Merge pull request #215 from timkjr/fix/past-sessions-history-quality

fix(history): three Past Sessions data-quality bugs (automated-session noise, cross-contaminated previews, blank restart-heavy rows)

src/services/unified-session-service.ts

@@ -226,6 +226,16 @@ export function mergeUnifiedSessions(sources: UnifiedSources): UnifiedSessionIte
   // different UUID. Backfill from the already-passed history: first try the claudeSessionId
   // join, then the newest transcript in the same workingDir. Never overwrite a non-empty
   // firstPrompt (so rows keyed to their own transcript are untouched).
+  //
+  // The workingDir guess is a last resort and MUST be skipped for any item that
+  // already has its own 'history' entry (step 1 above already gave it a real,
+  // direct scan of its own transcript). Without this guard, a history row whose
+  // OWN extraction genuinely failed (oversized first message, etc.) silently
+  // inherited the newest OTHER session's opening line from the same directory —
+  // not a blank, but actively wrong: old sessions displayed today's conversation
+  // as if it were their own. A row with no 'history' source at all (its
+  // transcript hasn't been linked/scanned under its own id yet) has no such
+  // direct attempt to prefer, so the guess remains a reasonable stand-in there.
   const firstPromptByUuid = new Map<string, string>();
   const firstPromptByWorkingDir = new Map<string, { prompt: string; ms: number }>();
   // COD-145: lastPrompt rides the same backfill (build parallel indexes; never overwrite).
@@ -254,12 +264,13 @@ export function mergeUnifiedSessions(sources: UnifiedSources): UnifiedSessionIte
     }
   }
   for (const item of map.values()) {
+    const hasOwnHistoryEntry = item.sources.includes('history');
     if (!item.firstPrompt) {
       // never overwrite an existing non-empty prompt
       const byUuid = item.claudeSessionId ? firstPromptByUuid.get(item.claudeSessionId) : undefined;
       if (byUuid) {
         item.firstPrompt = byUuid;
-      } else if (item.workingDir) {
+      } else if (item.workingDir && !hasOwnHistoryEntry) {
         const byDir = firstPromptByWorkingDir.get(item.workingDir);
         if (byDir) item.firstPrompt = byDir.prompt;
       }
@@ -268,7 +279,7 @@ export function mergeUnifiedSessions(sources: UnifiedSources): UnifiedSessionIte
       const byUuid = item.claudeSessionId ? lastPromptByUuid.get(item.claudeSessionId) : undefined;
       if (byUuid) {
         item.lastPrompt = byUuid;
-      } else if (item.workingDir) {
+      } else if (item.workingDir && !hasOwnHistoryEntry) {
         const byDir = lastPromptByWorkingDir.get(item.workingDir);
         if (byDir) item.lastPrompt = byDir.prompt;
       }

src/web/routes/session-routes.ts

@@ -2478,6 +2478,46 @@ export function registerSessionRoutes(
     return undefined;
   }
 
+  /**
+   * The `entrypoint` field Claude Code stamps on its own message records:
+   * 'cli' for a real interactive session, something else (e.g. 'sdk-py') for
+   * an SDK/automated invocation. Used to exclude non-interactive transcripts
+   * (CI review bots, etc.) from the resumable history list — they were never
+   * something a user can resume into.
+   *
+   * Scans every `"type":"user"`/`"type":"assistant"` line with an entrypoint
+   * field — not just the first one — and returns 'cli' the moment ANY of them
+   * carries it. A transcript is excluded only when every entrypoint-bearing
+   * message says something else; "first field wins" would misattribute a
+   * transcript that started under an older Claude Code version (no entrypoint
+   * on its true first message) and later picked up a non-'cli' entrypoint on
+   * some later message, wrongly hiding a genuinely interactive session. This
+   * deliberately errs toward keeping a session visible: one real interactive
+   * message anywhere is enough. Returns undefined ("unknown", fail-open) only
+   * when nothing scanned carries the field at all.
+   */
+  function extractTranscriptEntrypoint(text: string): string | undefined {
+    let start = 0;
+    let sawNonCli: string | undefined;
+    while (start < text.length) {
+      const end = text.indexOf('\n', start);
+      const line = end === -1 ? text.slice(start) : text.slice(start, end);
+      start = end === -1 ? text.length : end + 1;
+      if (!line.includes('"type":"user"') && !line.includes('"type":"assistant"')) continue;
+      if (!line.includes('"entrypoint"')) continue;
+      try {
+        const entry = JSON.parse(line);
+        if ((entry.type === 'user' || entry.type === 'assistant') && typeof entry.entrypoint === 'string') {
+          if (entry.entrypoint === 'cli') return 'cli';
+          sawNonCli ??= entry.entrypoint;
+        }
+      } catch {
+        // Malformed/truncated line — skip
+      }
+    }
+    return sawNonCli;
+  }
+
   /**
    * Extract the text of the LAST user message from a JSONL transcript chunk
    * (COD-145). Mirrors `extractFirstUserPrompt` exactly — same user-message
@@ -2690,7 +2730,7 @@ export function registerSessionRoutes(
     return finalExists ? current : process.env.HOME || '/tmp';
   }
 
