opencode-lore 0.4.2 → 0.4.4

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "opencode-lore",
-  "version": "0.4.2",
+  "version": "0.4.4",
   "type": "module",
   "license": "MIT",
   "description": "Three-tier memory architecture for OpenCode — distillation, not summarization",

package/src/db.ts

@@ -2,7 +2,7 @@ import { Database } from "bun:sqlite";
 import { join } from "path";
 import { mkdirSync } from "fs";
 
-const SCHEMA_VERSION = 3;
+const SCHEMA_VERSION = 4;
 
 const MIGRATIONS: string[] = [
   `
@@ -130,6 +130,17 @@ const MIGRATIONS: string[] = [
   -- VACUUM must run outside a transaction and cannot be in a multi-statement
   -- exec, so it is handled specially in the migrate() function.
   `,
+  `
+  -- Version 4: Persistent session state for error recovery.
+  -- Stores forceMinLayer so it survives OpenCode restarts. Without this,
+  -- a "prompt too long" error recovery (escalate to layer 2) is lost if
+  -- the process restarts before the next turn.
+  CREATE TABLE IF NOT EXISTS session_state (
+    session_id TEXT PRIMARY KEY,
+    force_min_layer INTEGER NOT NULL DEFAULT 0,
+    updated_at INTEGER NOT NULL
+  );
+  `,
 ];
 
 function dataDir() {
@@ -229,3 +240,35 @@ export function isFirstRun(): boolean {
     .get() as { count: number };
   return row.count === 0;
 }
+
+// ---------------------------------------------------------------------------
+// Persistent session state (error recovery)
+// ---------------------------------------------------------------------------
+
+/**
+ * Load persisted forceMinLayer for a session. Returns 0 if none stored.
+ */
+export function loadForceMinLayer(sessionID: string): number {
+  const row = db()
+    .query("SELECT force_min_layer FROM session_state WHERE session_id = ?")
+    .get(sessionID) as { force_min_layer: number } | null;
+  return row?.force_min_layer ?? 0;
+}
+
+/**
+ * Persist forceMinLayer for a session. Deletes the row when layer is 0
+ * (consumed) to avoid unbounded growth.
+ */
+export function saveForceMinLayer(sessionID: string, layer: number): void {
+  if (layer === 0) {
+    db()
+      .query("DELETE FROM session_state WHERE session_id = ?")
+      .run(sessionID);
+  } else {
+    db()
+      .query(
+        "INSERT OR REPLACE INTO session_state (session_id, force_min_layer, updated_at) VALUES (?, ?, ?)",
+      )
+      .run(sessionID, layer, Date.now());
+  }
+}

package/src/gradient.ts

@@ -1,5 +1,5 @@
 import type { Message, Part } from "@opencode-ai/sdk";
-import { db, ensureProject } from "./db";
+import { db, ensureProject, loadForceMinLayer, saveForceMinLayer } from "./db";
 import { config } from "./config";
 import { formatDistillations } from "./prompt";
 import { normalize } from "./markdown";
@@ -53,9 +53,12 @@ let calibratedOverhead: number | null = null;
 // lore-curator) from corrupting the main session's sticky-layer guard and
 // delta-estimation state when their transform() calls return layer 0.
 //
-// DB persistence is unnecessary: UNCALIBRATED_SAFETY=1.5 safely handles
-// the first turn of a resumed session. The Map is bounded — there are never
-// more than a handful of active sessions at once.
+// forceMinLayer is the one field that MUST survive process restarts: when the
+// API returns "prompt is too long", the error handler sets forceMinLayer=2.
+// If OpenCode restarts before the next turn, the escalation is lost and the
+// overflow repeats. forceMinLayer is persisted to SQLite (session_state table)
+// and loaded on first access. All other state rebuilds from the first API
+// response via UNCALIBRATED_SAFETY.
 // ---------------------------------------------------------------------------
 
