@amaster.ai/dsh-storage 0.1.33 → 0.1.35

package/README.md

@@ -55,7 +55,7 @@ The logical message id rides in `metadata.id`; message rows use a deterministic
 
 For `user/message` rows, `content` remains the text projection, while `metadata.message` retains the original user message as received, including `source` and ordered content blocks. Image/file blocks keep whatever attachment reference, path, URL, or inline data the incoming message actually contains; this mirror does not fetch or rewrite attachments. Re-projection writes the same raw message into the existing row.
 
-The `tool_calls` array retains dsh's tool-call blocks (`id`, `name`, `arguments`); the matching entry gains the structured `result` block and, on failure, an `error` identity. The plugin reads persisted model rows to match results across process restarts and upserts the same row ID, so duplicate results do not add rows or increment the session's message count. Existing historical `type=tool` rows are left intact.
+The `tool_calls` array retains dsh's tool-call blocks (`id`, `name`, `arguments`); the matching entry gains the structured `result` block and, on failure, an `error` identity. Live call IDs are cached, and restored calls can resolve their message ID from the model surface: both paths read one message primary key. Older calls are searched in keyset pages of at most 100 ID/tool-call candidates, matched in the plugin and discarded before fetching the next page. Tool-result lookup SQL does not search JSON, join tables, or nest SELECT statements. Duplicate results update the same row and do not increment message counts. Existing historical `type=tool` rows are left intact.
 
 When an assistant message contains dsh `reasoning` blocks, `thoughts` is an array of `{ content }` entries in block order. If the adapter supplies a matching signature in the message's `source.replayState.blocks`, that entry also carries `thoughtSignature`. The full model `source`, including opaque `replayState`, is preserved in row metadata for lossless provider replay. The backend's `readMessages()` also accepts older `{ description }` entries and JSON string thoughts, returning them as `{ content }`. `subject` and a per-thought timestamp are not invented when the upstream message does not supply them.
 