-  /** Read the first 16KB of a file for content sniffing. */
+  /** Read the first `buf.length` bytes of a file for content sniffing. */
   async function readFileHead(path: string, buf: Buffer): Promise<string | null> {
     try {
       const fd = await fs.open(path, 'r');
@@ -2733,7 +2773,12 @@ export function registerSessionRoutes(
 
   // Scan a single project directory and return all valid history sessions in it.
   // Reused by both the global overview and the single-folder drill-down.
-  async function scanProjectDir(projPath: string, projDir: string, headBuf: Buffer): Promise<HistorySession[]> {
+  async function scanProjectDir(
+    projPath: string,
+    projDir: string,
+    smallHeadBuf: Buffer,
+    headBuf: Buffer
+  ): Promise<HistorySession[]> {
     const out: HistorySession[] = [];
     const stat = await fs.stat(projPath).catch(() => null);
     if (!stat?.isDirectory()) return out;
@@ -2751,22 +2796,45 @@ export function registerSessionRoutes(
       if (!fileStat) continue;
       if (fileStat.size < 4000) continue;
 
-      let firstPrompt: string | undefined;
-      const head = await readFileHead(filePath, headBuf);
       const hasConversation = (text: string) =>
         text.includes('"type":"user"') || text.includes('"type":"assistant"') || text.includes('"type":"summary"');
 
+      // Two-tier head read: try the cheap smallHeadBuf (16KB) size first -- enough
+      // for the vast majority of transcripts -- and only escalate to the full
+      // headBuf (128KB) when that wasn't enough. Reading 128KB unconditionally for
+      // EVERY file in the directory roughly quadrupled the cost of a full scan
+      // (measured against a real ~/.claude/projects tree: ~4x both bytes read and
+      // wall time) to fix a problem only ~28% of files actually have. Escalating
+      // resolves the restart-bookkeeping case (the reason 128KB exists at all)
+      // without ever touching the tail-read fallback below for most of that 28%.
+      let head = await readFileHead(filePath, smallHeadBuf);
       let foundContent = head ? hasConversation(head) : false;
+      let firstPrompt = head ? extractFirstUserPrompt(head) : undefined;
+      if ((!foundContent || !firstPrompt) && head !== null && fileStat.size > smallHeadBuf.length) {
+        const biggerHead = await readFileHead(filePath, headBuf);
+        if (biggerHead) {
+          head = biggerHead;
+          if (!foundContent) foundContent = hasConversation(head);
+          if (!firstPrompt) firstPrompt = extractFirstUserPrompt(head);
+        }
+      }
+
       let tail: string | null = null;
-      if (!foundContent && fileStat.size > 16384) {
+      // `head === null` (a failed read -- e.g. EMFILE while scanning hundreds of
+      // files) must also get a shot at the tail, not just "file bigger than the
+      // head buffer". Losing this dropped the session from history entirely
+      // instead of giving it a second chance, for any file at or under the head
+      // buffer size whose head read happened to fail.
+      if (!foundContent && (head === null || fileStat.size > headBuf.length)) {
         const tailBuf = Buffer.alloc(32768);
         tail = await readFileTail(filePath, tailBuf, fileStat.size);
         if (tail) foundContent = hasConversation(tail);
       }
       if (!foundContent) continue;
 
