, opts, sid, config, pi, ctx, runtime) {
-    runtime.pulsing = true; // animate the status line while the (sync) pipeline runs
-    runtime.setEffect?.("pulse", "accent", 1500); // v0.8.3: ambient border pulse during compaction
-    // S21.2: reset the per-compaction memory-op counter so the post-compact
-    // consolidate pass only fires when memory rows actually changed during the
-    // compaction window (turn_end → auto-review may have written some).
-    runtime.memoriesTouchedThisCompaction = 0;
-    const result = compactSession({
-        sessionId: sid,
-        messages: view,
-        keepFrom,
-        summary: opts.summary,
-        timestamp: Date.now(),
-        onTier: runtime.makeTierCallback(ctx),
-        compressionPressure: opts.compressionPressure,
-    }, runtime.store);
-    runtime.pulsing = false;
-    if (result.skipped)
-        return { skipped: true };
-    if (!result.deduped) {
-        runtime.rt.persistedThisSession = true;
-        runtime.rt.lastCheckpointId = result.checkpointId;
-    }
-    runtime.rt.lastCompactedFrom = result.compactedFrom;
-    runtime.rt.lastCompactedTokens = result.tokenEstimate;
-    runtime.rt.dedupAttempts++;
-    // Honest "tokens saved" for this session-instance only:
-    //   new checkpoint      → original − stored
-    //   deduped onto existing → whole original region (nothing new stored)
-    // Resets to 0 on session_start (rt is rebuilt) — so a fresh session shows 0
-    // while the repo's cumulative saved (SQLite meta) keeps the running total.
-    const saved = result.deduped
-        ? result.originalTokenEstimate
-        : Math.max(0, result.originalTokenEstimate - result.tokenEstimate);
-    runtime.rt.tokensSaved += saved;
-    runtime.rt.compactCount += 1;
-    incCompactCount(runtime.currentStateDir);
-    if (result.deduped) {
-        runtime.rt.cacheHitTokens += saved;
-        incCacheHitTokens(saved, runtime.currentStateDir);
-    }
-    runtime.rt.lastCompactAt = Date.now();
-    if (result.deduped)
-        runtime.rt.dedupSkips++;
-    // Grow the rolling "saved" goal so the progress bar always has a fresh
-    // denominator (we don't want it pinned at 100% once we pass an old target).
-    if (runtime.rt.tokensSaved > runtime.savedGoal)
-        runtime.savedGoal = Math.ceil((runtime.rt.tokensSaved * 1.25) / 10_000) * 10_000;
-    // Live toolbar activity: what file/region just got compacted or deduped.
-    // Rendered via the rotating ticker line (see snapshot); the ring buffer is
-    // cycled one-per-repaint so the single line scrolls through recent files.
-    const files = result.filesModified ?? [];
-    const fileLabel = files.length
-        ? files.map((f) => f.split("/").pop() ?? f).slice(0, 2).join(", ")
-        : result.regionHash.slice(0, 8);
-    runtime.lastActivityAt = Date.now();
-    // Explain-why line: surfaced while fresh. Pulls the dedup reason (which for
-    // L2 includes the cosine sim) so the user sees WHY a region was kept/dropped.
-    runtime.lastWhy = result.deduped
-        ? `why: deduped@${result.dedupReason ?? "tier"}`
-        : `why: compacted → ${result.checkpointId}`;
-    // Recall/activity ticker: record this event in the ring buffer.
-    const savedK = (saved / 1000).toFixed(1);
-    runtime.pushTicker(result.deduped
-        ? `${C.green}♻${C.reset} deduped ${fileLabel} · ${savedK}k saved`
-        : `${C.cyan}🗜${C.reset} ${result.checkpointId} · +${savedK}k · ${fileLabel}`);
-    // The per-tier trace has settled into the final outcome — fold it back into
-    // the activity line and stop showing the live trace.
-    runtime.tierTrace = undefined;
-    // Record session activity + a daily-log entry in the per-repo SQLite store
-    // (foundation for resume-sessions / daily-log features). Best-effort — never
-    // block a compaction on bookkeeping.
