context-mode 1.0.151 → 1.0.152

package/build/store.js

@@ -262,7 +262,7 @@ function findMinSpan(positionLists) {
         return Infinity;
     if (positionLists.length === 1)
         return 0;
-    const sorted = positionLists.map((p) => [...p].sort((a, b) => a - b));
+    const sorted = positionLists;
     const ptrs = new Array(sorted.length).fill(0);
     let minSpan = Infinity;
     while (true) {
@@ -507,7 +507,8 @@ export class ContentStore {
         chunks.timestamp,
         sources.label,
         bm25(chunks, 5.0, 1.0) AS rank,
-        highlight(chunks, 1, char(2), char(3)) AS highlighted
+        highlight(chunks, 1, char(2), char(3)) AS highlighted,
+        chunks.session_id
       FROM chunks
       JOIN sources ON sources.id = chunks.source_id
       WHERE chunks MATCH ?
@@ -522,7 +523,8 @@ export class ContentStore {
         chunks.timestamp,
         sources.label,
         bm25(chunks, 5.0, 1.0) AS rank,
-        highlight(chunks, 1, char(2), char(3)) AS highlighted
+        highlight(chunks, 1, char(2), char(3)) AS highlighted,
+        chunks.session_id
       FROM chunks
       JOIN sources ON sources.id = chunks.source_id
       WHERE chunks MATCH ? AND sources.label LIKE ? ESCAPE '\\'
@@ -537,7 +539,8 @@ export class ContentStore {
         chunks.timestamp,
         sources.label,
         bm25(chunks, 5.0, 1.0) AS rank,
-        highlight(chunks, 1, char(2), char(3)) AS highlighted
+        highlight(chunks, 1, char(2), char(3)) AS highlighted,
+        chunks.session_id
       FROM chunks
       JOIN sources ON sources.id = chunks.source_id
       WHERE chunks MATCH ? AND sources.label = ?
@@ -552,7 +555,8 @@ export class ContentStore {
         chunks_trigram.timestamp,
         sources.label,
         bm25(chunks_trigram, 5.0, 1.0) AS rank,
-        highlight(chunks_trigram, 1, char(2), char(3)) AS highlighted
+        highlight(chunks_trigram, 1, char(2), char(3)) AS highlighted,
+        chunks_trigram.session_id
       FROM chunks_trigram
       JOIN sources ON sources.id = chunks_trigram.source_id
       WHERE chunks_trigram MATCH ?
@@ -567,7 +571,8 @@ export class ContentStore {
         chunks_trigram.timestamp,
         sources.label,
         bm25(chunks_trigram, 5.0, 1.0) AS rank,
-        highlight(chunks_trigram, 1, char(2), char(3)) AS highlighted
+        highlight(chunks_trigram, 1, char(2), char(3)) AS highlighted,
+        chunks_trigram.session_id
       FROM chunks_trigram
       JOIN sources ON sources.id = chunks_trigram.source_id
       WHERE chunks_trigram MATCH ? AND sources.label LIKE ? ESCAPE '\\'
@@ -582,7 +587,8 @@ export class ContentStore {
         chunks_trigram.timestamp,
         sources.label,
         bm25(chunks_trigram, 5.0, 1.0) AS rank,
-        highlight(chunks_trigram, 1, char(2), char(3)) AS highlighted
+        highlight(chunks_trigram, 1, char(2), char(3)) AS highlighted,
+        chunks_trigram.session_id
       FROM chunks_trigram
       JOIN sources ON sources.id = chunks_trigram.source_id
       WHERE chunks_trigram MATCH ? AND sources.label = ?
@@ -598,7 +604,8 @@ export class ContentStore {
         chunks.timestamp,
         sources.label,
         bm25(chunks, 5.0, 1.0) AS rank,
-        highlight(chunks, 1, char(2), char(3)) AS highlighted
+        highlight(chunks, 1, char(2), char(3)) AS highlighted,
+        chunks.session_id
       FROM chunks
       JOIN sources ON sources.id = chunks.source_id
       WHERE chunks MATCH ? AND chunks.content_type = ?