-      if (head) firstPrompt = extractFirstUserPrompt(head);
-      if (!firstPrompt && fileStat.size > 65536) {
+      // firstPrompt was already attempted from head (both tiers) above; this is
+      // purely the tail fallback for whatever's left unresolved.
+      if (!firstPrompt && (head === null || fileStat.size > headBuf.length)) {
         if (!tail) {
           const tailBuf = Buffer.alloc(32768);
           tail = await readFileTail(filePath, tailBuf, fileStat.size);
@@ -2776,15 +2844,37 @@ export function registerSessionRoutes(
 
       // COD-145: last (most recent) user prompt lives near the END of the file, so
       // prefer the tail. For large files where no tail was read yet, read one
-      // (mirrors the firstPrompt > 65536 block). Small files fit in `head`, which
-      // then contains the whole transcript — scan it for the last match instead.
-      if (!tail && fileStat.size > 65536) {
+      // (mirrors the firstPrompt > headBuf.length block). Small files fit in `head`,
+      // which then contains the whole transcript — scan it for the last match instead.
+      if (!tail && fileStat.size > headBuf.length) {
         const tailBuf = Buffer.alloc(32768);
         tail = await readFileTail(filePath, tailBuf, fileStat.size);
       }
       const lastPrompt =
         (tail ? extractLastUserPrompt(tail) : undefined) ?? (head ? extractLastUserPrompt(head) : undefined);
 
+      // Automated/SDK-driven invocations (CI review bots, etc.) write transcripts
+      // into the same ~/.claude/projects tree as interactive sessions but were
+      // never something a user can resume into — no PTY, no running process, and
+      // their "conversation" is typically a single one-shot prompt (often with a
+      // full diff embedded, which is exactly why it dwarfs this scanner's read
+      // windows and shows up above as blank or as an identical boilerplate
+      // sentence across many rows). Checked last, so it reuses whatever `head`/
+      // `tail` the prompt extraction above already read rather than triggering
+      // an extra file read. Missing entrypoint (older transcripts) reads as
+      // interactive — fail open, matching every other gating check in this
+      // codebase.
+      //
+      // head and tail are checked independently and merged with "cli wins" (not
+      // a first-truthy-value `??` chain): a large file's head might land on a
+      // non-'cli' message while a real interactive message sits in the tail (or
+      // vice versa), and either one being 'cli' is enough to keep the session.
+      const headEntrypoint = head ? extractTranscriptEntrypoint(head) : undefined;
+      const tailEntrypoint = tail ? extractTranscriptEntrypoint(tail) : undefined;
+      const entrypoint =
+        headEntrypoint === 'cli' || tailEntrypoint === 'cli' ? 'cli' : (headEntrypoint ?? tailEntrypoint);
+      if (entrypoint && entrypoint !== 'cli') continue;
+
       out.push({
         sessionId,
         workingDir,
@@ -2801,7 +2891,13 @@ export function registerSessionRoutes(
   app.get('/api/history/sessions', async (req) => {
     const query = req.query as { projectKey?: string; offset?: string; limit?: string };
     const projectsDir = join(process.env.HOME || '/tmp', '.claude', 'projects');
-    const headBuf = Buffer.alloc(16384);
+    // scanProjectDir tries smallHeadBuf (16KB, the original size) first for every
+    // file and only escalates to headBuf (128KB) when that wasn't enough — see the
+    // comment at the escalation site in scanProjectDir for why unconditional 128KB
+    // reads were too expensive to keep. 128KB matches the existing precedent
+    // elsewhere in this file (line ~1431).
+    const smallHeadBuf = Buffer.alloc(16384);
+    const headBuf = Buffer.alloc(131072);
     // Multi-user: this scans the host-wide ~/.claude/projects tree, so a non-admin
     // must only see history whose decoded workingDir is inside their own case space.
     // Do NOT trust the caller-supplied projectKey — confine on the decoded path.
@@ -2819,7 +2915,7 @@ export function registerSessionRoutes(
       const offset = Math.max(0, parseInt(query.offset || '0', 10) || 0);
       const limit = Math.min(100, Math.max(1, parseInt(query.limit || '20', 10) || 20));
       const projPath = join(projectsDir, query.projectKey);
-      let all = await scanProjectDir(projPath, query.projectKey, headBuf);
+      let all = await scanProjectDir(projPath, query.projectKey, smallHeadBuf, headBuf);
       // Confine to the caller's workspace (a projectKey maps to a single foreign cwd).
       if (scopeHistory) all = all.filter((r) => isWorkingDirAllowed(user, r.workingDir));
       all.sort((a, b) => new Date(b.lastModified).getTime() - new Date(a.lastModified).getTime());
@@ -2832,7 +2928,7 @@ export function registerSessionRoutes(
       const projectDirs = await fs.readdir(projectsDir);
       for (const projDir of projectDirs) {
         const projPath = join(projectsDir, projDir);
-        const list = await scanProjectDir(projPath, projDir, headBuf);
+        const list = await scanProjectDir(projPath, projDir, smallHeadBuf, headBuf);
         results.push(...list);
       }
     } catch {
@@ -2908,11 +3004,13 @@ export function registerSessionRoutes(
     const history: HistoryInput[] = [];
     try {
       const projectsDir = join(process.env.HOME || '/tmp', '.claude', 'projects');
-      const headBuf = Buffer.alloc(16384);
+      // See the sibling allocation above for why there are two sizes.
+      const smallHeadBuf = Buffer.alloc(16384);
+      const headBuf = Buffer.alloc(131072);
       const projectDirs = await fs.readdir(projectsDir);
       for (const projDir of projectDirs) {
         const projPath = join(projectsDir, projDir);
-        const list = await scanProjectDir(projPath, projDir, headBuf);
+        const list = await scanProjectDir(projPath, projDir, smallHeadBuf, headBuf);
         for (const h of list) {
           history.push({
             sessionId: h.sessionId,

