pi-mega-compact 0.11.4 → 0.11.6

package/dist/extensions/mega-events/separated-prompt.js

@@ -0,0 +1,248 @@
+/**
+ * extensions/mega-events/separated-prompt.ts — A1+A2 PLAN_V2 Phase 2+3
+ * Message Separation + Cache-Optimized Prompt Builder:
+ *   - buildSeparatedPrompt(): Basic 4-layer message separation (Phase 2).
+ *   - buildCacheOptimizedPrompt(): Insert Layer 2 (vector-optimized stable
+ *     context) between summary (Layer 1) and thread (Layer 3), ordered by
+ *     stripe stability DESC (Phase 3).
+ *   - detectTopicShift(): Cosine-similarity-based topic-shift detection.
+ *   - cosineSimilarity(): cosine similarity of two Float32Array embeddings.
+ *
+ * Layer order (Phase 3): 0 (system) -> 1 (summary) -> 2 (cache stripes) ->
+ * 3 (thread: user/assistant turns) -> 4 (tool results at tail).
+ *
+ * Feature-gated by MEGACOMPACT_CACHE_STRIPING (default OFF).
+ * MEGACOMPACT_MESSAGE_SEPARATION must also be ON for any layering to occur.
+ * Flag-OFF = byte-identical to pre-sprint — returns messages unchanged.
+ */
+import { openStore } from "../../src/store/sqlite/utils.js";
+import { getStateDir } from "../../src/store.js";
+// ─── Constants ───────────────────────────────────────────────────────────────
+/** Default stripe limit: top 10 most stable chunks. */
+const DEFAULT_STRIPE_LIMIT = 10;
+/** Topic-shift threshold: similarity below this triggers a stripe refresh. */
+const TOPIC_SHIFT_THRESHOLD = 0.7;
+// ─── Build separated prompt (Phase 2) ────────────────────────────────────────
+/**
+ * Reorganize messages into stable layers for better prompt caching.
+ *
+ * Layer order: 0 (system) -> 1 (summary) -> 3 (thread: user/assistant turns) ->
+ * 4 (tool results at tail).
+ *
+ * Feature gate: MEGACOMPACT_MESSAGE_SEPARATION env var.  When OFF (default),
+ * returns `messages` unchanged (byte-identical to pre-sprint).
+ */
+export function buildSeparatedPrompt(messages, _opts) {
+    const flag = process.env.MEGACOMPACT_MESSAGE_SEPARATION;
+    if (flag === "0" || flag === "false" || flag === undefined || flag === "") {
+        return messages;
+    }
+    // pi's AgentMessage union has no "system" role — the system prompt lives in
+    // AgentState.systemPrompt, separate from this array. The cache-relevant,
+    // low-risk transformation is moving volatile tool results/executions to the
+    // tail so the stable prefix (user/assistant/summaries/custom) stays contiguous.
+    // Discriminate by role only — never reach into `.content`/`.tool_calls`,
+    // which are variant-specific and require narrowing.
+    const main = [];
+    const tail = [];
+    for (const m of messages) {
+        if (m.role === "toolResult" || m.role === "bashExecution") {
+            tail.push(m);
+        }
+        else {
+            main.push(m);
+        }
+    }
+    if (tail.length === 0)
+        return messages; // nothing to reorder — byte-identical
+    return [...main, ...tail];
+}
+// ─── Cosine similarity ───────────────────────────────────────────────────────
+/**
+ * Compute cosine similarity between two Float32Array embeddings.
+ * Returns a value in [0, 1] where 1 = identical direction.
+ * Returns 0 if either vector is zero or inputs are mismatched.
+ */
+export function cosineSimilarity(a, b) {
+    if (a.length !== b.length || a.length === 0)
+        return 0;
+    let dot = 0;
+    let normA = 0;
+    let normB = 0;
+    for (let i = 0; i < a.length; i++) {
+        dot += a[i] * b[i];
+        normA += a[i] * a[i];
+        normB += b[i] * b[i];
+    }
+    const denom = Math.sqrt(normA) * Math.sqrt(normB);
+    if (denom === 0)
+        return 0;
+    return dot / denom;
+}
+// ─── Buffer ↔ Float32Array helpers ───────────────────────────────────────────
+/**
+ * Decode a Float32Array from a Buffer (little-endian float32 blob).
+ * Returns null on empty/undefined input.
+ */
+export function decodeEmbeddingBlob(buf) {
+    if (!buf || buf.length === 0)
+        return null;
+    return new Float32Array(buf.buffer, buf.byteOffset, buf.byteLength / 4);
+}
+/**
+ * Encode a Float32Array into a Buffer for DB storage.
+ */
+export function encodeEmbeddingBlob(vec) {
+    return Buffer.from(vec.buffer, vec.byteOffset, vec.byteLength);
+}
+// ─── Topic-shift detection ───────────────────────────────────────────────────
+/**
+ * Detect whether the conversation topic has shifted between the current turn
+ * and the previous turn. Uses cosine similarity between topic embeddings.
+ *
+ * Returns `true` when similarity < TOPIC_SHIFT_THRESHOLD (0.7), meaning the
+ * stable cache prefix may need reordering.
+ *
+ * Returns `false` (no shift) when either embedding is missing.
+ */
+export function detectTopicShift(currentEmb, previousEmb) {
+    if (!currentEmb || !previousEmb)
+        return false;
+    return cosineSimilarity(currentEmb, previousEmb) < TOPIC_SHIFT_THRESHOLD;
+}
+// ─── Load embeddings from store ──────────────────────────────────────────────
+/**
+ * Load the topic embedding for a given turn from the embedding_cache table.
+ * Returns null if no embedding is stored for that turn.
+ */
+export function loadTopicEmbedding(stateDir, conversationId, turnIndex) {