@@ -613,7 +620,8 @@ export class ContentStore {
         chunks.timestamp,
         sources.label,
         bm25(chunks, 5.0, 1.0) AS rank,
-        highlight(chunks, 1, char(2), char(3)) AS highlighted
+        highlight(chunks, 1, char(2), char(3)) AS highlighted,
+        chunks.session_id
       FROM chunks
       JOIN sources ON sources.id = chunks.source_id
       WHERE chunks MATCH ? AND sources.label LIKE ? ESCAPE '\\' AND chunks.content_type = ?
@@ -628,7 +636,8 @@ export class ContentStore {
         chunks.timestamp,
         sources.label,
         bm25(chunks, 5.0, 1.0) AS rank,
-        highlight(chunks, 1, char(2), char(3)) AS highlighted
+        highlight(chunks, 1, char(2), char(3)) AS highlighted,
+        chunks.session_id
       FROM chunks
       JOIN sources ON sources.id = chunks.source_id
       WHERE chunks MATCH ? AND sources.label = ? AND chunks.content_type = ?
@@ -643,7 +652,8 @@ export class ContentStore {
         chunks_trigram.timestamp,
         sources.label,
         bm25(chunks_trigram, 5.0, 1.0) AS rank,
-        highlight(chunks_trigram, 1, char(2), char(3)) AS highlighted
+        highlight(chunks_trigram, 1, char(2), char(3)) AS highlighted,
+        chunks_trigram.session_id
       FROM chunks_trigram
       JOIN sources ON sources.id = chunks_trigram.source_id
       WHERE chunks_trigram MATCH ? AND chunks_trigram.content_type = ?
@@ -658,7 +668,8 @@ export class ContentStore {
         chunks_trigram.timestamp,
         sources.label,
         bm25(chunks_trigram, 5.0, 1.0) AS rank,
-        highlight(chunks_trigram, 1, char(2), char(3)) AS highlighted
+        highlight(chunks_trigram, 1, char(2), char(3)) AS highlighted,
+        chunks_trigram.session_id
       FROM chunks_trigram
       JOIN sources ON sources.id = chunks_trigram.source_id
       WHERE chunks_trigram MATCH ? AND sources.label LIKE ? ESCAPE '\\' AND chunks_trigram.content_type = ?
@@ -673,7 +684,8 @@ export class ContentStore {
         chunks_trigram.timestamp,
         sources.label,
         bm25(chunks_trigram, 5.0, 1.0) AS rank,
-        highlight(chunks_trigram, 1, char(2), char(3)) AS highlighted
+        highlight(chunks_trigram, 1, char(2), char(3)) AS highlighted,
+        chunks_trigram.session_id
       FROM chunks_trigram
       JOIN sources ON sources.id = chunks_trigram.source_id
       WHERE chunks_trigram MATCH ? AND sources.label = ? AND chunks_trigram.content_type = ?
@@ -902,6 +914,7 @@ export class ContentStore {
             contentType: r.content_type,
             highlighted: r.highlighted,
             timestamp: r.timestamp ?? undefined,
+            sessionId: r.session_id ?? "",
         }));
     }
     #sourceFilterParam(source, sourceMatchMode) {
@@ -1088,13 +1101,21 @@ export class ContentStore {
             .map(({ result }) => result);
     }
     // ── Unified Fallback Search ──
-    searchWithFallback(query, limit = 3, source, contentType, sourceMatchMode = "like") {
+    searchWithFallback(query, limit = 3, source, contentType, sourceMatchMode = "like", sessionIdAllowSet) {
         // Step 0: Auto-refresh stale file-backed sources before searching
         this.#refreshStaleSources();
+        // When a session-id allow-set is in play (issue #737 project filter),
+        // fetch a larger candidate pool from the FTS5 layers so the post-filter
+        // can still deliver `limit` matches even if many candidates are excluded.
+        // The cap is bounded — even at the largest installs the chunk count
+        // dwarfs `limit * 8`, and the surplus is dropped on the post-filter.