-    try {
-        const root = resolveRepoRoot(ctx.cwd);
-        touchSession(sid, root, runtime.currentStateDir);
-        logDaily(sid, "compact", result.checkpointId, saved, runtime.currentStateDir);
-    }
-    catch {
-        /* non-fatal: stats bookkeeping only */
-    }
-    // S21.2: best-effort consolidation of near-duplicate memories for this repo.
-    // Runs after the per-repo stats touch so `consolidateMemories` can use the
-    // same stateDir. Non-fatal — a failed consolidate never blocks a compaction.
-    // Only runs when new memory ops landed in this pass (otherwise the prior
-    // compaction's consolidate already had its shot — re-running would just
-    // touch every row again with no merges).
-    if (!result.deduped && runtime.memoriesTouchedThisCompaction > 0) {
-        try {
-            const root = resolveRepoRoot(ctx.cwd);
-            void consolidateMemories(runtime.currentStateDir, root).then((n) => {
-                if (n > 0)
-                    runtime.pushTicker(`${C.green}∫${C.reset} consolidated ${n} memory dup${n === 1 ? "" : "s"}`);
-            }, () => {
-                /* swallow: consolidate failures must never surface to the user */
-            });
-        }
-        catch {
-            /* non-fatal */
-        }
-    }
-    // S24 review-on-compact: when pressure is high, the just-compacted region is
-    // exactly the context worth remembering, so review it immediately rather than
-    // waiting for the next turn-cadence tick. Uses the shared runMemoryReview
-    // helper (fire-and-forget; doCompact is sync). Best-effort + non-fatal. Only
-    // fires above the `high` band so low-pressure compactions don't pay the cost.
-    if (!result.deduped && config.memoryAutoReview && runtime.pressureBand !== "low" && runtime.pressureBand !== "medium") {
-        void runMemoryReview(runtime, view, "pressure");
-    }
-    // Sentinel marker: a non-LLM bookkeeping entry so subsequent triggers can
-    // skip re-vectorizing an already-compacted region (zero token cost).
-    // v0.8.6: gate on !result.deduped so the marker ONLY lands when a genuinely
-    // new checkpoint was created. Without this, every dedup re-fire appended a
-    // fresh sentinel to the real transcript, bloating it and perturbing the
-    // provider KV-cache prefix (the alternating cache-miss regression). Matches
-    // the RAPTOR + vector-index blocks above, which are already !deduped-gated.
-    if (!result.deduped) {
-        pi.appendEntry(MARKER_TYPE, {
-            checkpointId: result.checkpointId,
-            regionHash: result.regionHash,
-            tokenEstimate: result.tokenEstimate,
-            deduped: result.deduped,
-        });
-    }
-    // Fix D: refresh the RAPTOR tree for this session so live recall (search) can
-    // serve high-level summaries. Best-effort + non-fatal: never block compaction.
-    // Budget-guarded (RAPTOR_BUDGET_MS) so it can't hang a large session.
-    if (config.raptorEnabled && !result.deduped) {
-        try {
-            const dd = loadDedupConfig();
-            const all = vectorList(runtime.store, sid);
-            const leaves = all.map((cp) => ({
-                id: cp.checkpointId,
-                messages: [],
-                sourceText: cp.normalizedText ?? cp.summary ?? cp.regionHash,
-                embedding: cp.embedding,
-            }));
-            if (leaves.length >= 2) {
-                // S42D: skip the rebuild when the last build is fresh (within
-                // RAPTOR_FRESHNESS_HOURS) and the checkpoint count hasn't drifted by
-                // more than 20%. avoids re-clustering on every compaction when the
-                // tree is still representative. 0 disables (always rebuild).
