perf(memory): covering index on messages and load_history_filtered CTE refactor

CHANGELOG.md

@@ -23,6 +23,10 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 - `[profile.ci]` build profile with thin LTO and 16 codegen-units for faster CI release builds (#878)
 - `schema` feature flag in `zeph-llm` gating `schemars` dependency and typed output API (#879)
 
+### Performance
+- Added composite covering index `(conversation_id, id)` on `messages` table (migration 015); replaces single-column index for filter+order access patterns in `oldest_message_ids` and `load_history_filtered` (#895)
+- Replaced double-sort subquery in `load_history_filtered` with a CTE — eliminates redundant `ORDER BY` on the derived table (#896)
+
 ### Changed
 - Split 3177-line `src/main.rs` into focused modules: `runner.rs` (dispatch), `agent_setup.rs` (tool/MCP/feature setup), `tracing_init.rs`, `tui_bridge.rs`, `channel.rs`, `tests.rs` — `main.rs` reduced to 26 LOC (#839)
 - Split 1791-line `crates/zeph-core/src/bootstrap.rs` into submodule directory: `config.rs`, `health.rs`, `mcp.rs`, `provider.rs`, `skills.rs`, `tests.rs` — `bootstrap/mod.rs` reduced to 278 LOC (#840)

crates/zeph-memory/README.md

@@ -19,7 +19,7 @@ Includes a document ingestion subsystem for loading, chunking, and storing user
 
 | Module | Description |
 |--------|-------------|
-| `sqlite` | SQLite storage for conversations and messages; visibility-aware queries (`load_history_filtered`, `messages_by_ids`, `keyword_search`); durable compaction via `replace_conversation()` |
+| `sqlite` | SQLite storage for conversations and messages; visibility-aware queries (`load_history_filtered` via CTE, `messages_by_ids`, `keyword_search`); durable compaction via `replace_conversation()`; composite covering index `(conversation_id, id)` on messages for efficient history reads |
 | `sqlite::history` | Input history persistence for CLI channel |
 | `sqlite::acp_sessions` | ACP session and event persistence for session resume and lifecycle tracking |
 | `qdrant` | Qdrant client for vector upsert and search |

crates/zeph-memory/migrations/015_messages_covering_index.sql

@@ -0,0 +1,3 @@
+-- Covering index for queries that filter by conversation_id and order/limit by id.
+-- Replaces the single-column idx_messages_conversation_id for these access patterns.
+CREATE INDEX IF NOT EXISTS idx_messages_conversation_id_id ON messages(conversation_id, id);

crates/zeph-memory/src/sqlite/messages.rs

@@ -161,14 +161,14 @@ impl SqliteStore {
         let uv = user_visible.map(i64::from);
 
         let rows: Vec<(String, String, String, i64, i64)> = sqlx::query_as(
-            "SELECT role, content, parts, agent_visible, user_visible FROM (\
+            "WITH recent AS (\
                 SELECT role, content, parts, agent_visible, user_visible, id FROM messages \
                 WHERE conversation_id = ? \
                   AND (? IS NULL OR agent_visible = ?) \
                   AND (? IS NULL OR user_visible = ?) \
                 ORDER BY id DESC \
                 LIMIT ?\
-             ) ORDER BY id ASC",
+             ) SELECT role, content, parts, agent_visible, user_visible FROM recent ORDER BY id ASC",
         )
         .bind(conversation_id)
         .bind(av)

docs/src/architecture/performance.md

@@ -43,6 +43,12 @@ The TUI applies two optimizations to maintain responsive input during heavy stre
 - **Event loop batching**: `biased` `tokio::select!` prioritizes keyboard/mouse input over agent events. Agent events are drained via `try_recv` loop, coalescing multiple streaming chunks into a single frame redraw.
 - **Per-message render cache**: Syntax highlighting and markdown parsing results are cached with content-hash keys. Only messages with changed content are re-parsed. Cache invalidation triggers: content mutation, terminal resize, and view mode toggle.
 
+## SQLite Message Index
+
+Migration `015_messages_covering_index.sql` replaces the single-column `conversation_id` index on the `messages` table with a composite covering index on `(conversation_id, id)`. History queries filter by `conversation_id` and order by `id`, so the covering index satisfies both clauses from the index alone, eliminating the post-filter sort step.
+
+The `load_history_filtered` query uses a CTE to express the base filter before applying ordering and limit, replacing the previous double-sort subquery pattern.
+
 ## SQLite WAL Mode
 
 SQLite is opened with WAL (Write-Ahead Logging) mode, enabling concurrent reads during writes and improving throughput for the message persistence hot path.