 type SessionState = {
@@ -102,6 +105,11 @@ function getSessionState(sessionID: string): SessionState {
   let state = sessionStates.get(sessionID);
   if (!state) {
     state = makeSessionState();
+    // Restore persisted forceMinLayer from DB — survives process restarts.
+    // Critical for "prompt too long" recovery: the error handler sets
+    // forceMinLayer=2, but if OpenCode restarts before the next turn,
+    // the in-memory escalation would be lost without this.
+    state.forceMinLayer = loadForceMinLayer(sessionID) as SafetyLayer;
     sessionStates.set(sessionID, state);
   }
   return state;
@@ -213,10 +221,12 @@ export function getLastLayer(sessionID?: string): SafetyLayer {
 export function setForceMinLayer(layer: SafetyLayer, sessionID?: string) {
   if (sessionID) {
     getSessionState(sessionID).forceMinLayer = layer;
+    saveForceMinLayer(sessionID, layer);
   } else {
     // Fallback for tests / callers without session ID: set on all active sessions
-    for (const state of sessionStates.values()) {
+    for (const [sid, state] of sessionStates.entries()) {
       state.forceMinLayer = layer;
+      saveForceMinLayer(sid, layer);
     }
   }
 }
@@ -225,8 +235,12 @@ export function setForceMinLayer(layer: SafetyLayer, sessionID?: string) {
 export function resetCalibration(sessionID?: string) {
   calibratedOverhead = null;
   if (sessionID) {
+    saveForceMinLayer(sessionID, 0); // clear persisted state
     sessionStates.delete(sessionID);
   } else {
+    for (const sid of sessionStates.keys()) {
+      saveForceMinLayer(sid, 0);
+    }
     sessionStates.clear();
   }
 }
@@ -812,11 +826,12 @@ function transformInner(input: {
   // --- Force escalation (reactive error recovery) ---
   // When the API previously rejected with "prompt is too long", skip layers
   // below the forced minimum to ensure enough trimming on the next attempt.
-  // One-shot: consumed here and reset to 0.
+  // One-shot: consumed here and reset to 0 (both in-memory and on disk).
   const sid = input.sessionID ?? input.messages[0]?.info.sessionID;
   const sessState = sid ? getSessionState(sid) : makeSessionState();
   let effectiveMinLayer = sessState.forceMinLayer;
   sessState.forceMinLayer = 0;
+  if (sid && effectiveMinLayer > 0) saveForceMinLayer(sid, 0);
 
   // --- Approach A: Cache-preserving passthrough ---
   // Use exact token count from the previous API response when available.
@@ -876,7 +891,12 @@ function transformInner(input: {
     expectedInput = messageTokens + overhead + ltmTokens;
   }
 
-  if (effectiveMinLayer === 0 && expectedInput <= maxInput) {
+  // When uncalibrated, apply safety multiplier to the layer-0 decision too.
+  // chars/3 undercounts by ~1.63x on real sessions — without this, a session
+  // estimated at 146K passes layer 0 but actually costs 214K → overflow.
+  const layer0Input = calibrated ? expectedInput : expectedInput * UNCALIBRATED_SAFETY;
+
+  if (effectiveMinLayer === 0 && layer0Input <= maxInput) {
     // All messages fit — return unmodified to preserve append-only prompt-cache pattern.
     // Raw messages are strictly better context than lossy distilled summaries.
     const messageTokens = calibrated

package/src/index.ts

@@ -15,10 +15,51 @@ import {
   setForceMinLayer,
   getLastTransformedCount,
 } from "./gradient";
-import { formatKnowledge } from "./prompt";
+import { formatKnowledge, formatDistillations } from "./prompt";
 import { createRecallTool } from "./reflect";
 import { shouldImport, importFromFile, exportToFile } from "./agents-file";
 
+/**
+ * Detect whether an error from session.error is a context overflow ("prompt too long").
+ * Matches both APIError wrapper shape (error.data.message) and direct shape (error.message).
+ */
+export function isContextOverflow(rawError: unknown): boolean {
+  const error = rawError as
+    | { name?: string; message?: string; data?: { message?: string } }
+    | undefined;
+  const errorMessage = error?.data?.message ?? error?.message ?? "";
+  return (
+    typeof errorMessage === "string" &&
+    (errorMessage.includes("prompt is too long") ||
+      errorMessage.includes("context length exceeded") ||
+      errorMessage.includes("maximum context length") ||
+      errorMessage.includes("ContextWindowExceededError") ||
+      errorMessage.includes("too many tokens"))
+  );
+}
+
+/**
+ * Build the synthetic recovery message injected after a context overflow.
+ * Contains the distilled session history so the model can continue.
+ */
+export function buildRecoveryMessage(
+  summaries: Array<{ observations: string; generation: number }>,
+): string {
+  const historyText = summaries.length > 0
+    ? formatDistillations(summaries)
+    : "";
+
+  return [
+    "<system-reminder>",
+    "The previous turn failed with a context overflow error (prompt too long).",
+    "Lore has automatically compressed the conversation history.",
+    "Review the session history below and continue where you left off.",
+    "",
+    historyText || "(No distilled history available — check recent messages for context.)",
+    "</system-reminder>",
+  ].join("\n");
+}
+
 export const LorePlugin: Plugin = async (ctx) => {
   const projectPath = ctx.worktree || ctx.directory;
   await load(ctx.directory);
@@ -226,45 +267,50 @@ export const LorePlugin: Plugin = async (ctx) => {
           | undefined;
         if (errorSessionID && await shouldSkip(errorSessionID)) return;
 