+    const db = openStore(stateDir);
+    try {
+        const row = db.prepare(`SELECT embedding FROM embedding_cache WHERE content_hash = ?`).get(`topic:${conversationId}:${turnIndex}`);
+        if (!row)
+            return null;
+        return decodeEmbeddingBlob(row.embedding);
+    }
+    finally {
+        db.close();
+    }
+}
+/**
+ * Store a topic embedding for a given turn in the embedding_cache table.
+ * Idempotent (INSERT OR REPLACE).
+ */
+export function storeTopicEmbedding(stateDir, conversationId, turnIndex, embedding) {
+    const db = openStore(stateDir);
+    try {
+        const contentHash = `topic:${conversationId}:${turnIndex}`;
+        const blob = encodeEmbeddingBlob(embedding);
+        db.prepare(`INSERT OR REPLACE INTO embedding_cache (content_hash, embedding, computed_at) VALUES (?, ?, ?)`).run(contentHash, blob, Date.now());
+    }
+    finally {
+        db.close();
+    }
+}
+// ─── Cache stripe management ─────────────────────────────────────────────────
+/**
+ * Refresh cache stripe assignments from the cache_stripes table for a given
+ * conversation/epoch. Returns rows ordered by stability DESC.
+ *
+ * This is called at the start of each turn (or on topic-shift) to ensure the
+ * stripe assignments reflect the latest stability scores.
+ */
+export function refreshStripeAssignments(stateDir, epochId, limit = DEFAULT_STRIPE_LIMIT) {
+    const db = openStore(stateDir);
+    try {
+        const rows = db.prepare(`SELECT chunk_id, stripe, stability, assigned_at, epoch_id
+       FROM cache_stripes
+       WHERE epoch_id = ?
+       ORDER BY stability DESC
+       LIMIT ?`).all(epochId, limit);
+        return rows;
+    }
+    finally {
+        db.close();
+    }
+}
+// ─── Build cache-optimized prompt (Phase 3) ──────────────────────────────────
+/**
+ * Build a cache-optimized prompt with 5 layers:
+ *   0 (system) -> 1 (summary) -> 2 (cache stripes) -> 3 (thread) -> 4 (tool)
+ *
+ * Feature gate: MEGACOMPACT_CACHE_STRIPING (default OFF). When OFF, delegates
+ * to buildSeparatedPrompt (byte-identical to Phase 2 behavior).
+ *
+ * When ON but MEGACOMPACT_MESSAGE_SEPARATION is OFF, also delegates to
+ * buildSeparatedPrompt (which returns messages unchanged).
+ */
+export function buildCacheOptimizedPrompt(messages, opts) {
+    const flag = process.env.MEGACOMPACT_CACHE_STRIPING;
+    if (flag === "0" || flag === "false" || flag === undefined || flag === "") {
+        // Flag OFF: delegate to buildSeparatedPrompt (byte-identical).
+        return buildSeparatedPrompt(messages, opts);
+    }
+    // Build the base 4-layer structure first.
+    const base = buildSeparatedPrompt(messages, opts);
+    // If base equals messages, separation is OFF — return unchanged.
+    if (base === messages)
+        return base;
+    const stateDir = opts?.stateDir ?? getStateDir();
+    // Layer 1 = leading branch/compaction summaries (the stable summary layer).
+    // There is no "system" role in pi's AgentMessage union, so Layer 0 is empty.
+    // Count leading summaries to find where to insert the cache-stripe layer.
+    let layer1End = 0;
+    for (let i = 0; i < base.length; i++) {
+        const r = base[i].role;
+        if (r === "branchSummary" || r === "compactionSummary") {
+            layer1End = i + 1;
+        }
+        else {
+            break;
+        }
+    }
+    // Lay out the result: Layer 1 (summaries) first.
+    const result = base.slice(0, layer1End);
+    // Insert Layer 2: cache stripes (stable context ordered by stability DESC).
+    // Try to find the most recent epoch.
+    const db = openStore(stateDir);
+    let epochId = null;
+    try {
+        const epoch = db.prepare(`SELECT epoch_id FROM checkpoint_epochs ORDER BY created_at DESC LIMIT 1`).get();
+        epochId = epoch?.epoch_id ?? null;
+    }
+    finally {
+        db.close();
+    }
+    if (epochId) {
+        const stripes = refreshStripeAssignments(stateDir, epochId);
+        if (stripes.length > 0) {
+            // Build a single user message with concatenated stripe content.
+            const stripeBlocks = [];
+            for (const s of stripes) {
+                // Load chunk content from context_chunks table.
+                const db2 = openStore(stateDir);
+                try {
+                    const chunk = db2.prepare(`SELECT normalized_text FROM context_chunks WHERE id = ?`).get(s.chunk_id);
+                    if (chunk?.normalized_text) {
+                        stripeBlocks.push(chunk.normalized_text);
+                    }
+                }
+                finally {
+                    db2.close();
+                }
+            }
+            if (stripeBlocks.length > 0) {
+                result.push({
+                    role: "user",
+                    content: stripeBlocks.join("\n\n---\n\n"),
+                    timestamp: Date.now(),
+                });
+            }
+        }
+    }
+    // Append remaining layers (3: thread, 4: tool results).
+    for (let i = layer1End; i < base.length; i++) {
+        result.push(base[i]);
+    }
+    return result;
+}