-                if (dd.RAPTOR_FRESHNESS_HOURS > 0 &&
-                    isRaptorTreeFresh(sid, runtime.currentStateDir, dd.RAPTOR_FRESHNESS_HOURS, all.length)) {
-                    runtime.logger?.info("raptor_skip_fresh", { sessionId: sid });
-                }
-                else {
-                    // S25: stamp the tree with the newest checkpoint epoch so the
-                    // freshness guard in raptorSearchHits can reject stale trees after a
-                    // later compaction adds newer checkpoints.
-                    const builtAt = all.length > 0 ? Math.max(...all.map((c) => c.timestamp)) : Date.now();
-                    runRaptor(leaves, {
-                        stateDir: runtime.currentStateDir,
-                        sessionId: sid,
-                        budgetMs: dd.RAPTOR_BUDGET_MS,
-                        clustersPerLevel: dd.RAPTOR_CLUSTERS_PER_LEVEL,
-                        consistencyThreshold: dd.RAPTOR_CONSISTENCY,
-                        logger: runtime.logger,
-                        builtAt: Number.isFinite(builtAt) ? builtAt : Date.now(),
-                    });
-                }
-            }
-        }
-        catch {
-            /* non-fatal: tree refresh never blocks a compaction */
-        }
-    }
-    // Slice 2: best-effort mirror of the new checkpoint into the async global
-    // PGlite/HNSW vector index. Fires once per compaction (not per-add), so the
-    // shared global dir is never hammered by concurrent test workers.
-    // Non-fatal: a WASM init failure degrades to the sync scan silently.
-    if (!result.deduped) {
-        try {
-            const all = vectorList(runtime.store, sid);
-            const latest = all.find((cp) => cp.checkpointId === result.checkpointId);
-            if (latest?.embedding) {
-                void indexUpsertEmbedding(repoKey(runtime.currentStateDir), sid, latest.checkpointId, latest.embedding).catch(() => {
-                    /* non-fatal: index refresh never blocks a compaction */
-                });
-            }
-        }
-        catch {
-            /* non-fatal: index refresh never blocks a compaction */
-        }
-    }
-    runtime.setStatus(ctx, runtime.rt.persistedThisSession
-        ? `mega-compact: ${result.checkpointId} · ${saved} tok saved`
-        : `mega-compact: ready`);
-    runtime.logger.info("compact", {
-        sessionId: sid,
-        checkpointId: result.checkpointId ?? "(deduped)",
-        deduped: result.deduped,
-        tokenEstimate: saved,
-        compactedFrom: result.compactedFrom,
-    });
-    runtime.dashboard.event("compact", {
-        sessionId: sid,
-        checkpointId: result.checkpointId ?? "(deduped)",
-        deduped: result.deduped,
-        tokenEstimate: saved,
-        compactedFrom: result.compactedFrom,
-    });
-    runtime.snapshot(ctx);
-    return { skipped: false, result, keepFrom, saved };
-}
-/**
- * Predict whether pi's `ctx.compact()` would throw a no-op error — "Already
- * compacted" or "Nothing to compact (session too small)" — so the auto-trigger
- * can SKIP the call instead of surfacing a hard, user-facing error.
- *
- * Why we can't intercept or suppress it: pi's public `compact()` computes
- * `prepareCompaction()` and throws *before* it emits `session_before_compact`,
- * so our handler there never runs on the no-op path. And `ctx.compact()`'s
- * `onError` callback fires only AFTER pi has already emitted a `compaction_end`
- * event carrying the error message (which the interactive UI renders) — so
- * `onError` cannot mute it either. The only robust fix is to not call
- * `ctx.compact()` when pi would no-op. (pi's own `_runAutoCompaction` path is
- * silent on this same condition; the public path we're forced through is the
- * one that throws.)
- *
- * Skipping is correct, not a compromise: by the time this runs, `runCompact()`
- * has already persisted the recall checkpoint (Path A). The durable on-disk
- * trim is only useful when pi can actually summarize a region; a transcript
- * under pi's `keepRecentTokens` budget is small enough that reloading it on
- * resume isn't a token-growth problem, so the durable trim is unnecessary
- * there anyway.