package/lib/backends/database.js

@@ -24,6 +24,14 @@ import { createHash } from 'node:crypto';
 function pk(value) {
     return createHash('sha256').update(value).digest('hex').slice(0, 36);
 }
+const ARCHIVE_IN_PROGRESS = 'dsh-storage:archive-in-progress';
+const TOKEN_BASELINE_KEY = 'dsh-storage:token-baseline';
+function metadataOf(value) {
+    const parsed = typeof value === 'string' ? JSON.parse(value) : value;
+    return parsed && typeof parsed === 'object' && !Array.isArray(parsed)
+        ? parsed
+        : {};
+}
 /**
  * Parse the prisma-style sqlserver URL
  * (sqlserver://host:1433;database=d;user=u;password=p;encrypt=true;trustServerCertificate=true)
@@ -93,6 +101,11 @@ export class DatabaseBackend {
     constructor(config) {
         this.config = config;
     }
+    client() {
+        if (!this.prisma)
+            throw new Error('dsh-storage database is not initialized');
+        return this.prisma;
+    }
     /** Load the pre-generated client for the configured provider (seam for tests). */
     async loadClient() {
         const mod = await import(new URL(`../generated/${this.config.provider}/client.js`, import.meta.url).href);
@@ -103,8 +116,15 @@ export class DatabaseBackend {
             this.loadClient(),
             createAdapter(this.config),
         ]);
-        this.prisma = new PrismaClient({ adapter });
-        await this.prisma.$connect();
+        const prisma = new PrismaClient({ adapter });
+        try {
+            await prisma.$connect();
+            this.prisma = prisma;
+        }
+        catch (error) {
+            await prisma.$disconnect?.().catch(() => { });
+            throw error;
+        }
     }
     /** JSON column value, text-serialized on SQL Server (no Prisma Json type). */
     jsonField(value) {
@@ -130,31 +150,125 @@ export class DatabaseBackend {
             createdAt: row.createdAt,
         };
     }
-    async upsertMessage(row) {
-        if (!this.prisma)
-            return;
+    async writeMessage(client, row) {
         // NUL-separated: a space-joined pair would be ambiguous ('a b','c' ≡ 'a','b c').
         const id = pk(`message ${row.sessionId}\0${row.id}`);
         const data = this.toDbData(row);
-        await this.prisma.aiMessage.upsert({
+        await client.aiMessage.upsert({
             where: { id },
             create: { id, sessionId: row.sessionId, historyId: row.historyId, ...data },
             update: data,
         });
     }
+    async upsertMessage(row) {
+        await this.writeMessage(this.client(), row);
+    }
     async readMessages(sessionId, messageId) {
-        if (!this.prisma)
-            return [];
-        // ponytail: tool-call lookup scans one session; add an indexed call lookup if that becomes slow.
-        // Scan only the active segment for tool results; exact-ID reads still find
-        // archived rows so redelivery does not count the same message twice.
-        const rows = await this.prisma.aiMessage.findMany({
-            where: {
-                sessionId,
-                deletedAt: null,
-                ...(messageId ? { id: pk(`message ${sessionId}\0${messageId}`) } : { historyId: null }),
-            },
+        return this.readRows({
+            sessionId,
+            deletedAt: null,
+            id: pk(`message ${sessionId}\0${messageId}`),
+        });
+    }
+    async readToolCallMessages(sessionId, callId) {
+        const prisma = this.client();
+        const provider = this.config.provider;
+        const sqlserver = provider === 'sqlserver';
+        // SQL Server formats datetime2 with all seven fractional digits: a raw JS
+        // Date would lose the 100ns precision needed by keyset pagination.
+        const sortTime = provider === 'mysql'
+            ? "DATE_FORMAT(created_at, '%Y-%m-%d %H:%i:%s.%f')"
+            : provider === 'postgresql'
+                ? "to_char(created_at, 'YYYY-MM-DD HH24:MI:SS.US')"
+                : sqlserver
+                    ? "FORMAT(created_at, 'yyyy-MM-ddTHH:mm:ss.fffffff')"
+                    : 'CAST(created_at AS TEXT)';
+        let cursor;
+        let nullPhase = false;
+        for (;;) {
+            const params = provider === 'mysql' || provider === 'sqlite' ? [sessionId, sessionId] : [sessionId];
+            let before = '';
+            if (provider === 'postgresql') {
+                // PostgreSQL defaults to NULLS FIRST, which cannot be paginated with a
+                // single mixed-time cursor. Walk non-NULL timestamps first, then IDs
+                // of NULL rows, instead of relying on a NULLS LAST index definition.
+                if (nullPhase) {
+                    before =
+                        cursor?.time === null ? 'AND created_at IS NULL AND id < ?' : 'AND created_at IS NULL';
+                    if (cursor)
+                        params.push(cursor.id);
+                }
+                else {
+                    before = 'AND created_at IS NOT NULL';
+                    if (cursor?.time) {
+                        before += ' AND (created_at < ? OR (created_at = ? AND id < ?))';
+                        params.push(cursor.time, cursor.time, cursor.id);
+                    }
+                }
+            }
+            else if (cursor) {
+                if (cursor.time === null) {
+                    before = 'AND created_at IS NULL AND id < ?';
+                    params.push(cursor.id);
+                }
+                else {
+                    before = 'AND (created_at < ? OR (created_at = ? AND id < ?) OR created_at IS NULL)';
+                    params.push(cursor.time, cursor.time, cursor.id);
+                }
+            }
+            const orderBy = provider === 'postgresql' && nullPhase ? 'id DESC' : 'created_at DESC, id DESC';
+            let sql = `SELECT ${sqlserver ? 'TOP 100 ' : ''}id, ${sqlserver ? 'session_id, ' : ''}tool_calls, ${sortTime} AS sort_time
+        FROM ai_messages WHERE session_id = ?
+        ${provider === 'sqlite' ? 'AND session_id COLLATE BINARY = ?' : ''}
+        ${provider === 'mysql' ? 'AND BINARY session_id = BINARY ?' : ''}
+        AND deleted_at IS NULL AND history_id IS NULL AND type = 'model' AND tool_calls IS NOT NULL
+        ${before} ORDER BY ${orderBy} ${sqlserver ? '' : 'LIMIT 100'}`;
+            if (sqlserver || provider === 'postgresql') {
+                let index = 0;
+                sql = sql.replaceAll('?', () => `${sqlserver ? '@P' : '$'}${++index}`);
+            }
+            const page = await prisma.$queryRawUnsafe(sql, ...params);
+            for (const row of page) {
+                // The SQL Server index can seek on session_id without a collation
+                // expression; verify exact identity here instead of losing index use.
+                if (sqlserver && row.session_id !== undefined && row.session_id !== sessionId)
+                    continue;
+                let calls = row.tool_calls;
+                if (typeof calls === 'string') {
+                    try {
+                        calls = JSON.parse(calls);
+                    }
+                    catch {
+                        continue;
+                    }
+                }
+                if (Array.isArray(calls) && calls.some((call) => call?.id === callId)) {
+                    return this.readRows({
+                        sessionId,
+                        deletedAt: null,
+                        historyId: null,
+                        type: 'model',
+                        id: row.id,
+                    });
+                }
+            }
+            if (page.length < 100) {
+                if (provider === 'postgresql' && !nullPhase) {
+                    nullPhase = true;
+                    cursor = undefined;
+                    continue;
+                }
+                return [];
+            }
+            const last = page.at(-1);
+            cursor = { id: last.id, time: last.sort_time };
+        }
+    }
+    async readRows(where) {
+        const rows = await this.client().aiMessage.findMany({
+            where,
             orderBy: { createdAt: 'asc' },
+            take: 1,
         });
         const json = (value) => (typeof value === 'string' ? JSON.parse(value) : value);
         return rows.map((row) => {
@@ -191,9 +305,7 @@ export class DatabaseBackend {
         });
     }
     async readSession(sessionId) {
-        if (!this.prisma)
-            return null;
-        const row = await this.prisma.aiChatHistory.findFirst({
+        const row = await this.client().aiChatHistory.findFirst({
             where: { sessionId, deletedAt: null, messages: { none: {} } },
         });
         if (!row)
@@ -213,14 +325,12 @@ export class DatabaseBackend {
                     : row.metadata,
         };
     }
-    async upsertSession(row) {
-        if (!this.prisma)
-            return;
+    async writeSession(client, row) {
         // Find-then-write is race-free because the plugin serializes event
         // processing per session (see the per-session chain in index.ts). The
         // cuid primary key matches the source project and early-scaffold rows,
         // so existing data is continued, never duplicated.
-        const existing = await this.prisma.aiChatHistory.findFirst({
+        const existing = await client.aiChatHistory.findFirst({
             where: { sessionId: row.sessionId, deletedAt: null, messages: { none: {} } },
             select: { id: true },
         });
@@ -234,16 +344,25 @@ export class DatabaseBackend {
             metadata: this.jsonField(row.metadata),
         };
         if (existing) {
-            await this.prisma.aiChatHistory.update({ where: { id: existing.id }, data });
+            await client.aiChatHistory.update({ where: { id: existing.id }, data });
         }
         else {
-            await this.prisma.aiChatHistory.create({ data });
+            await client.aiChatHistory.create({ data });
         }
     }
+    async upsertSession(row) {
+        await this.writeSession(this.client(), row);
+    }
+    async upsertEvent(message, session) {
+        await this.client().$transaction(async (tx) => {
+            if (message)
+                await this.writeMessage(tx, message);
+            await this.writeSession(tx, session);
+        });
+    }
     async archiveSession(sessionId) {
-        if (!this.prisma)
-            throw new Error('dsh-storage database is not initialized');
-        return this.prisma.$transaction(async (tx) => {
+        const client = this.client();
+        const { expected, historyId } = await client.$transaction(async (tx) => {
             const where = { sessionId, historyId: null, deletedAt: null };
             const stats = await tx.aiMessage.aggregate({
                 where,
@@ -251,30 +370,148 @@ export class DatabaseBackend {
                 _min: { createdAt: true },
                 _max: { createdAt: true },
             });
-            if (stats._count._all === 0)
-                return null;
-            // Active rollups have no attached messages; an archive gains them below.
-            const active = await tx.aiChatHistory.findFirst({
-                where: { sessionId, deletedAt: null, messages: { none: {} } },
-                select: { id: true },
+            const expected = stats._count._all;
+            // A previous batch may have attached messages already. Resume that marked
+            // history before considering a newer active rollup or an empty finalization.
+            const interrupted = await tx.aiMessage.findFirst({
+                where: { sessionId, historyId: { not: null }, deletedAt: null },
+                orderBy: [{ updatedAt: 'desc' }, { id: 'desc' }],
+                select: { historyId: true },
             });
+            if (interrupted?.historyId) {
+                const candidate = await tx.aiChatHistory.findFirst({
+                    where: { id: interrupted.historyId },
+                    select: { metadata: true },
+                });
+                if (metadataOf(candidate?.metadata)[ARCHIVE_IN_PROGRESS] === true)
+                    return { expected, historyId: interrupted.historyId };
+            }
+            if (expected === 0)
+                return { expected, historyId: null };
             const rollup = {
-                messageCount: stats._count._all,
+                messageCount: expected,
                 firstMessageAt: stats._min.createdAt,
                 lastMessageAt: stats._max.createdAt,
             };
-            const history = active
-                ? await tx.aiChatHistory.update({ where: { id: active.id }, data: rollup })
-                : await tx.aiChatHistory.create({
-                    data: { sessionId, ...rollup, totalTokens: 0n },
-                    select: { id: true },
+            // Active rollups have no attached messages; an archive gains them below.
+            const active = await tx.aiChatHistory.findFirst({
+                where: { sessionId, deletedAt: null, messages: { none: {} } },
+                select: { id: true, metadata: true, totalTokens: true },
+            });
+            if (active) {
+                await tx.aiChatHistory.update({
+                    where: { id: active.id },
+                    data: {
+                        ...rollup,
+                        metadata: this.jsonField({
+                            ...metadataOf(active.metadata),
+                            [TOKEN_BASELINE_KEY]: Number(active.totalTokens ?? 0n),
+                            [ARCHIVE_IN_PROGRESS]: true,
+                        }),
+                    },
+                });
+                return { expected, historyId: active.id };
+            }
+            const history = await tx.aiChatHistory.create({
+                data: {
+                    sessionId,
+                    ...rollup,
+                    totalTokens: 0n,
+                    metadata: this.jsonField({ [ARCHIVE_IN_PROGRESS]: true }),
+                },
+                select: { id: true },
+            });
+            return { expected, historyId: history.id };
+        });
+        if (!historyId)
+            return null;
+        // Keep each write batch short instead of holding one transaction over the
+        // entire active segment. The final aggregate is the correctness barrier.
+        let updated = 0;
+        while (updated < expected) {
+            const candidates = await client.aiMessage.findMany({
+                where: { sessionId, historyId: null, deletedAt: null },
+                select: { id: true },
+                take: 100,
+            });
+            if (candidates.length === 0)
+                break;
+            const batch = await client.aiMessage.updateMany({
+                where: { id: { in: candidates.map((row) => row.id) } },
+                data: { historyId },
+            });
+            if (batch.count === 0)
+                break;
+            updated += batch.count;
+        }
+        if (updated !== expected)
+            throw new Error('active messages changed during archive');
+        await client.$transaction(async (tx) => {
+            const stats = await tx.aiMessage.aggregate({
+                where: { historyId, deletedAt: null },
+                _count: { _all: true },
+                _min: { createdAt: true },
+                _max: { createdAt: true },
+            });
+            const current = await tx.aiChatHistory.findFirst({
+                where: { id: historyId },
+                select: { metadata: true, totalTokens: true },
+            });
+            const latestActive = await tx.aiChatHistory.findFirst({
+                where: { sessionId, id: { not: historyId }, deletedAt: null, messages: { none: {} } },
+                orderBy: { lastMessageAt: 'desc' },
+                select: { totalTokens: true, metadata: true },
+            });
+            const currentTokens = current?.totalTokens ?? 0n;
+            const baselineValue = metadataOf(latestActive?.metadata)[TOKEN_BASELINE_KEY];
+            const baseline = Number(baselineValue);
+            let totalTokens = currentTokens;
+            if (latestActive && Number.isSafeInteger(baseline) && baseline >= 0) {
+                // The baseline records the active rollup's seed total. Everything
+                // after it belongs to this archive, including restart-only deltas.
+                const delta = latestActive.totalTokens - BigInt(baseline);
+                if (delta > 0n)
+                    totalTokens += delta;
+            }
+            else if (latestActive && latestActive.totalTokens > currentTokens) {
+                // Legacy rows without a baseline keep the cumulative compatibility path.
+                totalTokens = latestActive.totalTokens;
+            }
+            const metadata = { ...metadataOf(current?.metadata) };
+            delete metadata[ARCHIVE_IN_PROGRESS];
+            await tx.aiChatHistory.update({
+                where: { id: historyId },
+                data: {
+                    messageCount: stats._count._all,
+                    totalTokens,
+                    firstMessageAt: stats._min.createdAt,
+                    lastMessageAt: stats._max.createdAt,
+                    metadata: this.jsonField(metadata),
+                },
+            });
+            // A recovery may have created a second active rollup during the failed
+            // batches. Do not let its pre-archive counts seed the next live segment.
+            await tx.aiChatHistory.updateMany({
+                where: { sessionId, id: { not: historyId }, deletedAt: null, messages: { none: {} } },
+                data: { messageCount: 0, totalTokens: 0n, firstMessageAt: null, lastMessageAt: null },
+            });
+            const activeRollups = await tx.aiChatHistory.findMany({
+                where: { sessionId, id: { not: historyId }, deletedAt: null, messages: { none: {} } },
+                select: { id: true, metadata: true },
+            });
+            for (const active of activeRollups ?? []) {
+                // A completed archive must not leak its retry marker or stale baseline
+                // into the next live segment, or the next archive will skip freezing.
+                const activeMetadata = { ...metadataOf(active.metadata) };
+                delete activeMetadata[ARCHIVE_IN_PROGRESS];
+                delete activeMetadata[TOKEN_BASELINE_KEY];
+                await tx.aiChatHistory.update({
+                    where: { id: active.id },
+                    data: { metadata: this.jsonField(activeMetadata) },
                 });
-            const updated = await tx.aiMessage.updateMany({ where, data: { historyId: history.id } });
-            if (updated.count !== stats._count._all) {
-                throw new Error('active messages changed during archive');
             }
-            return history.id;
         });
+        return historyId;
     }
     async close() {
         await this.prisma?.$disconnect?.();