+        const fetchLimit = sessionIdAllowSet ? Math.max(limit * 8, 40) : limit;
+        const sessionFilter = this.#makeSessionFilter(sessionIdAllowSet);
         // Step 1: RRF fusion (porter OR + trigram OR → merge)
-        const rrfResults = this.#rrfSearch(query, limit, source, contentType, sourceMatchMode);
-        if (rrfResults.length > 0) {
-            const reranked = this.#applyProximityReranking(rrfResults, query);
+        const rrfResults = this.#rrfSearch(query, fetchLimit, source, contentType, sourceMatchMode);
+        const rrfFiltered = sessionFilter ? rrfResults.filter(sessionFilter) : rrfResults;
+        if (rrfFiltered.length > 0) {
+            const reranked = this.#applyProximityReranking(rrfFiltered.slice(0, limit), query);
             return reranked.map((r) => ({ ...r, matchLayer: "rrf" }));
         }
         // Step 2: Fuzzy correction → RRF re-run
@@ -1109,14 +1130,29 @@ export class ContentStore {
         const correctedWords = words.map((w) => this.fuzzyCorrect(w) ?? w);
         const correctedQuery = correctedWords.join(" ");
         if (correctedQuery !== original) {
-            const fuzzyResults = this.#rrfSearch(correctedQuery, limit, source, contentType, sourceMatchMode);
-            if (fuzzyResults.length > 0) {
-                const reranked = this.#applyProximityReranking(fuzzyResults, correctedQuery);
+            const fuzzyResults = this.#rrfSearch(correctedQuery, fetchLimit, source, contentType, sourceMatchMode);
+            const fuzzyFiltered = sessionFilter ? fuzzyResults.filter(sessionFilter) : fuzzyResults;
+            if (fuzzyFiltered.length > 0) {
+                const reranked = this.#applyProximityReranking(fuzzyFiltered.slice(0, limit), correctedQuery);
                 return reranked.map((r) => ({ ...r, matchLayer: "rrf-fuzzy" }));
             }
         }
         return [];
     }
+    /**
+     * Build the session-id post-filter for the FTS5 candidate pool. Legacy
+     * chunks indexed before per-session attribution carry `session_id=''` and
+     * stay visible across projects so user-indexed content remains reachable
+     * after opting into the shared-DB mode (#737).
+     */
+    #makeSessionFilter(allowSet) {
+        if (!allowSet)
+            return null;
+        return (r) => {
+            const sid = r.sessionId ?? "";
+            return sid === "" || allowSet.has(sid);
+        };
+    }
     /** Number of sources auto-refreshed in the last searchWithFallback call. */
     lastRefreshCount = 0;
     /**
@@ -1181,6 +1217,22 @@ export class ContentStore {
     listSources() {
         return this.#stmtListSources.all();
     }
+    /**
+     * Aggregate snapshot of the persistent content store. Returns total
+     * chunk count, source count, and the most recent indexed_at timestamp.
+     * Used by ctx_stats so callers can see observability state in the same
+     * round trip instead of inferring it from snapshot diffs.
+     */
+    getIndexState() {
+        const row = this.#db
+            .prepare("SELECT COALESCE(SUM(chunk_count), 0) AS total_chunks, COUNT(*) AS total_sources, MAX(indexed_at) AS last_indexed_at FROM sources")
+            .get();
+        return {
+            totalChunks: row.total_chunks ?? 0,
+            totalSources: row.total_sources ?? 0,
+            lastIndexedAt: row.last_indexed_at ?? undefined,
+        };
+    }
     /**
      * Get all chunks for a given source by ID — bypasses FTS5 MATCH entirely.
      * Use this for inventory/listing where you need all sections, not search.

package/build/types.d.ts

@@ -75,6 +75,13 @@ export interface SearchResult {
     matchLayer?: "porter" | "trigram" | "fuzzy" | "rrf" | "rrf-fuzzy";
     highlighted?: string;
     timestamp?: string;
+    /**
+     * Session-id attribution copied from the `chunks.session_id` column
+     * (legacy unattributed chunks carry an empty string). Used by the
+     * `ctx_search` per-project filter (#737) to scope shared-DB results
+     * to the current project via the 2-step IN-clause strategy.