- *
- * Mirrors pi's `prepareCompaction()` return-undefined conditions (compaction.js):
- *  (1) last entry is a compaction      → "Already compacted"
- *  (2) <2 cut-point messages since the last compaction → nothing to summarize
- *      (a cut point = any non-toolResult message — user/assistant/bash/custom/
- *       branchSummary/compactionSummary — matching pi's isCutPointMessage)
- *  (3) transcript tokens since the last compaction < keepRecentTokens → pi
- *      keeps everything → nothing to summarize
- * `keepRecentTokens` isn't readable from the extension API, so (3) uses the pi
- * default (20000) as a conservative floor; raise it via
- * `MEGACOMPACT_DURABLE_TRIM_FLOOR` if you raise pi's `compact.keepRecentTokens`.
- *
- * Best-effort: on any read error returns true (skip) — skipping a durable trim
- * is always safe; calling `ctx.compact()` on a no-op throws to the user.
+ * compact.ts — delegate-shell for the compaction pipeline (3WF-2 split).
+ *
+ * This file is now a thin barrel: the real bodies live in `./compact/` so the
+ * module stays well under the extensions/ 400-line soft cap. The public API is
+ * UNCHANGED — every existing importer (via `../mega-pipeline.js`, which does
+ * `export * from "./mega-pipeline/compact.js"`) keeps working with ZERO changes
+ * at their call sites. Exports preserved exactly:
+ *   - `RunCompactResult` (type)
+ *   - `runCompact`
+ *   - `piCompactWouldNoop`
  */
-export function piCompactWouldNoop(ctx) {
-    try {
-        const branch = ctx.sessionManager.getBranch();
-        if (branch.length === 0)
-            return true;
-        // (1) already compacted — pi throws "Already compacted"
-        if (branch[branch.length - 1].type === "compaction")
-            return true;
-        // boundaryStart = index just after the most recent compaction entry (or 0)
-        let boundaryStart = 0;
-        for (let i = branch.length - 1; i >= 0; i--) {
-            if (branch[i].type === "compaction") {
-                boundaryStart = i + 1;
-                break;
-            }
-        }
-        let cutPoints = 0;
-        let tokens = 0;
-        for (let i = boundaryStart; i < branch.length; i++) {
-            const e = branch[i];
-            if (e.type === "compaction")
-                continue;
-            let isCut = false;
-            for (const m of sessionEntryToContextMessages(e)) {
-                // pi's isCutPointMessage: every role except toolResult
-                if (m.role !== "toolResult")
-                    isCut = true;
-                const c = m.content;
-                const text = typeof c === "string" ? c
-                    : Array.isArray(c)
-                        ? c.map((b) => b?.text ?? "").join(" ")
-                        : "";
-                if (text)
-                    tokens += estimateBlockTokens(text);
-            }
-            if (isCut)
-                cutPoints++;
-        }
-        // (2) need >=2 cut points so the kept cut isn't the first message
-        if (cutPoints < 2)
-            return true;
-        // (3) transcript under pi's keepRecentTokens budget → pi keeps everything
-        if (tokens < durableTrimFloorTokens())
-            return true;
-        return false;
-    }
-    catch {
-        return true; // safe: skip the durable trim rather than risk a user-facing throw
-    }
-}
-/** pi's default keepRecentTokens (compaction settings). Override with
- *  MEGACOMPACT_DURABLE_TRIM_FLOOR if you raise pi's compact.keepRecentTokens. */
-function durableTrimFloorTokens() {
-    const raw = process.env.MEGACOMPACT_DURABLE_TRIM_FLOOR;
-    if (raw !== undefined && Number.isFinite(Number(raw)))
-        return Number(raw);
-    return 20_000;
-}
+export { runCompact } from "./compact/run.js";
+export { piCompactWouldNoop } from "./compact/noop.js";

package/dist/extensions/mega-pipeline/recall/impl.js

@@ -0,0 +1,258 @@
+/**
+ * recall/impl.ts — unified Layer-5 recall pipeline implementation (3WF-3 split).