-        // Detect "prompt is too long" API errors and auto-recover:
-        // 1. Force the gradient transform to escalate on the next call (skip layer 0/1)
-        // 2. Force distillation to capture all temporal data before compaction
-        // 3. Trigger compaction so the session recovers without user intervention
+        // Detect "prompt is too long" API errors and auto-recover.
         const rawError = (event.properties as Record<string, unknown>).error;
-        // Diagnostic: log the full error shape so we can verify our detection matches
         console.error("[lore] session.error received:", JSON.stringify(rawError, null, 2));
 
-        const error = rawError as
-          | { name?: string; message?: string; data?: { message?: string } }
-          | undefined;
-        // Match both shapes: error.data.message (APIError wrapper) and error.message (direct)
-        const errorMessage = error?.data?.message ?? error?.message ?? "";
-        const isPromptTooLong =
-          typeof errorMessage === "string" &&
-          (errorMessage.includes("prompt is too long") ||
-            errorMessage.includes("context length exceeded") ||
-            errorMessage.includes("maximum context length") ||
-            errorMessage.includes("ContextWindowExceededError") ||
-            errorMessage.includes("too many tokens"));
-
-        console.error(
-          `[lore] session.error isPromptTooLong=${isPromptTooLong} (name=${error?.name}, message=${errorMessage.substring(0, 120)})`,
-        );
-
-        if (isPromptTooLong) {
+        if (isContextOverflow(rawError) && errorSessionID) {
           console.error(
-            `[lore] detected 'prompt too long' error — forcing distillation + layer escalation (session: ${errorSessionID?.substring(0, 16)})`,
+            `[lore] detected context overflow — auto-recovering (session: ${errorSessionID.substring(0, 16)})`,
           );
-          // Force layer 2 on next transform — layers 0 and 1 were already too large.
-          // The gradient at layers 2-4 will compress the context enough for the next turn.
-          // Do NOT call session.summarize() here — it sends all messages to the model,
-          // which would overflow again and create a stuck compaction loop.
+
+          // 1. Force layer 2 on next transform (persisted to DB — survives restarts).
           setForceMinLayer(2, errorSessionID);
 
-          if (errorSessionID) {
-            // Force distillation to capture all undistilled messages into the temporal
-            // store so they're preserved even if the session is later compacted manually.
-            await backgroundDistill(errorSessionID, true);
+          // 2. Distill all undistilled messages so nothing is lost.
+          await backgroundDistill(errorSessionID, true);
+
+          // 3. Auto-recover: inject a synthetic message that goes through the normal
+          //    chat path. The gradient transform fires with forceMinLayer=2, compressing
+          //    the context to fit. The model receives the distilled summaries and
+          //    continues where it left off — no user intervention needed.
+          try {
+            const summaries = distillation.loadForSession(projectPath, errorSessionID);
+            const recoveryText = buildRecoveryMessage(
+              summaries.map(s => ({ observations: s.observations, generation: s.generation })),
+            );
+
+            console.error(
+              `[lore] sending auto-recovery message to session ${errorSessionID.substring(0, 16)}`,
+            );
+            await ctx.client.session.prompt({
+              path: { id: errorSessionID },
+              body: {
+                parts: [{ type: "text", text: recoveryText, synthetic: true }],
+              },
+            });
+            console.error(
+              `[lore] auto-recovery message sent successfully`,
+            );
+          } catch (recoveryError) {
+            // Recovery is best-effort — don't let it crash the event handler.
+            // The persisted forceMinLayer will still help on the user's next message.
+            console.error(
+              `[lore] auto-recovery failed (forceMinLayer still persisted):`,
+              recoveryError,
+            );
           }
         }
       }