+     */
+    sessionId?: string;
 }
 /**
  * Aggregate statistics for a ContentStore instance.

package/build/util/project-dir.d.ts

@@ -18,12 +18,13 @@ import type { PlatformId } from "../adapters/types.js";
  *     (shell-set, survives `process.chdir`) before falling back.
  */
 /**
- * Detect whether a path lives inside the Claude Code plugin install tree —
- * specifically `<home>/.claude/plugins/cache/<plugin>/<plugin>/<version>/`
- * or the marketplace mirror `<home>/.claude/plugins/marketplaces/...`.
+ * Detect whether a path lives inside an agent plugin install tree —
+ * specifically `<home>/.claude/plugins/cache/<plugin>/<plugin>/<version>/`,
+ * `<home>/.codex/plugins/cache/<plugin>/<plugin>/<version>/`, or the
+ * marketplace mirror under `<home>/.{claude,codex}/plugins/marketplaces/...`.
  *
  * Cross-OS: matches both POSIX (`/`) and Windows (`\`) path separators.
- * Independent of `home` location — we only care about the `.claude/plugins/`
+ * Independent of `home` location — we only care about the agent plugin
  * suffix pattern.
  */
 export declare function isPluginInstallPath(p: string): boolean;
@@ -65,24 +66,26 @@ export declare function resolveProjectDirFromTranscript(opts: {
  * session log when the spawned MCP child inherits a non-project cwd
  * (e.g. $HOME when Codex was launched from anywhere outside the project).
  *
- * Codex writes its session transcripts to
- * `${CODEX_HOME ?? ~/.codex}/sessions/<uuid>.jsonl`. The first line is a
- * `SessionMeta` JSON struct whose `meta.cwd` field carries the literal
- * project directory the CLI was launched from (see refs/platforms/codex/
- * codex-rs SessionMeta). Codex publishes NO workspace env var to its child
- * MCP processes — so unlike Claude/Pi/Cursor, we have no env signal at all.
- * The session log is the strongest available signal.
+ * Codex writes its session transcripts to either
+ * `${CODEX_HOME ?? ~/.codex}/sessions/<uuid>.jsonl` (CLI) or a dated desktop
+ * layout such as
+ * `${CODEX_HOME ?? ~/.codex}/sessions/YYYY/MM/DD/rollout-*.jsonl`.
+ * The cwd appears on `meta.cwd` for the CLI shape and on
+ * `payload.cwd` in `type: "session_meta"` records for Codex Desktop. Codex
+ * publishes NO workspace env var to its child MCP processes — so unlike
+ * Claude/Pi/Cursor, we have no env signal at all. The session log is the
+ * strongest available signal.
  *
  * Mirror of `resolveProjectDirFromTranscript` for Claude Code; differences:
- *   • Sessions live flat in `${codexHome}/sessions/*.jsonl` (no per-project
- *     encoded subdir like Claude's `~/.claude/projects/<encoded>/`).
- *   • The cwd is on `meta.cwd` (nested), not top-level `cwd`.
+ *   • Sessions may live flat or in a dated hierarchy (no per-project encoded
+ *     subdir like Claude's `~/.claude/projects/<encoded>/`).
+ *   • The cwd is nested on `meta.cwd` or `payload.cwd`, not top-level `cwd`.
  *
  * Returns `null` when:
  *   • `codexHome` or its `sessions/` subdir does not exist.
- *   • No `.jsonl` files exist or none has a parseable `meta.cwd` string.
+ *   • No `.jsonl` files exist or none has a parseable cwd string.
  *   • The newest log is older than `transcriptMaxAgeMs` (multi-window guard).
- *   • The resolved `meta.cwd` points at a plugin install path (poisoned).
+ *   • The resolved cwd points at a plugin install path (poisoned).
  */
 export declare function resolveCodexSessionCwd(opts?: {
     /** Defaults to `process.env.CODEX_HOME ?? path.join(os.homedir(), ".codex")`. */

package/build/util/project-dir.js

@@ -47,18 +47,19 @@ const LEGACY_NON_STRICT_CANDIDATES = [
  *     (shell-set, survives `process.chdir`) before falling back.