+ *
+ * Behavior is UNCHANGED from the pre-split recall.ts. `doRecall` is the ONE path
+ * that injects (sync). `doRecallAsync` augments with optional cross-repo HNSW
+ * on resume / /mega-recall --cross-repo. Both mutate the shared MegaRuntime
+ * (token accounting, ticker, dashboard events). The shell recall.ts re-exports
+ * these names so `export * from "./mega-pipeline/recall.js"` stays byte-stable.
+ */
+import { sessionEntryToContextMessages } from "@earendil-works/pi-coding-agent";
+import { recallAndInline, recallAndInlineAsync, formatRecallBlock, } from "../../../src/recall.js";
+import { normalizeSessionId } from "../../../src/store.js";
+import { incRecallInjected, incCacheHitTokens, getIndexDir, } from "../../../src/store/sqlite.js";
+import { ensureConversationIdFor, recordTurnWrite, recordRecallWrite, } from "../../mega-turn-store.js";
+import { C } from "../../mega-runtime.js";
+import { recordRecallLatency } from "../../mega-runtime/vc-observer.js";
+/**
+ * Unified recall (Layer 5). The ONE path that injects. Returns the recall
+ * result; callers decide whether to stage it for before_agent_start (resume)
+ * or report it (command).
+ */
+export function doRecall(runtime, config, ctx, query, source) {
+    runtime.bindRepo(ctx.cwd);
+    const sid = normalizeSessionId(ctx.sessionManager.getSessionId());
+    // Live window text for inline dedupe (Fix C): drop recalled checkpoints that
+    // are already resident in the session, so recall never re-injects context the
+    // model can already see. Best-effort — an empty window just skips dedupe.
+    const liveWindow = config.windowDedupe ? extractLiveWindow(ctx) : undefined;
+    const recallStartMs = Date.now();
+    const result = recallAndInline({
+        sessionId: sid,
+        query,
+        limit: config.autoInlineK,
+        source,
+        skipInjected: true,
+        recallMaxTokens: config.recallMaxTokens,
+        windowDedupe: config.windowDedupe,
+        liveWindow,
+        dedupSim: config.dedupSim,
+    }, runtime.store);
+    runtime.dashboard.event("recall", {
+        source,
+        query: query.slice(0, 120),
+        injected: result.toInject.length,
+        empty: result.empty,
+    });
+    if (config.ragRecallMetrics && result.hydeInfo) {
+        runtime.dashboard.event("hyde_executed", {
+            sessionId: sid,
+            ran: result.hydeInfo.ran,
+            skipped: result.hydeInfo.skipped,
+            reason: result.hydeInfo.reason,
+            hypotheticalDoc: result.hydeInfo.hypotheticalDoc.slice(0, 400),
+            generationMs: result.hydeInfo.generationMs,
+            rawHitCount: result.hydeInfo.rawHitCount,
+            hydeHitCount: result.hydeInfo.hydeHitCount,
+            fusedHitCount: result.hydeInfo.fusedHitCount,
+            lift: result.hydeInfo.lift,
+        });
+    }
+    if (config.ragRecallMetrics && result.recallMetrics) {
+        runtime.dashboard.event("recall_metrics", {
+            sessionId: sid,
+            hitCount: result.recallMetrics.hitCount,
+            score: result.recallMetrics.score,
+            pass: result.recallMetrics.pass,
+            relevance: result.recallMetrics.relevance,
+            coverage: result.recallMetrics.coverage,
+            diversity: result.recallMetrics.diversity,
+            specificity: result.recallMetrics.specificity,
+        });
+    }
+    if (!result.empty && result.toInject.length > 0) {
+        const top = result.toInject[0];
+        const scorePct = Math.round((top.score ?? 0) * 100);
+        const files = top.checkpoint.filesModified ?? [];
+        const label = files.length
+            ? files
+                .map((f) => f.split("/").pop() ?? f)
+                .slice(0, 2)
+                .join(", ")
+            : top.checkpoint.checkpointId;
+        runtime.pushTicker(`${C.amber}↩${C.reset} recalled ${top.checkpoint.checkpointId} · ${scorePct}% · ${label}`);
+        runtime.lastWhy = `why: recalled@${scorePct}% (${result.toInject.length} chkpt)`;
+    }
+    let sumTokens = 0;
+    for (const h of result.toInject)
+        sumTokens += h.checkpoint.tokenEstimate;
+    if (result.toInject.length > 0) {
+        runtime.rt.recallInjections += result.toInject.length;
+        runtime.rt.cacheHitTokens += sumTokens;
+        incRecallInjected(result.toInject.length, runtime.currentStateDir);
+        incCacheHitTokens(sumTokens, runtime.currentStateDir);
+    }
+    // S43: record recall provenance — which checkpoints/summaries served this
+    // turn, their score + source path. Linked to the turn row written at
+    // turn_end via the conversation+turnIndex. Best-effort + non-fatal.