package/lib/generated/mysql/internal/class.js

@@ -15,7 +15,7 @@ const config = {
     "clientVersion": "7.9.1",
     "engineVersion": "e922089b7d7502aff4249d5da3420f6fa55fc6ad",
     "activeProvider": "mysql",
-    "inlineSchema": "// Prisma schema for the dsh-storage database backend — MySQL variant (Prisma 7).\n//\n// Table layout matches the source project's ai_messages / ai_chat_histories\n// (packages/core/prisma/schema.prisma) so existing data stays compatible:\n//   - `type` values come from the dsh session event map\n//     ('user' | 'model' | 'tool' | ...) instead of gemini-cli message roles.\n//   - `agent_id` distinguishes the main agent from subagent writers.\n//   - no user_id: tenancy is carried by session_id, not a column.\n//\n// The client is pre-generated at package build time (`pnpm generate`, output\n// src/generated/mysql, compiled into lib/generated) — consumers do NOT run\n// `prisma generate`. To create/upgrade tables, run `prisma db push` /\n// migrations against this file; the connection URL lives in prisma.config.ts\n// (Prisma 7 dropped datasource.url from the schema).\n\ngenerator client {\n  provider            = \"prisma-client\"\n  output              = \"../src/generated/mysql\"\n  importFileExtension = \"js\"\n}\n\ndatasource db {\n  provider = \"mysql\"\n}\n\n/// User and assistant messages; tool results update the assistant row's tool_calls.\nmodel AiMessage {\n  id        String    @id @default(cuid()) @db.Char(36)\n  sessionId String    @map(\"session_id\") @db.VarChar(100)\n  historyId String?   @map(\"history_id\") @db.Char(36)\n  type      String    @db.VarChar(50)\n  content   String    @db.LongText\n  thoughts  Json?\n  model     String?   @db.VarChar(100)\n  tokens    Json?\n  toolCalls Json?     @map(\"tool_calls\")\n  agentId   String?   @map(\"agent_id\") @db.VarChar(100)\n  metadata  Json?\n  createdAt DateTime  @default(now()) @map(\"created_at\") @db.Timestamp(0)\n  updatedAt DateTime  @default(now()) @updatedAt @map(\"updated_at\") @db.Timestamp(0)\n  deletedAt DateTime? @map(\"deleted_at\") @db.Timestamp(0)\n\n  // Relations\n  history AiChatHistory? @relation(fields: [historyId], references: [id], onDelete: NoAction)\n\n  @@index([sessionId], map: \"idx_session\")\n  @@index([historyId], map: \"idx_history\")\n  @@index([createdAt], map: \"idx_created\")\n  @@index([deletedAt], map: \"idx_deleted\")\n  @@index([agentId], map: \"idx_agent\")\n  @@index([sessionId, agentId, type, createdAt], map: \"idx_session_agent_type_created\")\n  @@map(\"ai_messages\")\n}\n\n/// Per-session rollup, maintained incrementally by the projector.\nmodel AiChatHistory {\n  id             String    @id @default(cuid()) @db.Char(36)\n  sessionId      String    @map(\"session_id\") @db.VarChar(100)\n  title          String?   @db.VarChar(255)\n  summary        String?   @db.Text\n  messageCount   Int       @default(0) @map(\"message_count\")\n  totalTokens    BigInt    @default(0) @map(\"total_tokens\")\n  firstMessageAt DateTime? @map(\"first_message_at\") @db.DateTime(6)\n  lastMessageAt  DateTime? @map(\"last_message_at\") @db.DateTime(6)\n  metadata       Json?\n  deletedAt      DateTime? @map(\"deleted_at\") @db.DateTime(6)\n\n  // Relations\n  messages AiMessage[]\n\n  @@index([sessionId], map: \"idx_chat_history_session\")\n  @@index([deletedAt], map: \"idx_chat_history_deleted\")\n  @@map(\"ai_chat_histories\")\n}\n",
+    "inlineSchema": "// Prisma schema for the dsh-storage database backend — MySQL variant (Prisma 7).\n//\n// Table layout matches the source project's ai_messages / ai_chat_histories\n// (packages/core/prisma/schema.prisma) so existing data stays compatible:\n//   - `type` values come from the dsh session event map\n//     ('user' | 'model' | 'tool' | ...) instead of gemini-cli message roles.\n//   - `agent_id` distinguishes the main agent from subagent writers.\n//   - no user_id: tenancy is carried by session_id, not a column.\n//\n// The client is pre-generated at package build time (`pnpm generate`, output\n// src/generated/mysql, compiled into lib/generated) — consumers do NOT run\n// `prisma generate`. To create/upgrade tables, run `prisma db push` /\n// migrations against this file; the connection URL lives in prisma.config.ts\n// (Prisma 7 dropped datasource.url from the schema).\n\ngenerator client {\n  provider            = \"prisma-client\"\n  output              = \"../src/generated/mysql\"\n  importFileExtension = \"js\"\n}\n\ndatasource db {\n  provider = \"mysql\"\n}\n\n/// User and assistant messages; tool results update the assistant row's tool_calls.\nmodel AiMessage {\n  id        String    @id @default(cuid()) @db.Char(36)\n  sessionId String    @map(\"session_id\") @db.VarChar(100)\n  historyId String?   @map(\"history_id\") @db.Char(36)\n  type      String    @db.VarChar(50)\n  content   String    @db.LongText\n  thoughts  Json?\n  model     String?   @db.VarChar(100)\n  tokens    Json?\n  toolCalls Json?     @map(\"tool_calls\")\n  agentId   String?   @map(\"agent_id\") @db.VarChar(100)\n  metadata  Json?\n  createdAt DateTime  @default(now()) @map(\"created_at\") @db.Timestamp(0)\n  updatedAt DateTime  @default(now()) @updatedAt @map(\"updated_at\") @db.Timestamp(0)\n  deletedAt DateTime? @map(\"deleted_at\") @db.Timestamp(0)\n\n  // Relations\n  history AiChatHistory? @relation(fields: [historyId], references: [id], onDelete: NoAction)\n\n  @@index([sessionId], map: \"idx_session\")\n  @@index([historyId], map: \"idx_history\")\n  @@index([createdAt], map: \"idx_created\")\n  @@index([deletedAt], map: \"idx_deleted\")\n  @@index([agentId], map: \"idx_agent\")\n  @@index([sessionId, agentId, type, createdAt], map: \"idx_session_agent_type_created\")\n  @@index([sessionId, historyId, deletedAt, type, createdAt, id], map: \"idx_tool_call_lookup\")\n  @@index([sessionId, deletedAt, updatedAt, id], map: \"idx_archive_recovery\")\n  @@map(\"ai_messages\")\n}\n\n/// Per-session rollup, maintained incrementally by the projector.\nmodel AiChatHistory {\n  id             String    @id @default(cuid()) @db.Char(36)\n  sessionId      String    @map(\"session_id\") @db.VarChar(100)\n  title          String?   @db.VarChar(255)\n  summary        String?   @db.Text\n  messageCount   Int       @default(0) @map(\"message_count\")\n  totalTokens    BigInt    @default(0) @map(\"total_tokens\")\n  firstMessageAt DateTime? @map(\"first_message_at\") @db.DateTime(6)\n  lastMessageAt  DateTime? @map(\"last_message_at\") @db.DateTime(6)\n  metadata       Json?\n  deletedAt      DateTime? @map(\"deleted_at\") @db.DateTime(6)\n\n  // Relations\n  messages AiMessage[]\n\n  @@index([sessionId], map: \"idx_chat_history_session\")\n  @@index([deletedAt], map: \"idx_chat_history_deleted\")\n  @@map(\"ai_chat_histories\")\n}\n",
     "runtimeDataModel": {
         "models": {},
         "enums": {},