  */
 /**
- * Detect whether a path lives inside the Claude Code plugin install tree —
- * specifically `<home>/.claude/plugins/cache/<plugin>/<plugin>/<version>/`
- * or the marketplace mirror `<home>/.claude/plugins/marketplaces/...`.
+ * Detect whether a path lives inside an agent plugin install tree —
+ * specifically `<home>/.claude/plugins/cache/<plugin>/<plugin>/<version>/`,
+ * `<home>/.codex/plugins/cache/<plugin>/<plugin>/<version>/`, or the
+ * marketplace mirror under `<home>/.{claude,codex}/plugins/marketplaces/...`.
  *
  * Cross-OS: matches both POSIX (`/`) and Windows (`\`) path separators.
- * Independent of `home` location — we only care about the `.claude/plugins/`
+ * Independent of `home` location — we only care about the agent plugin
  * suffix pattern.
  */
 export function isPluginInstallPath(p) {
     if (!p)
         return false;
-    return /[/\\]\.claude[/\\]plugins[/\\](cache|marketplaces)[/\\]/.test(p);
+    return /[/\\]\.(claude|codex)[/\\]plugins[/\\](cache|marketplaces)[/\\]/.test(p);
 }
 /**
  * Read the per-session project dir from Claude Code's transcript files.
@@ -162,46 +163,76 @@ export function resolveProjectDirFromTranscript(opts) {
  * session log when the spawned MCP child inherits a non-project cwd
  * (e.g. $HOME when Codex was launched from anywhere outside the project).
  *
- * Codex writes its session transcripts to
- * `${CODEX_HOME ?? ~/.codex}/sessions/<uuid>.jsonl`. The first line is a
- * `SessionMeta` JSON struct whose `meta.cwd` field carries the literal
- * project directory the CLI was launched from (see refs/platforms/codex/
- * codex-rs SessionMeta). Codex publishes NO workspace env var to its child
- * MCP processes — so unlike Claude/Pi/Cursor, we have no env signal at all.
- * The session log is the strongest available signal.
+ * Codex writes its session transcripts to either
+ * `${CODEX_HOME ?? ~/.codex}/sessions/<uuid>.jsonl` (CLI) or a dated desktop
+ * layout such as
+ * `${CODEX_HOME ?? ~/.codex}/sessions/YYYY/MM/DD/rollout-*.jsonl`.
+ * The cwd appears on `meta.cwd` for the CLI shape and on
+ * `payload.cwd` in `type: "session_meta"` records for Codex Desktop. Codex
+ * publishes NO workspace env var to its child MCP processes — so unlike
+ * Claude/Pi/Cursor, we have no env signal at all. The session log is the
+ * strongest available signal.
  *
  * Mirror of `resolveProjectDirFromTranscript` for Claude Code; differences:
- *   • Sessions live flat in `${codexHome}/sessions/*.jsonl` (no per-project
- *     encoded subdir like Claude's `~/.claude/projects/<encoded>/`).
- *   • The cwd is on `meta.cwd` (nested), not top-level `cwd`.
+ *   • Sessions may live flat or in a dated hierarchy (no per-project encoded
+ *     subdir like Claude's `~/.claude/projects/<encoded>/`).
+ *   • The cwd is nested on `meta.cwd` or `payload.cwd`, not top-level `cwd`.
  *
  * Returns `null` when:
  *   • `codexHome` or its `sessions/` subdir does not exist.
- *   • No `.jsonl` files exist or none has a parseable `meta.cwd` string.
+ *   • No `.jsonl` files exist or none has a parseable cwd string.
  *   • The newest log is older than `transcriptMaxAgeMs` (multi-window guard).
- *   • The resolved `meta.cwd` points at a plugin install path (poisoned).
+ *   • The resolved cwd points at a plugin install path (poisoned).