+    // Persists telemetry (HyDE + recall metrics) even when recall returned
+    // no hits, so empty-recall HyDE invocations are still visible in the
+    // dashboard Turns/Metrics tabs.
+    const hasTelemetry = result.hydeInfo != null || result.recallMetrics != null;
+    if (result.toInject.length > 0 || hasTelemetry) {
+        try {
+            const convId = ensureConversationIdFor(config, sid, runtime.currentStateDir);
+            const turnId = recordTurnWrite(config, {
+                conversationId: convId,
+                sessionId: sid,
+                turnIndex: runtime.currentTurn,
+                role: "assistant",
+                startedAt: Date.now(),
+                hyde: result.hydeInfo ?? undefined,
+                recallMetrics: result.recallMetrics ?? undefined,
+            }, runtime.currentStateDir);
+            if (result.toInject.length > 0) {
+                recordRecallWrite(config, turnId, result.toInject.map((h) => ({
+                    checkpointId: h.checkpoint.checkpointId,
+                    score: h.score,
+                    source: h.raptorLevel !== undefined
+                        ? "raptor"
+                        : h.repoId
+                            ? "cross-repo"
+                            : "flat",
+                    raptorLevel: h.raptorLevel,
+                })), runtime.currentStateDir);
+            }
+        }
+        catch {
+            /* non-fatal: recall provenance never breaks the recall path */
+        }
+    }
+    // VC0A: record recall latency on the eval observer (mode A) so the dashboard
+    // histogram reflects real data. No-op when the observer is absent (flag off /
+    // construction failure).
+    try {
+        recordRecallLatency(runtime, Date.now() - recallStartMs, sid, 0);
+    }
+    catch {
+        /* non-fatal: latency recording never breaks recall */
+    }
+    return result;
+}
+/**
+ * S17: async recall with optional cross-repo augmentation. Used on resume
+ * (session_start) and /mega-recall --cross-repo — NEVER from the mid-turn
+ * context handler (that stays sync). Runs the sync same-repo scan first; if it
+ * returns < config.autoInlineK hits AND crossRepo is enabled, awaits the PGlite
+ * HNSW cross-repo path and merges (source-labeled, deduped by checkpointId). The
+ * recallMaxTokens cap + windowDedupe apply to the merged set so cross-repo can
+ * never net-inflate the window. Cross-repo uses a stricter cosine floor
+ * (config.crossRepoCosine) than same-repo. Non-fatal: any async failure returns
+ * the same-repo result unchanged.
+ */
+export async function doRecallAsync(runtime, config, ctx, query, source, opts = {}) {
+    runtime.bindRepo(ctx.cwd);
+    const sid = normalizeSessionId(ctx.sessionManager.getSessionId());
+    const liveWindow = config.windowDedupe ? extractLiveWindow(ctx) : undefined;
+    // Sync same-repo first (fast, never blocks).