package/lib/generated/postgresql/internal/class.js

@@ -15,7 +15,7 @@ const config = {
     "clientVersion": "7.9.1",
     "engineVersion": "e922089b7d7502aff4249d5da3420f6fa55fc6ad",
     "activeProvider": "postgresql",
-    "inlineSchema": "// Prisma schema for the dsh-storage database backend — PostgreSQL variant (Prisma 7).\n//\n// Same models as schema.mysql.prisma, with PostgreSQL native types\n// (`@db.Text` instead of `@db.LongText`, `@db.Timestamp` instead of\n// `@db.DateTime`). Client pre-generation and table creation work as described\n// in schema.mysql.prisma's header.\n\ngenerator client {\n  provider            = \"prisma-client\"\n  output              = \"../src/generated/postgresql\"\n  importFileExtension = \"js\"\n}\n\ndatasource db {\n  provider = \"postgresql\"\n}\n\n/// User and assistant messages; tool results update the assistant row's tool_calls.\nmodel AiMessage {\n  id        String    @id @default(cuid()) @db.Char(36)\n  sessionId String    @map(\"session_id\") @db.VarChar(100)\n  historyId String?   @map(\"history_id\") @db.Char(36)\n  type      String    @db.VarChar(50)\n  content   String    @db.Text\n  thoughts  Json?\n  model     String?   @db.VarChar(100)\n  tokens    Json?\n  toolCalls Json?     @map(\"tool_calls\")\n  agentId   String?   @map(\"agent_id\") @db.VarChar(100)\n  metadata  Json?\n  createdAt DateTime  @default(now()) @map(\"created_at\") @db.Timestamp(0)\n  updatedAt DateTime  @default(now()) @updatedAt @map(\"updated_at\") @db.Timestamp(0)\n  deletedAt DateTime? @map(\"deleted_at\") @db.Timestamp(0)\n\n  // Relations\n  history AiChatHistory? @relation(fields: [historyId], references: [id], onDelete: NoAction)\n\n  @@index([sessionId], map: \"idx_session\")\n  @@index([historyId], map: \"idx_history\")\n  @@index([createdAt], map: \"idx_created\")\n  @@index([deletedAt], map: \"idx_deleted\")\n  @@index([agentId], map: \"idx_agent\")\n  @@index([sessionId, agentId, type, createdAt], map: \"idx_session_agent_type_created\")\n  @@map(\"ai_messages\")\n}\n\n/// Per-session rollup, maintained incrementally by the projector.\nmodel AiChatHistory {\n  id             String    @id @default(cuid()) @db.Char(36)\n  sessionId      String    @map(\"session_id\") @db.VarChar(100)\n  title          String?   @db.VarChar(255)\n  summary        String?   @db.Text\n  messageCount   Int       @default(0) @map(\"message_count\")\n  totalTokens    BigInt    @default(0) @map(\"total_tokens\")\n  firstMessageAt DateTime? @map(\"first_message_at\") @db.Timestamp(6)\n  lastMessageAt  DateTime? @map(\"last_message_at\") @db.Timestamp(6)\n  metadata       Json?\n  deletedAt      DateTime? @map(\"deleted_at\") @db.Timestamp(6)\n\n  // Relations\n  messages AiMessage[]\n\n  @@index([sessionId], map: \"idx_chat_history_session\")\n  @@index([deletedAt], map: \"idx_chat_history_deleted\")\n  @@map(\"ai_chat_histories\")\n}\n",
+    "inlineSchema": "// Prisma schema for the dsh-storage database backend — PostgreSQL variant (Prisma 7).\n//\n// Same models as schema.mysql.prisma, with PostgreSQL native types\n// (`@db.Text` instead of `@db.LongText`, `@db.Timestamp` instead of\n// `@db.DateTime`). Client pre-generation and table creation work as described\n// in schema.mysql.prisma's header.\n\ngenerator client {\n  provider            = \"prisma-client\"\n  output              = \"../src/generated/postgresql\"\n  importFileExtension = \"js\"\n}\n\ndatasource db {\n  provider = \"postgresql\"\n}\n\n/// User and assistant messages; tool results update the assistant row's tool_calls.\nmodel AiMessage {\n  id        String    @id @default(cuid()) @db.Char(36)\n  sessionId String    @map(\"session_id\") @db.VarChar(100)\n  historyId String?   @map(\"history_id\") @db.Char(36)\n  type      String    @db.VarChar(50)\n  content   String    @db.Text\n  thoughts  Json?\n  model     String?   @db.VarChar(100)\n  tokens    Json?\n  toolCalls Json?     @map(\"tool_calls\")\n  agentId   String?   @map(\"agent_id\") @db.VarChar(100)\n  metadata  Json?\n  createdAt DateTime  @default(now()) @map(\"created_at\") @db.Timestamp(0)\n  updatedAt DateTime  @default(now()) @updatedAt @map(\"updated_at\") @db.Timestamp(0)\n  deletedAt DateTime? @map(\"deleted_at\") @db.Timestamp(0)\n\n  // Relations\n  history AiChatHistory? @relation(fields: [historyId], references: [id], onDelete: NoAction)\n\n  @@index([sessionId], map: \"idx_session\")\n  @@index([historyId], map: \"idx_history\")\n  @@index([createdAt], map: \"idx_created\")\n  @@index([deletedAt], map: \"idx_deleted\")\n  @@index([agentId], map: \"idx_agent\")\n  @@index([sessionId, agentId, type, createdAt], map: \"idx_session_agent_type_created\")\n  @@index([sessionId, historyId, deletedAt, type, createdAt, id], map: \"idx_tool_call_lookup\")\n  @@index([sessionId, deletedAt, updatedAt, id], map: \"idx_archive_recovery\")\n  @@map(\"ai_messages\")\n}\n\n/// Per-session rollup, maintained incrementally by the projector.\nmodel AiChatHistory {\n  id             String    @id @default(cuid()) @db.Char(36)\n  sessionId      String    @map(\"session_id\") @db.VarChar(100)\n  title          String?   @db.VarChar(255)\n  summary        String?   @db.Text\n  messageCount   Int       @default(0) @map(\"message_count\")\n  totalTokens    BigInt    @default(0) @map(\"total_tokens\")\n  firstMessageAt DateTime? @map(\"first_message_at\") @db.Timestamp(6)\n  lastMessageAt  DateTime? @map(\"last_message_at\") @db.Timestamp(6)\n  metadata       Json?\n  deletedAt      DateTime? @map(\"deleted_at\") @db.Timestamp(6)\n\n  // Relations\n  messages AiMessage[]\n\n  @@index([sessionId], map: \"idx_chat_history_session\")\n  @@index([deletedAt], map: \"idx_chat_history_deleted\")\n  @@map(\"ai_chat_histories\")\n}\n",
     "runtimeDataModel": {
         "models": {},
         "enums": {},

package/lib/generated/sqlite/internal/class.js

@@ -15,7 +15,7 @@ const config = {
     "clientVersion": "7.9.1",
     "engineVersion": "e922089b7d7502aff4249d5da3420f6fa55fc6ad",
     "activeProvider": "sqlite",
-    "inlineSchema": "// Prisma schema for the dsh-storage database backend — SQLite variant (Prisma 7).\n//\n// Same models as schema.mysql.prisma, without native type attributes\n// (SQLite has no `@db.*` vocabulary here). Client pre-generation and table\n// creation work as described in schema.mysql.prisma's header. Connection URL\n// shape: file:/absolute/path/to/db.sqlite.\n\ngenerator client {\n  provider            = \"prisma-client\"\n  output              = \"../src/generated/sqlite\"\n  importFileExtension = \"js\"\n}\n\ndatasource db {\n  provider = \"sqlite\"\n}\n\n/// User and assistant messages; tool results update the assistant row's tool_calls.\nmodel AiMessage {\n  id        String    @id @default(cuid())\n  sessionId String    @map(\"session_id\")\n  historyId String?   @map(\"history_id\")\n  type      String\n  content   String\n  thoughts  Json?\n  model     String?\n  tokens    Json?\n  toolCalls Json?     @map(\"tool_calls\")\n  agentId   String?   @map(\"agent_id\")\n  metadata  Json?\n  createdAt DateTime  @default(now()) @map(\"created_at\")\n  updatedAt DateTime  @default(now()) @updatedAt @map(\"updated_at\")\n  deletedAt DateTime? @map(\"deleted_at\")\n\n  // Relations\n  history AiChatHistory? @relation(fields: [historyId], references: [id], onDelete: NoAction)\n\n  @@index([sessionId], map: \"idx_session\")\n  @@index([historyId], map: \"idx_history\")\n  @@index([createdAt], map: \"idx_created\")\n  @@index([deletedAt], map: \"idx_deleted\")\n  @@index([agentId], map: \"idx_agent\")\n  @@index([sessionId, agentId, type, createdAt], map: \"idx_session_agent_type_created\")\n  @@map(\"ai_messages\")\n}\n\n/// Per-session rollup, maintained incrementally by the projector.\nmodel AiChatHistory {\n  id             String    @id @default(cuid())\n  sessionId      String    @map(\"session_id\")\n  title          String?\n  summary        String?\n  messageCount   Int       @default(0) @map(\"message_count\")\n  totalTokens    BigInt    @default(0) @map(\"total_tokens\")\n  firstMessageAt DateTime? @map(\"first_message_at\")\n  lastMessageAt  DateTime? @map(\"last_message_at\")\n  metadata       Json?\n  deletedAt      DateTime? @map(\"deleted_at\")\n\n  // Relations\n  messages AiMessage[]\n\n  @@index([sessionId], map: \"idx_chat_history_session\")\n  @@index([deletedAt], map: \"idx_chat_history_deleted\")\n  @@map(\"ai_chat_histories\")\n}\n",
+    "inlineSchema": "// Prisma schema for the dsh-storage database backend — SQLite variant (Prisma 7).\n//\n// Same models as schema.mysql.prisma, without native type attributes\n// (SQLite has no `@db.*` vocabulary here). Client pre-generation and table\n// creation work as described in schema.mysql.prisma's header. Connection URL\n// shape: file:/absolute/path/to/db.sqlite.\n\ngenerator client {\n  provider            = \"prisma-client\"\n  output              = \"../src/generated/sqlite\"\n  importFileExtension = \"js\"\n}\n\ndatasource db {\n  provider = \"sqlite\"\n}\n\n/// User and assistant messages; tool results update the assistant row's tool_calls.\nmodel AiMessage {\n  id        String    @id @default(cuid())\n  sessionId String    @map(\"session_id\")\n  historyId String?   @map(\"history_id\")\n  type      String\n  content   String\n  thoughts  Json?\n  model     String?\n  tokens    Json?\n  toolCalls Json?     @map(\"tool_calls\")\n  agentId   String?   @map(\"agent_id\")\n  metadata  Json?\n  createdAt DateTime  @default(now()) @map(\"created_at\")\n  updatedAt DateTime  @default(now()) @updatedAt @map(\"updated_at\")\n  deletedAt DateTime? @map(\"deleted_at\")\n\n  // Relations\n  history AiChatHistory? @relation(fields: [historyId], references: [id], onDelete: NoAction)\n\n  @@index([sessionId], map: \"idx_session\")\n  @@index([historyId], map: \"idx_history\")\n  @@index([createdAt], map: \"idx_created\")\n  @@index([deletedAt], map: \"idx_deleted\")\n  @@index([agentId], map: \"idx_agent\")\n  @@index([sessionId, agentId, type, createdAt], map: \"idx_session_agent_type_created\")\n  @@index([sessionId, historyId, deletedAt, type, createdAt, id], map: \"idx_tool_call_lookup\")\n  @@index([sessionId, deletedAt, updatedAt, id], map: \"idx_archive_recovery\")\n  @@map(\"ai_messages\")\n}\n\n/// Per-session rollup, maintained incrementally by the projector.\nmodel AiChatHistory {\n  id             String    @id @default(cuid())\n  sessionId      String    @map(\"session_id\")\n  title          String?\n  summary        String?\n  messageCount   Int       @default(0) @map(\"message_count\")\n  totalTokens    BigInt    @default(0) @map(\"total_tokens\")\n  firstMessageAt DateTime? @map(\"first_message_at\")\n  lastMessageAt  DateTime? @map(\"last_message_at\")\n  metadata       Json?\n  deletedAt      DateTime? @map(\"deleted_at\")\n\n  // Relations\n  messages AiMessage[]\n\n  @@index([sessionId], map: \"idx_chat_history_session\")\n  @@index([deletedAt], map: \"idx_chat_history_deleted\")\n  @@map(\"ai_chat_histories\")\n}\n",
     "runtimeDataModel": {
         "models": {},
         "enums": {},