  */
 export function resolveCodexSessionCwd(opts) {
     const codexHome = opts?.codexHome ?? process.env.CODEX_HOME ?? path.join(os.homedir(), ".codex");
     const sessionsDir = path.join(codexHome, "sessions");
     if (!fs.existsSync(sessionsDir))
         return null;
+    const MAX_SCAN_DEPTH = 4; // sessions/YYYY/MM/DD/<file>.jsonl plus one spare.
+    const MAX_SCAN_ENTRIES = 10_000;
+    let visitedEntries = 0;
     let bestPath;
     let bestMtime = 0;
-    try {
-        for (const f of fs.readdirSync(sessionsDir)) {
-            if (!f.endsWith(".jsonl"))
-                continue;
-            const fp = path.join(sessionsDir, f);
+    const visit = (dir, depth) => {
+        if (visitedEntries >= MAX_SCAN_ENTRIES)
+            return;
+        let entries;
+        try {
+            entries = fs.readdirSync(dir);
+        }
+        catch {
+            return;
+        }
+        entries.sort().reverse();
+        for (const entry of entries) {
+            if (visitedEntries >= MAX_SCAN_ENTRIES)
+                return;
+            visitedEntries++;
+            const fp = path.join(dir, entry);
+            let stat;
             try {
-                const m = fs.statSync(fp).mtimeMs;
-                if (m > bestMtime) {
-                    bestMtime = m;
-                    bestPath = fp;
-                }
+                stat = fs.statSync(fp);
+            }
+            catch {
+                continue;
+            }
+            if (stat.isDirectory()) {
+                if (depth < MAX_SCAN_DEPTH)
+                    visit(fp, depth + 1);
+                continue;
+            }
+            if (!stat.isFile() || !entry.endsWith(".jsonl"))
+                continue;
+            const m = stat.mtimeMs;
+            if (m > bestMtime) {
+                bestMtime = m;
+                bestPath = fp;
             }
-            catch { /* skip */ }
         }
+    };
+    try {
+        visit(sessionsDir, 0);
     }
     catch {
         return null;
@@ -213,28 +244,31 @@ export function resolveCodexSessionCwd(opts) {
         if (nowMs - bestMtime > opts.transcriptMaxAgeMs)
             return null;
     }
-    // Read first ~8KB; the SessionMeta JSON is line 1 and small. Stream-cap
-    // mirrors `resolveProjectDirFromTranscript` for memory safety on long logs.
+    // Read a bounded head chunk. Codex Desktop's first session_meta line can be
+    // larger than Claude/Codex CLI metadata because it includes dynamic tool and
+    // instruction fields, but the full transcript can still be tens of MB.
     try {
         const fd = fs.openSync(bestPath, "r");
         try {
-            const buf = Buffer.alloc(8192);
+            const buf = Buffer.alloc(1024 * 1024);
             const bytes = fs.readSync(fd, buf, 0, buf.length, 0);
             const text = buf.subarray(0, bytes).toString("utf-8");
-            const firstLine = text.split("\n", 1)[0];
-            if (!firstLine || !firstLine.trim())
-                return null;
-            try {
-                const obj = JSON.parse(firstLine);
-                const cwd = obj?.meta?.cwd;
-                if (typeof cwd !== "string" || cwd.length === 0)
-                    return null;
-                if (isPluginInstallPath(cwd))
-                    return null;
-                return cwd;
-            }
-            catch {
-                return null; /* malformed first line */
+            for (const line of text.split("\n").slice(0, 10)) {
+                if (!line.trim())
+                    continue;
+                try {
+                    const obj = JSON.parse(line);
+                    const cwd = obj?.meta?.cwd ??
+                        (obj?.type === "session_meta" ? obj?.payload?.cwd : undefined);
+                    if (typeof cwd !== "string" || cwd.length === 0)
+                        continue;
+                    if (isPluginInstallPath(cwd))
+                        return null;
+                    return cwd;
+                }
+                catch {
+                    return null; /* malformed session metadata line */
+                }
             }
         }
         finally {
@@ -244,6 +278,7 @@ export function resolveCodexSessionCwd(opts) {
     catch {
         return null; /* file vanished mid-read */
     }
+    return null;
 }
 /**
  * Pure project-dir resolver. Mirror of the env-var chain inside