+    const sameRepo = recallAndInline({
+        sessionId: sid,
+        query,
+        limit: config.autoInlineK,
+        source,
+        skipInjected: true,
+        recallMaxTokens: config.recallMaxTokens,
+        windowDedupe: config.windowDedupe,
+        liveWindow,
+        dedupSim: config.dedupSim,
+    }, runtime.store);
+    if (!config.crossRepoEnabled || !opts.crossRepo)
+        return sameRepo;
+    if (sameRepo.toInject.length >= config.autoInlineK)
+        return sameRepo; // same-repo satisfied
+    // Augment: cross-repo HNSW (async) with the stricter floor. Non-fatal.
+    try {
+        const x = await recallAndInlineAsync({
+            sessionId: sid,
+            query,
+            limit: config.autoInlineK,
+            source,
+            skipInjected: true,
+            recallMaxTokens: config.recallMaxTokens,
+            windowDedupe: config.windowDedupe,
+            liveWindow,
+            dedupSim: config.crossRepoCosine,
+            crossRepo: true,
+            // F2: resolve the machine-wide index dir via the shared resolver so the
+            // cross-repo injected-set dedup works even when MEGACOMPACT_INDEX_DIR is
+            // unset. The env var still wins when set (getIndexDir checks it first);
+            // the default (~/.mega-compact-index) is the same DB mega-commands and the
+            // dashboard read, so injection counts stay consistent. Without this, a
+            // bare `process.env` read returns undefined → cross-repo hits re-inject in
+            // every new session (the global injected-set is never consulted).
+            globalIndexDir: getIndexDir(),
+        }, runtime.store);
+        runtime.dashboard.event("recall-crossrepo", {
+            source,
+            query: query.slice(0, 120),
+            injected: x.toInject.length,
+            sourceRepos: x.toInject.map((h) => h.repoId).filter(Boolean),
+        });
+        // Merge, dedup by checkpointId, respect the same token cap by reformatting.
+        const seen = new Set(sameRepo.toInject.map((h) => h.checkpoint.checkpointId));
+        const merged = [...sameRepo.toInject];
+        for (const h of x.toInject) {
+            if (!seen.has(h.checkpoint.checkpointId)) {
+                merged.push(h);
+                seen.add(h.checkpoint.checkpointId);
+            }
+        }
+        const block = merged.length ? formatRecallBlock(merged) : "";
+        if (merged.length > 0) {
+            let sumTokens = 0;
+            for (const h of merged)
+                sumTokens += h.checkpoint.tokenEstimate;
+            runtime.rt.recallInjections += merged.length;
+            runtime.rt.cacheHitTokens += sumTokens;
+            incRecallInjected(merged.length, runtime.currentStateDir);
+            incCacheHitTokens(sumTokens, runtime.currentStateDir);
+        }
+        return {
+            toInject: merged,
+            report: merged.map((h) => `  • ${h.checkpoint.checkpointId}${h.repoId ? ` (from ${h.repoId.split("/").filter(Boolean).pop()})` : ""}`),
+            block,
+            empty: merged.length === 0,
+            // H1: merged cross-repo result reuses the same-repo pass's telemetry.
+            hydeInfo: sameRepo.hydeInfo,
+            recallMetrics: sameRepo.recallMetrics,
+        };
+    }
+    catch {
+        return sameRepo; // cross-repo failure → same-repo only (non-fatal)
+    }
+}
+/**
+ * Extract the live-window message texts from the session manager (Fix C),
+ * for inline-dedupe of recalled checkpoints. Best-effort: returns [] on any
+ * error so recall falls back to unbounded (still correct, just no dedupe).
+ * Mirrors recentUserQuery's use of sessionEntryToContextMessages.
+ */
+export function extractLiveWindow(ctx) {
+    try {
+        const entries = ctx.sessionManager.getEntries();
+        const texts = [];
+        for (const e of entries) {
+            for (const m of sessionEntryToContextMessages(e)) {
+                const c = m.content;
+                if (typeof c === "string")
+                    texts.push(c);
+                else if (Array.isArray(c))
+                    texts.push(c.map((b) => b.text ?? "").join(" "));
+            }
+        }
+        return texts;
+    }
+    catch {
+        return [];
+    }
+}