package/dist/extensions/mega-events/session-handlers.js

@@ -149,14 +149,21 @@ export function registerSessionHandlers(pi, runtime, config) {
         catch {
             /* non-fatal: intent polling never breaks the agent loop */
         }
-        const cpBlock = runtime.pendingRecallBlock;
-        const memBlock = runtime.pendingMemoryRecallBlock;
-        if (!cpBlock && !memBlock)
-            return;
-        runtime.pendingRecallBlock = undefined;
-        runtime.pendingMemoryRecallBlock = undefined;
-        const composed = [cpBlock, memBlock].filter(Boolean).join("\n\n");
-        return { systemPrompt: `${event.systemPrompt}\n\n${composed}` };
+        // S53: when recallTailInject is ON, the staged blocks are injected as a
+        // user-role tail message by the context handler (recall-tail.ts), NOT
+        // here. This prepend path is the legacy behavior (flag OFF = byte-
+        // identical pre-sprint). captureModel + rewind-intent consumption above
+        // stay unconditional regardless of the flag.
+        if (!config.recallTailInject) {
+            const cpBlock = runtime.pendingRecallBlock;
+            const memBlock = runtime.pendingMemoryRecallBlock;
+            if (!cpBlock && !memBlock)
+                return;
+            runtime.pendingRecallBlock = undefined;
+            runtime.pendingMemoryRecallBlock = undefined;
+            const composed = [cpBlock, memBlock].filter(Boolean).join("\n\n");
+            return { systemPrompt: `${event.systemPrompt}\n\n${composed}` };
+        }
     });
     pi.on("session_shutdown", async (_event, ctx) => {
         runtime.setStatus(ctx, undefined);

package/dist/extensions/mega-runtime/env-loader.js

@@ -0,0 +1,56 @@
+/**
+ * mega-runtime/env-loader.ts — load `.mega-compact.env` from the state dir.
+ *
+ * The /mega-setup wizard and the dashboard Setup tab write a `.mega-compact.env`
+ * file with the user's chosen embedder configuration. Without a loader, that
+ * file is write-only dead weight — config is read from process.env at startup
+ * (mega-config.ts:loadConfig), so the file was never applied.
+ *
+ * This loader runs BEFORE loadConfig in the extension entry. It parses the
+ * file (KEY=VALUE lines, # comments, quoted values) and applies each key to
+ * process.env ONLY if it is not already set — so shell profile / inline env
+ * vars always win over the file (the user can override without editing it).
+ *
+ * Non-fatal: a missing or malformed file never breaks the extension.
+ *
+ * PREVENT-PI-004: reads a local file only — no network.
+ * PREVENT-001: no JSON here; line parser is null/empty-safe.
+ */
+import { readFileSync, existsSync } from "node:fs";
+import { join } from "node:path";
+/**
+ * Load `.mega-compact.env` from `stateDir` into `process.env`.
+ * Keys already set in the environment win (file does not override shell).
+ */
+export function loadMegaEnv(stateDir) {
+    const envPath = join(stateDir, ".mega-compact.env");
+    if (!existsSync(envPath))
+        return;
+    let content;
+    try {
+        content = readFileSync(envPath, "utf-8");
+    }
+    catch {
+        return; // unreadable — non-fatal
+    }
+    for (const line of content.split("\n")) {
+        const trimmed = line.trim();
+        if (!trimmed || trimmed.startsWith("#"))
+            continue;
+        const eq = trimmed.indexOf("=");
+        if (eq < 0)
+            continue;
+        const key = trimmed.slice(0, eq).trim();
+        if (!key)
+            continue;
+        let val = trimmed.slice(eq + 1).trim();
+        if ((val.startsWith('"') && val.endsWith('"')) ||
+            (val.startsWith("'") && val.endsWith("'"))) {
+            val = val.slice(1, -1);
+        }
+        // Shell / inline env vars win over the file.
+        if (process.env[key] === undefined) {
+            process.env[key] = val;
+        }
+    }
+}