test/routes/session-routes.test.ts

@@ -1350,6 +1350,224 @@ describe('session-routes', () => {
       expect(row.workingDir).toBe(dotDir);
       expect(row.workingDir).not.toContain('//');
     });
+
+    it('excludes non-interactive (SDK-driven) transcripts from the history list', async () => {
+      // CI review bots and other automated tools write transcripts into the same
+      // ~/.claude/projects tree as interactive sessions (entrypoint "sdk-py" etc.)
+      // but were never something a user can resume into — no PTY, no running
+      // process. They cluttered Past Sessions as blank rows or identical
+      // boilerplate ("Review this change for security vulnerabilities...").
+      const home = process.env.HOME as string;
+      const projPath = join(home, '.claude', 'projects', 'proj-entrypoint-test');
+      await mkdir(projPath, { recursive: true });
+
+      const cliId = '33333333-3333-3333-3333-333333333333';
+      const sdkId = '44444444-4444-4444-4444-444444444444';
+      const noEntrypointId = '55555555-5555-5555-5555-555555555555';
+
+      const cliLine =
+        JSON.stringify({ type: 'user', entrypoint: 'cli', message: { role: 'user', content: 'a real question' } }) +
+        '\n';
+      const sdkLine =
+        JSON.stringify({
+          type: 'user',
+          entrypoint: 'sdk-py',
+          message: { role: 'user', content: 'Review this change for security vulnerabilities.' },
+        }) + '\n';
+      // Older transcripts predate the entrypoint field entirely — must still show.
+      const noEntrypointLine =
+        JSON.stringify({ type: 'user', message: { role: 'user', content: 'a pre-entrypoint session' } }) + '\n';
+
+      await writeFile(join(projPath, `${cliId}.jsonl`), cliLine + '#'.repeat(4200 - cliLine.length));
+      await writeFile(join(projPath, `${sdkId}.jsonl`), sdkLine + '#'.repeat(4200 - sdkLine.length));
+      await writeFile(
+        join(projPath, `${noEntrypointId}.jsonl`),
+        noEntrypointLine + '#'.repeat(4200 - noEntrypointLine.length)
+      );
+
+      const res = await harness.app.inject({
+        method: 'GET',
+        url: '/api/history/sessions?projectKey=proj-entrypoint-test',
+      });
+      expect(res.statusCode).toBe(200);
+      const ids = JSON.parse(res.body).data.sessions.map((s: { sessionId: string }) => s.sessionId);
+      expect(ids).toContain(cliId);
+      expect(ids).toContain(noEntrypointId);
+      expect(ids).not.toContain(sdkId);
+    });
+
+    it('ignores a bookkeeping line that happens to mention "entrypoint" outside a real message record', async () => {
+      // Scanning must anchor on "type":"user"/"assistant" lines specifically,
+      // not any line that happens to contain the substring "entrypoint".
+      const home = process.env.HOME as string;
+      const projPath = join(home, '.claude', 'projects', 'proj-entrypoint-bookkeeping-test');
+      await mkdir(projPath, { recursive: true });
+
+      const sessionId = '77777777-7777-7777-7777-777777777777';
+      const bookkeepingLine = JSON.stringify({ type: 'mode', mode: 'normal', entrypoint: 'sdk-py' }) + '\n';
+      const realLine =
+        JSON.stringify({ type: 'user', entrypoint: 'cli', message: { role: 'user', content: 'a real message' } }) +
+        '\n';
+
+      // scanProjectDir skips files under 4000 bytes.
+      const body = bookkeepingLine + realLine;
+      await writeFile(join(projPath, `${sessionId}.jsonl`), body + '#'.repeat(4200 - body.length));
+
+      const res = await harness.app.inject({
+        method: 'GET',
+        url: '/api/history/sessions?projectKey=proj-entrypoint-bookkeeping-test',
+      });
+      expect(res.statusCode).toBe(200);
+      const ids = JSON.parse(res.body).data.sessions.map((s: { sessionId: string }) => s.sessionId);