package/lib/generated/sqlserver/internal/class.js

@@ -15,7 +15,7 @@ const config = {
     "clientVersion": "7.9.1",
     "engineVersion": "e922089b7d7502aff4249d5da3420f6fa55fc6ad",
     "activeProvider": "sqlserver",
-    "inlineSchema": "// Prisma schema for the dsh-storage database backend — SQL Server variant (Prisma 7).\n//\n// Same models as schema.mysql.prisma, with SQL Server native types. Prisma's\n// sqlserver connector has NO Json type, so every JSON column is a plain text\n// column here — the backend serializes them on write automatically (derived\n// from `provider: sqlserver`; SQL Server's own ISJSON/JSON_VALUE still query\n// the text as JSON).\n//\n// Connection URL shape (parsed by the backend into a node-mssql config):\n//   sqlserver://host:1433;database=dbname;user=sa;password=p;encrypt=true;trustServerCertificate=true\n// Client pre-generation and table creation work as described in\n// schema.mysql.prisma's header.\n\ngenerator client {\n  provider            = \"prisma-client\"\n  output              = \"../src/generated/sqlserver\"\n  importFileExtension = \"js\"\n}\n\ndatasource db {\n  provider = \"sqlserver\"\n}\n\n/// User and assistant messages; tool results update the assistant row's tool_calls.\nmodel AiMessage {\n  id        String    @id @default(cuid()) @db.Char(36)\n  sessionId String    @map(\"session_id\") @db.VarChar(100)\n  historyId String?   @map(\"history_id\") @db.Char(36)\n  type      String    @db.VarChar(50)\n  content   String    @db.NVarChar(Max)\n  thoughts  String?   @db.NVarChar(Max)\n  model     String?   @db.VarChar(100)\n  tokens    String?   @db.NVarChar(Max)\n  toolCalls String?   @map(\"tool_calls\") @db.NVarChar(Max)\n  agentId   String?   @map(\"agent_id\") @db.VarChar(100)\n  metadata  String?   @db.NVarChar(Max)\n  createdAt DateTime  @default(now()) @map(\"created_at\") @db.DateTime2\n  updatedAt DateTime  @default(now()) @updatedAt @map(\"updated_at\") @db.DateTime2\n  deletedAt DateTime? @map(\"deleted_at\") @db.DateTime2\n\n  // Relations\n  history AiChatHistory? @relation(fields: [historyId], references: [id], onDelete: NoAction)\n\n  @@index([sessionId], map: \"idx_session\")\n  @@index([historyId], map: \"idx_history\")\n  @@index([createdAt], map: \"idx_created\")\n  @@index([deletedAt], map: \"idx_deleted\")\n  @@index([agentId], map: \"idx_agent\")\n  @@index([sessionId, agentId, type, createdAt], map: \"idx_session_agent_type_created\")\n  @@map(\"ai_messages\")\n}\n\n/// Per-session rollup, maintained incrementally by the projector.\nmodel AiChatHistory {\n  id             String    @id @default(cuid()) @db.Char(36)\n  sessionId      String    @map(\"session_id\") @db.VarChar(100)\n  title          String?   @db.VarChar(255)\n  summary        String?   @db.NVarChar(Max)\n  messageCount   Int       @default(0) @map(\"message_count\")\n  totalTokens    BigInt    @default(0) @map(\"total_tokens\")\n  firstMessageAt DateTime? @map(\"first_message_at\") @db.DateTime2\n  lastMessageAt  DateTime? @map(\"last_message_at\") @db.DateTime2\n  metadata       String?   @db.NVarChar(Max)\n  deletedAt      DateTime? @map(\"deleted_at\") @db.DateTime2\n\n  // Relations\n  messages AiMessage[]\n\n  @@index([sessionId], map: \"idx_chat_history_session\")\n  @@index([deletedAt], map: \"idx_chat_history_deleted\")\n  @@map(\"ai_chat_histories\")\n}\n",
+    "inlineSchema": "// Prisma schema for the dsh-storage database backend — SQL Server variant (Prisma 7).\n//\n// Same models as schema.mysql.prisma, with SQL Server native types. Prisma's\n// sqlserver connector has NO Json type, so every JSON column is a plain text\n// column here — the backend serializes them on write automatically (derived\n// from `provider: sqlserver`; SQL Server's own ISJSON/JSON_VALUE still query\n// the text as JSON).\n//\n// Connection URL shape (parsed by the backend into a node-mssql config):\n//   sqlserver://host:1433;database=dbname;user=sa;password=p;encrypt=true;trustServerCertificate=true\n// Client pre-generation and table creation work as described in\n// schema.mysql.prisma's header.\n\ngenerator client {\n  provider            = \"prisma-client\"\n  output              = \"../src/generated/sqlserver\"\n  importFileExtension = \"js\"\n}\n\ndatasource db {\n  provider = \"sqlserver\"\n}\n\n/// User and assistant messages; tool results update the assistant row's tool_calls.\nmodel AiMessage {\n  id        String    @id @default(cuid()) @db.Char(36)\n  sessionId String    @map(\"session_id\") @db.VarChar(100)\n  historyId String?   @map(\"history_id\") @db.Char(36)\n  type      String    @db.VarChar(50)\n  content   String    @db.NVarChar(Max)\n  thoughts  String?   @db.NVarChar(Max)\n  model     String?   @db.VarChar(100)\n  tokens    String?   @db.NVarChar(Max)\n  toolCalls String?   @map(\"tool_calls\") @db.NVarChar(Max)\n  agentId   String?   @map(\"agent_id\") @db.VarChar(100)\n  metadata  String?   @db.NVarChar(Max)\n  createdAt DateTime  @default(now()) @map(\"created_at\") @db.DateTime2\n  updatedAt DateTime  @default(now()) @updatedAt @map(\"updated_at\") @db.DateTime2\n  deletedAt DateTime? @map(\"deleted_at\") @db.DateTime2\n\n  // Relations\n  history AiChatHistory? @relation(fields: [historyId], references: [id], onDelete: NoAction)\n\n  @@index([sessionId], map: \"idx_session\")\n  @@index([historyId], map: \"idx_history\")\n  @@index([createdAt], map: \"idx_created\")\n  @@index([deletedAt], map: \"idx_deleted\")\n  @@index([agentId], map: \"idx_agent\")\n  @@index([sessionId, agentId, type, createdAt], map: \"idx_session_agent_type_created\")\n  @@index([sessionId, historyId, deletedAt, type, createdAt, id], map: \"idx_tool_call_lookup\")\n  @@index([sessionId, deletedAt, updatedAt, id], map: \"idx_archive_recovery\")\n  @@map(\"ai_messages\")\n}\n\n/// Per-session rollup, maintained incrementally by the projector.\nmodel AiChatHistory {\n  id             String    @id @default(cuid()) @db.Char(36)\n  sessionId      String    @map(\"session_id\") @db.VarChar(100)\n  title          String?   @db.VarChar(255)\n  summary        String?   @db.NVarChar(Max)\n  messageCount   Int       @default(0) @map(\"message_count\")\n  totalTokens    BigInt    @default(0) @map(\"total_tokens\")\n  firstMessageAt DateTime? @map(\"first_message_at\") @db.DateTime2\n  lastMessageAt  DateTime? @map(\"last_message_at\") @db.DateTime2\n  metadata       String?   @db.NVarChar(Max)\n  deletedAt      DateTime? @map(\"deleted_at\") @db.DateTime2\n\n  // Relations\n  messages AiMessage[]\n\n  @@index([sessionId], map: \"idx_chat_history_session\")\n  @@index([deletedAt], map: \"idx_chat_history_deleted\")\n  @@map(\"ai_chat_histories\")\n}\n",
     "runtimeDataModel": {
         "models": {},
         "enums": {},