package/dist/extensions/mega-runtime/reset-runtime.js

@@ -24,6 +24,10 @@ export function resetRuntimeImpl(self, sessionId) {
         dedupAttempts: 0,
         tokensSaved: 0,
         lastCompactAt: null,
+        lastRecallAt: null,
+        lastInjectAt: null,
+        _prevCacheHitPct: null,
+        _lastCacheHealthScore: undefined,
         lastNativeCompactAt: null,
         compactCount: 0,
         recallInjections: 0,
@@ -40,7 +44,9 @@ export function resetRuntimeImpl(self, sessionId) {
         poisonedAdviseSent: false,
         providerOutageAdvised: false,
         poisonedCompactSignatures: new Set(),
+        recallInjectedThisTurn: false,
         poisonedCount: 0,
+        extensionInitiatedTurn: false,
     };
     self.trimCache = null; // v0.8.6: never replay a stale trim into a new session
     self.statusKey = undefined;

package/dist/extensions/mega-runtime/runtime-snapshot.js

@@ -176,6 +176,7 @@ export function snapshotImpl(self, ctx) {
             armed,
             modelSnap,
             providerCachePct,
+            perTurnCacheHitPct: self.rt._prevCacheHitPct ?? undefined,
         });
         self.widgetData = _snapResult.widgetData;
         // S33: consume the flare after copying it into widgetData so it fires

package/dist/extensions/mega-runtime/runtime.js

@@ -49,6 +49,10 @@ export class MegaRuntime {
         dedupAttempts: 0,
         tokensSaved: 0,
         lastCompactAt: null,
+        lastRecallAt: null,
+        lastInjectAt: null,
+        _prevCacheHitPct: null,
+        _lastCacheHealthScore: undefined,
         lastNativeCompactAt: null,
         compactCount: 0,
         recallInjections: 0,
@@ -65,7 +69,10 @@ export class MegaRuntime {
         poisonedAdviseSent: false,
         providerOutageAdvised: false,
         poisonedCompactSignatures: new Set(),
+        // S53: tracks whether withRecallTail injected this turn (consumed at turn_end).
+        recallInjectedThisTurn: false,
         poisonedCount: 0,
+        extensionInitiatedTurn: false,
     };
     // v0.8.6 cache-stability: the cached live-trim view for the current
     // compaction epoch. Set after a fresh runCompact + computeLiveTrimCut, and

package/dist/extensions/mega-runtime/snapshot.js

@@ -131,6 +131,7 @@ export function computeMegaSnapshot(p) {
         cachePct,
         megaCacheFlare: p.megaCacheFlare,
         megaCacheFlarePct: p.megaCacheFlarePct,
+        perTurnCacheHitPct: p.perTurnCacheHitPct ?? cachePct,
         levelUpFlare: p.levelUpFlare,
         achievementFlare: p.achievementFlare,
         achievementFlareTitles: p.achievementFlareTitles,

package/dist/extensions/mega-runtime/widget.js

@@ -87,6 +87,11 @@ export function buildWidgetLines(wd, width, activeAgents) {
         `${C.gray}mem${C.reset} ${wd.embedderName} · ${wd.chk} chunks · ${C.blue}comp ${wd.compStr}${C.reset}`,
         `${C.gray}drift${C.reset} ${wd.driftStatus === "ok" ? C.green : C.amber}${wd.driftStatus}${C.reset}`,
         `${C.gray}compact${C.reset} ${sinceCompactStr(wd.sinceCompact)}`,
+        ...(wd.perTurnCacheHitPct != null
+            ? [
+                `${C.gray}cache hit${C.reset} this turn ${C.green}${Math.round(wd.perTurnCacheHitPct)}%${C.reset}`,
+            ]
+            : []),
     ].join(sep);
     // Wrap to terminal width and pad each line
     const wrapped = wrapLine(content, width - 2, panelBg); // 2-char indent