+      expect(ids).toContain(sessionId);
+    });
+
+    it('shows a session with ANY interactive (cli) message, even if an earlier message was automated', async () => {
+      // "First field wins" would have misattributed this: an old transcript
+      // whose true first message predates the entrypoint field, later resumed
+      // under something automated (entrypoint: 'sdk-py' on message 2), then
+      // continued interactively by a real person (entrypoint: 'cli' on message
+      // 3). Stopping at the first entrypoint-bearing line found ('sdk-py')
+      // would wrongly exclude a session a human genuinely used. One real
+      // interactive message anywhere is enough to keep it visible.
+      const home = process.env.HOME as string;
+      const projPath = join(home, '.claude', 'projects', 'proj-entrypoint-any-cli-test');
+      await mkdir(projPath, { recursive: true });
+
+      const sessionId = '99999999-9999-9999-9999-999999999999';
+      const firstLine =
+        JSON.stringify({ type: 'user', message: { role: 'user', content: 'pre-entrypoint-field message' } }) + '\n';
+      const automatedLine =
+        JSON.stringify({
+          type: 'user',
+          entrypoint: 'sdk-py',
+          message: { role: 'user', content: 'an automated follow-up' },
+        }) + '\n';
+      const interactiveLine =
+        JSON.stringify({
+          type: 'user',
+          entrypoint: 'cli',
+          message: { role: 'user', content: 'a real person continued this' },
+        }) + '\n';
+
+      const body = firstLine + automatedLine + interactiveLine;
+      await writeFile(join(projPath, `${sessionId}.jsonl`), body + '#'.repeat(4200 - body.length));
+
+      const res = await harness.app.inject({
+        method: 'GET',
+        url: '/api/history/sessions?projectKey=proj-entrypoint-any-cli-test',
+      });
+      expect(res.statusCode).toBe(200);
+      const ids = JSON.parse(res.body).data.sessions.map((s: { sessionId: string }) => s.sessionId);
+      expect(ids).toContain(sessionId);
+    });
+
+    it('still excludes a session where every entrypoint-bearing message is automated', async () => {
+      // Mirror of the previous test with no 'cli' message anywhere — proves the
+      // "any cli wins" fix isn't just failing open unconditionally.
+      const home = process.env.HOME as string;
+      const projPath = join(home, '.claude', 'projects', 'proj-entrypoint-all-automated-test');
+      await mkdir(projPath, { recursive: true });
+
+      const sessionId = 'aaaaaaaa-aaaa-aaaa-aaaa-aaaaaaaaaaaa';
+      const firstLine =
+        JSON.stringify({
+          type: 'user',
+          entrypoint: 'sdk-py',
+          message: { role: 'user', content: 'Review this change for security vulnerabilities.' },
+        }) + '\n';
+      const secondLine =
+        JSON.stringify({
+          type: 'assistant',
+          entrypoint: 'sdk-py',
+          message: { role: 'assistant', content: [{ type: 'text', text: 'Looking at the diff...' }] },
+        }) + '\n';
+
+      const body = firstLine + secondLine;
+      await writeFile(join(projPath, `${sessionId}.jsonl`), body + '#'.repeat(4200 - body.length));
+
+      const res = await harness.app.inject({
+        method: 'GET',
+        url: '/api/history/sessions?projectKey=proj-entrypoint-all-automated-test',
+      });
+      expect(res.statusCode).toBe(200);
+      const ids = JSON.parse(res.body).data.sessions.map((s: { sessionId: string }) => s.sessionId);
+      expect(ids).not.toContain(sessionId);
+    });
+
+    it('finds the real first prompt past a large run of pre-message bookkeeping lines', async () => {
+      // A session restarted many times over a long conversation accumulates a batch
+      // of small bookkeeping lines (mode/permission-mode/last-prompt/queue-operation)