package/lib/index.js

@@ -20,7 +20,7 @@
 import Schema from '@deepseek-ai/schemastery';
 import { DatabaseBackend } from './backends/database.js';
 import { compressionSnapshot } from './compaction.js';
-import { mergeToolResult, projectEvent, usageSampleOf } from './projector.js';
+import { mergeToolResult, projectEvent, toolResultCallId, usageSampleOf } from './projector.js';
 export const name = 'dsh-storage';
 // Event-tap only: no hard service dependency. Add 'sessionPersistence' to
 // `inject` if you want fail-fast when persistence is not composed.
@@ -35,6 +35,8 @@ export const Config = Schema.object({
 });
 /** Metadata key for our usage bookkeeping — namespaced so other writers' fields never collide with it. */
 const LAST_USAGE_KEY = 'dsh-storage:lastUsage';
+const TOKEN_BASELINE_KEY = 'dsh-storage:token-baseline';
+const ARCHIVE_IN_PROGRESS = 'dsh-storage:archive-in-progress';
 /** Shape-guarded restore of our usage sample from arbitrary metadata. */
 function usageSampleOfMetadata(value) {
     const v = value;
@@ -63,24 +65,85 @@ export function apply(ctx, config) {
         }),
     ];
     const sessions = new Map();
+    const toolCallMessages = new Map();
+    const TOOL_CALL_CACHE_LIMIT = 1000;
+    function rememberToolCalls(sessionId, row) {
+        if (row?.type !== 'model' || !Array.isArray(row.toolCalls))
+            return;
+        for (const call of row.toolCalls) {
+            if (typeof call?.id !== 'string')
+                continue;
+            const calls = toolCallMessages.get(sessionId) ?? new Map();
+            // Bound each session's cache; Map iteration evicts the oldest insertion first.
+            if (!calls.has(call.id) && calls.size >= TOOL_CALL_CACHE_LIMIT) {
+                const oldest = calls.keys().next().value;
+                if (oldest !== undefined)
+                    calls.delete(oldest);
+            }
+            calls.set(call.id, row.id);
+            toolCallMessages.set(sessionId, calls);
+        }
+    }
+    async function readToolResultRows(session, sessionId, event) {
+        const callId = toolResultCallId(event);
+        if (!callId)
+            return [];
+        let messageId = toolCallMessages.get(sessionId)?.get(callId);
+        if (!messageId && typeof session?.deriveMessages === 'function') {
+            const messages = session.deriveMessages();
+            for (let index = messages.length - 1; index >= 0; index--) {
+                const message = messages[index];
+                if (message.role === 'assistant' &&
+                    typeof message.id === 'string' &&
+                    Array.isArray(message.content) &&
+                    message.content.some((block) => block?.type === 'tool-call' && block.id === callId)) {
+                    messageId = message.id;
+                    break;
+                }
+            }
+        }
+        if (messageId) {
+            const rows = await backends[0].readMessages(sessionId, messageId);
+            if (rows.some((row) => Array.isArray(row.toolCalls) && row.toolCalls.some((call) => call?.id === callId)))
+                return rows;
+        }
+        // Cold resume and eviction fall back to bounded database candidate pages.
+        const rows = await backends[0].readToolCallMessages(sessionId, callId);
+        for (const row of rows)
+            rememberToolCalls(sessionId, row);
+        return rows;
+    }
     // Row writes are fire-and-forget during the run, but tracked so the
     // session/flush checkpoint and shutdown can drain them — a one-shot
     // (headless) run would otherwise lose the tail events to process exit.
     const pending = new Set();
     function track(task) {
         pending.add(task);
-        void task.finally(() => pending.delete(task));
+        void task.then(() => pending.delete(task), () => pending.delete(task));
         return task;
     }
     function guard(call) {
         // Mirroring must never break the agent loop: log and swallow.
         return Promise.all(backends.map((backend) => call(backend).catch((error) => console.error(`[dsh-storage] ${backend.name} error:`, error)))).then(() => { });
     }
-    // Backend init starts at plugin load; every write chains behind it so no
-    // row is lost to a still-connecting backend.
+    // All sessions share one in-flight init. A failed attempt is cleared so the
+    // next event can retry once; no read or write passes a failed init barrier.
     let started;
+    function ensureStarted() {
+        if (!started) {
+            const attempt = track(Promise.all(backends.map((backend) => backend.init?.())).then(() => { }));
+            started = attempt;
+            void attempt.catch(() => {
+                if (started === attempt)
+                    started = undefined;
+            });
+        }
+        return started;
+    }
     function fanout(call) {
-        return track(started.then(() => guard(call)));
+        // Event failures reject into the per-session queue, which is the no-throw
+        // boundary. Swallowing here would make a partial write look committed.
+        return track(ensureStarted().then(() => Promise.all(backends.map(call)).then(() => { })));
     }
     // Per-session serialization: all event processing for one session runs
     // through a promise chain, so rollup writes complete in event order
@@ -90,10 +153,10 @@ export function apply(ctx, config) {
     // so history must come from the database, not from zero).
     const chains = new Map();
     function enqueue(sessionId, task) {
-        // First task of a session also chains behind backend init, so seedAccum
-        // never reads from a still-connecting backend (prisma === null would
-        // masquerade as "row not found" and zero-seed over the stored rollup).
-        const next = (chains.get(sessionId) ?? started)
+        // Every task rechecks the init barrier: a failed earlier attempt must not
+        // become a resolved chain that lets later events zero-seed or no-op write.
+        const next = (chains.get(sessionId) ?? Promise.resolve())
+            .then(() => ensureStarted())
             .then(task)
             .catch((error) => console.error('[dsh-storage] mirror task error:', error));
         chains.set(sessionId, next);
@@ -115,7 +178,9 @@ export function apply(ctx, config) {
                         firstMessageAt: row.firstMessageAt ?? undefined,
                         lastMessageAt: row.lastMessageAt ?? undefined,
                         // Preserve every existing metadata field; we only own our key.
-                        metadata: { ...(row.metadata ?? {}) },
+                        // The baseline distinguishes cumulative active totals from a
+                        // restart's fresh-segment deltas during archive recovery.
+                        metadata: { [TOKEN_BASELINE_KEY]: row.totalTokens, ...(row.metadata ?? {}) },
                         usageSample: usageSampleOfMetadata(row.metadata?.[LAST_USAGE_KEY]),
                     };
                 }
@@ -127,13 +192,14 @@ export function apply(ctx, config) {
         }
         if (readFailure)
             throw readFailure;
-        return { messageCount: 0, totalTokens: 0, metadata: {} };
+        return { messageCount: 0, totalTokens: 0, metadata: { [TOKEN_BASELINE_KEY]: 0 } };
     }
     ctx.effect(() => {
-        started = track(guard(async (backend) => backend.init?.()));
+        void ensureStarted().catch((error) => console.error('[dsh-storage] database error:', error));
         return async () => {
             sessions.clear();
-            await Promise.all([...pending]);
+            toolCallMessages.clear();
+            await Promise.allSettled([...pending]);
             await guard(async (backend) => backend.close?.());
         };
     });
@@ -141,9 +207,20 @@ export function apply(ctx, config) {
         archiveSession(sessionId) {
             // Join the same per-session chain as event writes, then discard the old
             // in-memory rollup only after the database transaction commits.
-            const archived = (chains.get(sessionId) ?? started).then(async () => {
+            const archived = (chains.get(sessionId) ?? Promise.resolve())
+                .then(() => ensureStarted())
+                .then(async () => {
+                const accum = sessions.get(sessionId);
+                if (accum && !accum.metadata[ARCHIVE_IN_PROGRESS]) {
+                    // Freeze inside the session queue: all earlier events have now
+                    // committed, and this first baseline survives every retry until
+                    // the archive actually succeeds.
+                    accum.metadata[TOKEN_BASELINE_KEY] = accum.totalTokens;
+                    accum.metadata[ARCHIVE_IN_PROGRESS] = true;
+                }
                 const id = await backends[0].archiveSession(sessionId);
                 sessions.delete(sessionId);
+                toolCallMessages.delete(sessionId);
                 return id;
             });
             const settled = archived.then(() => { }, () => { });
@@ -169,18 +246,20 @@ export function apply(ctx, config) {
                 await fanout((backend) => backend.upsertMessage(snapshot));
                 return;
             }
-            let accum = sessions.get(sessionId);
-            if (!accum) {
-                accum = await seedAccum(sessionId);
-                sessions.set(sessionId, accum);
-            }
+            const stored = sessions.get(sessionId) ?? (await seedAccum(sessionId));
+            sessions.set(sessionId, stored);
+            // Stage every change on a copy. The live accumulator advances only
+            // after the message + rollup transaction commits.
+            const accum = { ...stored, metadata: { ...stored.metadata } };
             let row = projectEvent(session, event, sessionId);
             if (row && session.header?.createdAt !== undefined) {
                 row.metadata = { ...row.metadata, sessionCreatedAt: session.header.createdAt };
             }
-            const storedRows = row || event?.type === 'tool/result'
-                ? await backends[0].readMessages(sessionId, row?.id)
-                : [];
+            const storedRows = event?.type === 'tool/result'
+                ? await readToolResultRows(session, sessionId, event)
+                : row
+                    ? await backends[0].readMessages(sessionId, row.id)
+                    : [];
             if (event?.type === 'tool/result')
                 row = mergeToolResult(event, storedRows);
             const existing = row && storedRows.find((stored) => stored.id === row?.id);
@@ -201,28 +280,21 @@ export function apply(ctx, config) {
             if (!row && event?.type === 'tool/result') {
                 console.error('[dsh-storage] no matching model tool call for result');
             }
-            if (row) {
-                if (!existing && event?.type !== 'tool/result') {
-                    accum.messageCount += 1;
-                    accum.firstMessageAt ??= row.createdAt;
-                    accum.lastMessageAt = row.createdAt;
-                }
-                await fanout((backend) => backend.upsertMessage(row));
+            if (row && !existing && event?.type !== 'tool/result') {
+                accum.messageCount += 1;
+                accum.firstMessageAt ??= row.createdAt;
+                accum.lastMessageAt = row.createdAt;
             }
-            // Replacement usage accounting: the latest sample for a (turn, step)
-            // supersedes the previous one — fold only the delta, so the
-            // chunk↔message pair and progressive samples never double-count, and a
-            // failed step (usage chunk, no message) still counts. Samples of one
-            // step arrive adjacently, so a single latest sample is all we track.
+            // Every settled attempt/message is one model call; retries on the same
+            // step are additive rather than replacement snapshots.
             const sampled = usageSampleOf(event);
             let usageDelta = 0;
             if (sampled) {
-                const prev = accum.usageSample;
-                usageDelta =
-                    sampled.sample.input +
-                        sampled.sample.output -
-                        (prev && prev.key === sampled.key ? prev.input + prev.output : 0);
-                if (!prev || prev.key !== sampled.key || usageDelta !== 0) {
+                const sequence = Number(sampled.key);
+                const previous = Number(accum.usageSample?.key);
+                const duplicate = Number.isSafeInteger(sequence) && Number.isSafeInteger(previous) && sequence <= previous;
+                if (!duplicate) {
+                    usageDelta = sampled.sample.input + sampled.sample.output;
                     accum.usageSample = { key: sampled.key, ...sampled.sample };
                     accum.totalTokens += usageDelta;
                 }
@@ -251,8 +323,10 @@ export function apply(ctx, config) {
                     lastMessageAt: accum.lastMessageAt ?? null,
                     ...(Object.keys(metadata).length > 0 ? { metadata } : {}),
                 };
-                await fanout((backend) => backend.upsertSession(sessionRow));
+                await fanout((backend) => backend.upsertEvent(row, sessionRow));
             }
+            sessions.set(sessionId, accum);
+            rememberToolCalls(sessionId, row);
         });
     });
     ctx.on('session/disposed', (session) => {
@@ -263,6 +337,7 @@ export function apply(ctx, config) {
         // the dispose was meant to drop).
         enqueue(sessionId, async () => {
             sessions.delete(sessionId);
+            toolCallMessages.delete(sessionId);
         });
         // Delete the chain entry only once it has settled AND is still the same
         // entry — deleting earlier would let a recreated session race the
@@ -274,8 +349,19 @@ export function apply(ctx, config) {
                 chains.delete(sessionId);
         });
     });
-    // Durability checkpoint: the store awaits every session/flush listener.
-    ctx.on('session/flush', async () => {
-        await Promise.all([...pending]);
+    // Durability checkpoint: wait only for this session. Shutdown still drains
+    // the global pending set; an unrelated slow session must not block this one.
+    ctx.on('session/flush', async (session) => {
+        const chain = chains.get(session?.id ?? 'unknown');
+        if (chain) {
+            await chain;
+            return;
+        }
+        try {
+            await ensureStarted();
+        }
+        catch (error) {
+            console.error('[dsh-storage] database error:', error);
+        }
     });
 }