+      // per restart, ahead of the real first message. With enough restarts these can
+      // push the genuine first prompt past a 16KB head-read window even though the
+      // message itself is tiny — the row showed up blank despite having real content.
+      const home = process.env.HOME as string;
+      const projPath = join(home, '.claude', 'projects', 'proj-bookkeeping-test');
+      await mkdir(projPath, { recursive: true });
+
+      const sessionId = '66666666-6666-6666-6666-666666666666';
+      const bookkeepingLine = JSON.stringify({ type: 'mode', mode: 'normal', sessionId }) + '\n';
+      // > 16KB (the old head-read size) but well under 128KB (the new one).
+      const prefix = bookkeepingLine.repeat(Math.ceil(20000 / bookkeepingLine.length));
+      const realLine =
+        JSON.stringify({
+          type: 'user',
+          entrypoint: 'cli',
+          message: { role: 'user', content: 'the real first message' },
+        }) + '\n';
+      expect(prefix.length).toBeGreaterThan(16384);
+
+      await writeFile(join(projPath, `${sessionId}.jsonl`), prefix + realLine);
+
+      const res = await harness.app.inject({
+        method: 'GET',
+        url: '/api/history/sessions?projectKey=proj-bookkeeping-test',
+      });
+      expect(res.statusCode).toBe(200);
+      const row = JSON.parse(res.body).data.sessions.find((s: { sessionId: string }) => s.sessionId === sessionId);
+      expect(row).toBeDefined();
+      expect(row.firstPrompt).toBe('the real first message');
+    });
+
+    it('still falls back to the tail read when bookkeeping alone exceeds the new 128KB head window', async () => {
+      // Raising the head buffer to 128KB helps most restart-heavy sessions, but an
+      // even more extreme case (many more restarts) can still exceed it. This
+      // proves the tail-read fallback itself is intact after the threshold
+      // rewrite (`fileStat.size > headBuf.length` replacing the old hardcoded
+      // 16384/65536) — the fallback's own logic, not the exact threshold value,
+      // is what could have silently broken (e.g. a copy-paste slip that dropped
+      // the `> headBuf.length` check entirely). The real message sits near the
+      // end of the file, well inside the 32KB tail window, so a working fallback
+      // finds it; a broken one leaves the row blank exactly like the bug this
+      // whole fix addresses.
+      const home = process.env.HOME as string;
+      const projPath = join(home, '.claude', 'projects', 'proj-tail-fallback-test');
+      await mkdir(projPath, { recursive: true });
+
+      const sessionId = '88888888-8888-8888-8888-888888888888';
+      const bookkeepingLine = JSON.stringify({ type: 'mode', mode: 'normal', sessionId }) + '\n';
+      // Comfortably past the new 128KB head window (was 16KB), so the head read
+      // never reaches a single "type":"user"/"assistant"/"summary" line.
+      const prefix = bookkeepingLine.repeat(Math.ceil(140000 / bookkeepingLine.length));
+      const realLine =
+        JSON.stringify({
+          type: 'user',
+          entrypoint: 'cli',
+          message: { role: 'user', content: 'found via tail fallback' },
+        }) + '\n';
+      expect(prefix.length).toBeGreaterThan(131072);
+
+      await writeFile(join(projPath, `${sessionId}.jsonl`), prefix + realLine);
+
+      const res = await harness.app.inject({
+        method: 'GET',
+        url: '/api/history/sessions?projectKey=proj-tail-fallback-test',
+      });
+      expect(res.statusCode).toBe(200);
+      const row = JSON.parse(res.body).data.sessions.find((s: { sessionId: string }) => s.sessionId === sessionId);
+      expect(row).toBeDefined();
+      expect(row.firstPrompt).toBe('found via tail fallback');
+    });
   });
 