package/lib/projector.js

@@ -106,12 +106,32 @@ export function projectEvent(_session, event, sessionId) {
             return null;
     }
 }
+/** Identify the assistant tool call answered by either supported result shape. */
+export function toolResultCallId(event) {
+    const message = event?.data?.message;
+    const first = message?.content?.[0];
+    const callId = message?.toolCallId ??
+        message?.source?.callId ??
+        (first?.type === 'tool-result' ? first.toolCallId : undefined);
+    return typeof callId === 'string' ? callId : undefined;
+}
 /** Update the persisted assistant call identified by a tool result. */
 export function mergeToolResult(event, rows) {
-    const result = event?.data?.message?.content?.[0] ?? null;
-    const callId = event?.data?.message?.source?.callId ?? result?.toolCallId;
+    const message = event?.data?.message;
+    const content = Array.isArray(message?.content) ? message.content : [];
+    const legacy = content.length === 1 && content[0]?.type === 'tool-result' ? content[0] : undefined;
+    const callId = toolResultCallId(event);
     if (!callId)
         return null;
+    const result = legacy ??
+        (message
+            ? {
+                type: 'tool-result',
+                toolCallId: callId,
+                content,
+                ...(typeof message.isError === 'boolean' ? { isError: message.isError } : {}),
+            }
+            : null);
     const model = [...rows]
         .reverse()
         .find((row) => row.type === 'model' &&
@@ -133,21 +153,19 @@ export function mergeToolResult(event, rows) {
     };
 }
 /**
- * Extract the usage sample carried by one event, keyed by its step, or null
- * when the event carries none.
+ * Extract one model call's usage, keyed by its durable event sequence.
  *
  * V3 session log (dsh-session@0.1.6-alpha.2):
- * - `assistant/message.data.usage` — the settled step's final accounting;
- * - `assistant/attempt` — a failed/retried/cancelled step commits no surface
+ * - `assistant/message.data.usage` — the settled call's accounting;
+ * - `assistant/attempt` — a failed/retried/cancelled call commits no surface
  *   message; the adapter's usage chunk survives as a raw `chunk` record in the
  *   embedded compact `stream` (deltas are packed, but usage is a raw type).
  *
- * Samples for one `(turn, step)` are REPLACEMENTS, not additive: later
- * samples supersede earlier ones (a failed attempt's usage, then the settled
- * message's). The caller keeps the latest sample per step and folds only the
- * delta into its rollup. Totals include the cache buckets: cacheRead/cacheWrite
- * are billed in their own buckets (observed cacheReadTokens ≫ inputTokens in
- * dsh session logs — they are not subsets of inputTokens).
+ * Calls on one `(turn, step)` are additive: a retry is another billed model
+ * call, not a progressive snapshot. Only redelivery of the same event sequence
+ * is a duplicate. Totals include the cache buckets: cacheRead/cacheWrite are
+ * billed in their own buckets (observed cacheReadTokens ≫ inputTokens in dsh
+ * session logs — they are not subsets of inputTokens).
  */
 export function usageSampleOf(event) {
     const total = (u) => ({
@@ -167,7 +185,7 @@ export function usageSampleOf(event) {
     else {
         return null;
     }
-    if (!usage || data?.turn == null || data?.step == null)
+    if (!usage || !Number.isSafeInteger(event?.seq))
         return null;
-    return { key: `${data.turn}:${data.step}`, sample: total(usage) };
+    return { key: String(event.seq), sample: total(usage) };
 }

package/lib/types/backends/database.d.ts

@@ -24,16 +24,22 @@ export declare class DatabaseBackend implements StorageBackend {
     readonly name = "database";
     private prisma;
     constructor(config: DatabaseBackendConfig);
+    private client;
     /** Load the pre-generated client for the configured provider (seam for tests). */
     protected loadClient(): Promise<any>;
     init(): Promise<void>;
     /** JSON column value, text-serialized on SQL Server (no Prisma Json type). */
     private jsonField;
     private toDbData;
+    private writeMessage;
     upsertMessage(row: MessageRow): Promise<void>;
-    readMessages(sessionId: string, messageId?: string): Promise<MessageRow[]>;
+    readMessages(sessionId: string, messageId: string): Promise<MessageRow[]>;
+    readToolCallMessages(sessionId: string, callId: string): Promise<MessageRow[]>;
+    private readRows;
     readSession(sessionId: string): Promise<SessionRow | null>;
+    private writeSession;
     upsertSession(row: SessionRow): Promise<void>;
+    upsertEvent(message: MessageRow | null, session: SessionRow): Promise<void>;
     archiveSession(sessionId: string): Promise<string | null>;
     close(): Promise<void>;
 }