   // ========== POST /api/sessions (with resumeSessionId) ==========

test/services/unified-session-service.test.ts

@@ -270,6 +270,75 @@ describe('mergeUnifiedSessions', () => {
     expect(live!.firstPrompt).toBeUndefined();
   });
 
+  it('does NOT borrow a sibling transcript for a history-only row whose own extraction failed (no cross-contamination)', () => {
+    // A pure history row already got its own real scan (step 1 keys it under its
+    // OWN sessionId) — if that extraction genuinely failed (oversized first
+    // message, noise-filtered, etc.), the workingDir guess must not paper over
+    // it with an unrelated session's opening line. Regression: an old session
+    // in a shared workingDir was displaying TODAY's live session's firstPrompt
+    // as its own, because the guess didn't check whether this row already had
+    // its own (failed) attempt.
+    const merged = mergeUnifiedSessions({
+      history: [
+        // This session's own transcript scan found no usable prompt.
+        {
+          sessionId: 'old-uuid',
+          workingDir: '/shared',
+          sizeBytes: 5000,
+          lastModified: '2026-01-01T00:00:00.000Z',
+          firstPrompt: undefined,
+        },
+        // A much newer, unrelated session in the same directory.
+        {
+          sessionId: 'newer-uuid',
+          workingDir: '/shared',
+          sizeBytes: 6000,
+          lastModified: '2026-06-01T00:00:00.000Z',
+          firstPrompt: "today's real prompt",
+        },
+      ],
+    });
+    const old = merged.find((m) => m.sessionId === 'old-uuid');
+    expect(old).toBeDefined();
+    expect(old!.firstPrompt).toBeUndefined();
+  });
+
+  it('leaves a RESUMED session blank rather than borrowing a sibling, once its own transcript is aliased in', () => {
+    // The exact scenario COD-140's own comment lists first: a live/persisted row
+    // whose claudeSessionId aliases to an on-disk transcript. Once that alias
+    // successfully folds the transcript's own (failed) extraction into this row
+    // (sources includes 'history'), it must NOT then fall through to the
+    // workingDir guess and borrow an unrelated sibling's prompt -- same bug as
+    // the plain history-only case above, but for the resumed-session path the
+    // backfill mechanism was actually built for.
+    const merged = mergeUnifiedSessions({
+      live: [{ id: 'codeman-resumed', status: 'working', claudeSessionId: 'resumed-uuid', workingDir: '/shared' }],
+      history: [
+        // The resumed session's OWN transcript -- aliased in via claudeSessionId,
+        // but its own extraction found nothing.
+        {
+          sessionId: 'resumed-uuid',
+          workingDir: '/shared',
+          sizeBytes: 5000,
+          lastModified: '2026-01-01T00:00:00.000Z',
+          firstPrompt: undefined,
+        },
+        // An unrelated, newer sibling in the same directory.
+        {
+          sessionId: 'sibling-uuid',
+          workingDir: '/shared',
+          sizeBytes: 6000,
+          lastModified: '2026-06-01T00:00:00.000Z',
+          firstPrompt: "unrelated sibling's prompt",
+        },
+      ],
+    });
+    const resumed = merged.find((m) => m.sessionId === 'codeman-resumed');
+    expect(resumed).toBeDefined();
+    expect([...resumed!.sources].sort()).toEqual(['history', 'live']);
+    expect(resumed!.firstPrompt).toBeUndefined();
+  });
+
   // COD-145: lastPrompt backfill — mirrors the COD-140 firstPrompt path so the
   // most-recent user prompt also reaches live rows whose id ≠ transcript UUID.
   it('backfills lastPrompt onto a live session by claudeSessionId join (uuid-join)', () => {