package/lib/types/index.d.ts

@@ -22,29 +22,29 @@ import Schema from '@deepseek-ai/schemastery';
 import { type DatabaseProvider } from './backends/database.js';
 export declare const name = "dsh-storage";
 export declare const inject: string[];
-export declare const Config: Schema<Schemastery.ObjectS<{
-    enabled: Schema<boolean, boolean>;
-    database: Schema<Schemastery.ObjectS<{
-        enabled: Schema<boolean, boolean>;
-        provider: Schema<"mysql" | "postgresql" | "sqlite" | "sqlserver", "mysql" | "postgresql" | "sqlite" | "sqlserver">;
-        url: Schema<string, string>;
-    }>, Schemastery.ObjectT<{
-        enabled: Schema<boolean, boolean>;
-        provider: Schema<"mysql" | "postgresql" | "sqlite" | "sqlserver", "mysql" | "postgresql" | "sqlite" | "sqlserver">;
-        url: Schema<string, string>;
-    }>>;
-}>, Schemastery.ObjectT<{
-    enabled: Schema<boolean, boolean>;
-    database: Schema<Schemastery.ObjectS<{
-        enabled: Schema<boolean, boolean>;
-        provider: Schema<"mysql" | "postgresql" | "sqlite" | "sqlserver", "mysql" | "postgresql" | "sqlite" | "sqlserver">;
-        url: Schema<string, string>;
-    }>, Schemastery.ObjectT<{
-        enabled: Schema<boolean, boolean>;
-        provider: Schema<"mysql" | "postgresql" | "sqlite" | "sqlserver", "mysql" | "postgresql" | "sqlite" | "sqlserver">;
-        url: Schema<string, string>;
-    }>>;
-}>>;
+export declare const Config: Schema<Schemastery.ObjectS<NoInfer<{
+    enabled: Schema<boolean, boolean, "defined">;
+    database: Schema<Schemastery.ObjectS<NoInfer<{
+        enabled: Schema<boolean, boolean, "defined">;
+        provider: Schema<"mysql" | "postgresql" | "sqlite" | "sqlserver", "mysql" | "postgresql" | "sqlite" | "sqlserver", "defined">;
+        url: Schema<string, string, "defined">;
+    }>>, Schemastery.ObjectT<NoInfer<{
+        enabled: Schema<boolean, boolean, "defined">;
+        provider: Schema<"mysql" | "postgresql" | "sqlite" | "sqlserver", "mysql" | "postgresql" | "sqlite" | "sqlserver", "defined">;
+        url: Schema<string, string, "defined">;
+    }>>, "plain">;
+}>>, Schemastery.ObjectT<NoInfer<{
+    enabled: Schema<boolean, boolean, "defined">;
+    database: Schema<Schemastery.ObjectS<NoInfer<{
+        enabled: Schema<boolean, boolean, "defined">;
+        provider: Schema<"mysql" | "postgresql" | "sqlite" | "sqlserver", "mysql" | "postgresql" | "sqlite" | "sqlserver", "defined">;
+        url: Schema<string, string, "defined">;
+    }>>, Schemastery.ObjectT<NoInfer<{
+        enabled: Schema<boolean, boolean, "defined">;
+        provider: Schema<"mysql" | "postgresql" | "sqlite" | "sqlserver", "mysql" | "postgresql" | "sqlite" | "sqlserver", "defined">;
+        url: Schema<string, string, "defined">;
+    }>>, "plain">;
+}>>, "plain">;
 export interface StoragePluginConfig {
     enabled: boolean;
     database: {

package/lib/types/projector.d.ts

@@ -22,6 +22,8 @@ import type { MessageRow } from './types.js';
  * turn/step lifecycle, approvals, ...).
  */
 export declare function projectEvent(_session: any, event: any, sessionId: string): MessageRow | null;
+/** Identify the assistant tool call answered by either supported result shape. */
+export declare function toolResultCallId(event: any): string | undefined;
 /** Update the persisted assistant call identified by a tool result. */
 export declare function mergeToolResult(event: any, rows: MessageRow[]): MessageRow | null;
 export interface UsageSample {
@@ -29,21 +31,19 @@ export interface UsageSample {
     output: number;
 }
 /**
- * Extract the usage sample carried by one event, keyed by its step, or null
- * when the event carries none.
+ * Extract one model call's usage, keyed by its durable event sequence.
  *
  * V3 session log (dsh-session@0.1.6-alpha.2):
- * - `assistant/message.data.usage` — the settled step's final accounting;
- * - `assistant/attempt` — a failed/retried/cancelled step commits no surface
+ * - `assistant/message.data.usage` — the settled call's accounting;
+ * - `assistant/attempt` — a failed/retried/cancelled call commits no surface
  *   message; the adapter's usage chunk survives as a raw `chunk` record in the
  *   embedded compact `stream` (deltas are packed, but usage is a raw type).
  *
- * Samples for one `(turn, step)` are REPLACEMENTS, not additive: later
- * samples supersede earlier ones (a failed attempt's usage, then the settled
- * message's). The caller keeps the latest sample per step and folds only the
- * delta into its rollup. Totals include the cache buckets: cacheRead/cacheWrite
- * are billed in their own buckets (observed cacheReadTokens ≫ inputTokens in
- * dsh session logs — they are not subsets of inputTokens).
+ * Calls on one `(turn, step)` are additive: a retry is another billed model
+ * call, not a progressive snapshot. Only redelivery of the same event sequence
+ * is a duplicate. Totals include the cache buckets: cacheRead/cacheWrite are
+ * billed in their own buckets (observed cacheReadTokens ≫ inputTokens in dsh
+ * session logs — they are not subsets of inputTokens).
  */
 export declare function usageSampleOf(event: any): {
     key: string;

package/lib/types/types.d.ts

@@ -33,20 +33,23 @@ export interface SessionRow {
     metadata?: Record<string, unknown>;
 }
 /**
- * A storage backend receives already-projected rows. Backends must be
- * no-throw from the plugin's perspective: implementations report errors to
- * the plugin logger and swallow them, mirroring must never break the agent
- * loop.
+ * A storage backend receives already-projected rows. Backend errors reject
+ * into the plugin's per-session queue, whose catch/log boundary keeps
+ * mirroring failures out of the agent loop.
  */
 export interface StorageBackend {
     readonly name: string;
     init?(): Promise<void>;
     /** Read the stored session row, for seeding the live rollup on resume. */
     readSession?(sessionId: string): Promise<SessionRow | null>;
-    /** Read persisted messages, optionally narrowing to one logical message ID. */
-    readMessages(sessionId: string, messageId?: string): Promise<MessageRow[]>;
+    /** Read one persisted logical message, including archived rows for redelivery. */
+    readMessages(sessionId: string, messageId: string): Promise<MessageRow[]>;
+    /** Find the newest active model message containing this exact tool call. */
+    readToolCallMessages(sessionId: string, callId: string): Promise<MessageRow[]>;
     upsertMessage(row: MessageRow): Promise<void>;
     upsertSession(row: SessionRow): Promise<void>;
+    /** Atomically persist one event's optional message and its session rollup. */
+    upsertEvent(message: MessageRow | null, session: SessionRow): Promise<void>;
     /** Move active messages into a queryable history row; null when already empty. */
     archiveSession(sessionId: string): Promise<string | null>;
     close?(): Promise<void>;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@amaster.ai/dsh-storage",
-  "version": "0.1.33",
+  "version": "0.1.35",
   "description": "Session storage mirror for DeepSeek Harness: projects the session event stream into MySQL/PostgreSQL/SQLite/SQL Server (ai_messages / ai_chat_histories) without replacing dsh's local persistence.",
   "type": "module",
   "main": "lib/index.js",
@@ -33,7 +33,7 @@
     }
   },
   "dependencies": {
-    "@deepseek-ai/schemastery": "^3.18.1"
+    "@deepseek-ai/schemastery": "^3.18.5"
   },
   "peerDependencies": {
     "@deepseek-ai/cordis": "^4.0.1",
@@ -63,7 +63,7 @@
   },
   "devDependencies": {
     "@deepseek-ai/cordis": "4.0.5",
-    "@deepseek-ai/dsh-session": "0.2.1-rc.4",
+    "@deepseek-ai/dsh-session": "0.2.1-rc.5",
     "@prisma/adapter-libsql": "^7.9.1",
     "@prisma/adapter-mariadb": "^7.9.1",
     "@prisma/adapter-mssql": "^7.9.1",

package/prisma/schema.mysql.prisma

@@ -49,6 +49,8 @@ model AiMessage {
   @@index([deletedAt], map: "idx_deleted")
   @@index([agentId], map: "idx_agent")
   @@index([sessionId, agentId, type, createdAt], map: "idx_session_agent_type_created")
+  @@index([sessionId, historyId, deletedAt, type, createdAt, id], map: "idx_tool_call_lookup")
+  @@index([sessionId, deletedAt, updatedAt, id], map: "idx_archive_recovery")
   @@map("ai_messages")
 }
 

package/prisma/schema.postgresql.prisma

@@ -41,6 +41,8 @@ model AiMessage {
   @@index([deletedAt], map: "idx_deleted")
   @@index([agentId], map: "idx_agent")
   @@index([sessionId, agentId, type, createdAt], map: "idx_session_agent_type_created")
+  @@index([sessionId, historyId, deletedAt, type, createdAt, id], map: "idx_tool_call_lookup")
+  @@index([sessionId, deletedAt, updatedAt, id], map: "idx_archive_recovery")
   @@map("ai_messages")
 }
 

package/prisma/schema.sqlite.prisma

@@ -41,6 +41,8 @@ model AiMessage {
   @@index([deletedAt], map: "idx_deleted")
   @@index([agentId], map: "idx_agent")
   @@index([sessionId, agentId, type, createdAt], map: "idx_session_agent_type_created")
+  @@index([sessionId, historyId, deletedAt, type, createdAt, id], map: "idx_tool_call_lookup")
+  @@index([sessionId, deletedAt, updatedAt, id], map: "idx_archive_recovery")
   @@map("ai_messages")
 }
 

package/prisma/schema.sqlserver.prisma

@@ -47,6 +47,8 @@ model AiMessage {
   @@index([deletedAt], map: "idx_deleted")
   @@index([agentId], map: "idx_agent")
   @@index([sessionId, agentId, type, createdAt], map: "idx_session_agent_type_created")
+  @@index([sessionId, historyId, deletedAt, type, createdAt, id], map: "idx_tool_call_lookup")
+  @@index([sessionId, deletedAt, updatedAt, id], map: "idx_archive_recovery")
   @@map("ai_messages")
 }