diff --git a/apps/server/src/factory/ledger/listRuns.test.ts b/apps/server/src/factory/ledger/listRuns.test.ts
index eb563d8a4063..e1cdfdd46298 100644
--- a/apps/server/src/factory/ledger/listRuns.test.ts
+++ b/apps/server/src/factory/ledger/listRuns.test.ts
@@ -2,7 +2,7 @@
 import * as NodeFS from "node:fs";
 import * as NodePath from "node:path";
 
-import { afterEach, describe, expect, it } from "vite-plus/test";
+import { afterEach, describe, expect, it, vi } from "vite-plus/test";
 
 import type { State } from "../domain/lifecycle.ts";
 import { makeCause } from "./events.ts";
@@ -11,6 +11,12 @@ import { holdsSlot, listUnfinishedRuns } from "./listRuns.ts";
 import { RUNS_DIR, runIdFileName, type RunId } from "./runId.ts";
 import { T0, ticking, useTempDirs } from "./testSupport.ts";
 
+// Records every full-file read; `node:fs` namespaces are frozen, so the module itself is wrapped.
+vi.mock("node:fs", async (importOriginal) => {
+  const actual = await importOriginal<typeof import("node:fs")>();
+  return { ...actual, readFileSync: vi.fn(actual.readFileSync) };
+});
+
 const tempDir = useTempDirs();
 
 const handles: Array<Ledger> = [];
@@ -163,6 +169,126 @@ describe("listUnfinishedRuns: whether a session may be running", () => {
   });
 });
 
+const line = (event: string, data: unknown, ms = T0) =>
+  `${JSON.stringify({ ts: "2026-10-04T10:13:16.123Z", ms, event, data })}\n`;
+const stateLine = (to: string) => line("state_changed", { from: "ready", to });
+const finishedLine = line("finished", { outcome: "stuck" });
+
+const writeRun = (dataDir: string, id: RunId, content: string) => {
+  NodeFS.mkdirSync(NodePath.join(dataDir, RUNS_DIR), { recursive: true });
+  const path = NodePath.join(dataDir, RUNS_DIR, runIdFileName(id));
+  NodeFS.writeFileSync(path, content);
+  return path;
+};
+
+const fullReads = () => vi.mocked(NodeFS.readFileSync).mock.calls.map(([path]) => String(path));
+
+describe("listUnfinishedRuns: finished runs are skipped by their tail", () => {
+  it("never reads a large finished run whole, and still lists an unfinished sibling", () => {
+    const dataDir = tempDir();
+    const finished = writeRun(
+      dataDir,
+      rid(1),
+      stateLine("triage") +
+        line("step_changed", { step: "x".repeat(200) }).repeat(1000) +
+        finishedLine,
+    );
+    expect(NodeFS.statSync(finished).size).toBeGreaterThan(64 * 1024);
+    const open = writeRun(dataDir, rid(2), stateLine("building"));
+    vi.mocked(NodeFS.readFileSync).mockClear();
+    expect(listUnfinishedRuns(dataDir)).toEqual([
+      { id: rid(2), state: "building", unreadable: false, confirmed: false, session: false },
+    ]);
+    expect(fullReads()).toContain(open);
+    expect(fullReads()).not.toContain(finished);
+  });
+
+  it("skips a small finished run, also after a trailing blank line or CRLF", () => {
+    const dataDir = tempDir();
+    const small = writeRun(dataDir, rid(1), stateLine("triage") + finishedLine);
+    const blank = writeRun(dataDir, rid(2), stateLine("triage") + finishedLine + "\n\n");
+    const crlf = writeRun(dataDir, rid(3), stateLine("triage") + finishedLine + "\r\n");
+    vi.mocked(NodeFS.readFileSync).mockClear();
+    expect(listUnfinishedRuns(dataDir)).toEqual([]);
+    for (const path of [small, blank, crlf]) expect(fullReads()).not.toContain(path);
+  });
+
+  it("reads a file whose last line is cut off in full and lists it as unfinished", () => {
+    const dataDir = tempDir();
+    writeRun(dataDir, rid(1), stateLine("building") + finishedLine.slice(0, 40));
+    expect(listUnfinishedRuns(dataDir)).toEqual([
+      { id: rid(1), state: "building", unreadable: false, confirmed: false, session: false },
+    ]);
+  });
+
+  it("reads a file in full when its last line names finished but is not a valid entry", () => {
+    const dataDir = tempDir();
+    const path = writeRun(
+      dataDir,
+      rid(1),
+      stateLine("building") + '{"event":"finished","data":{}}\n',
+    );
+    vi.mocked(NodeFS.readFileSync).mockClear();
+    expect(listUnfinishedRuns(dataDir)).toEqual([
+      { id: rid(1), state: "building", unreadable: false, confirmed: false, session: false },
+    ]);
+    expect(fullReads()).toContain(path);
+  });
+
+  it("reads a file in full when a cut-off fragment follows the finished line", () => {
+    const dataDir = tempDir();
+    const path = writeRun(
+      dataDir,
+      rid(1),
+      stateLine("triage") + finishedLine + finishedLine.slice(0, 40),
+    );
+    vi.mocked(NodeFS.readFileSync).mockClear();
+    expect(listUnfinishedRuns(dataDir)).toEqual([]);
+    expect(fullReads()).toContain(path);
+  });
+
+  it("reads a file whose last complete line is not JSON in full and lists it as unfinished", () => {
+    const dataDir = tempDir();
+    writeRun(dataDir, rid(1), stateLine("building") + "garbage\n");
+    expect(listUnfinishedRuns(dataDir)).toEqual([
+      { id: rid(1), state: "building", unreadable: false, confirmed: false, session: false },
+    ]);
+  });
+
+  it("a finished line longer than the tail is still skipped, after a full read", () => {
+    const dataDir = tempDir();
+    const long = writeRun(
+      dataDir,
+      rid(1),
+      stateLine("triage") + line("finished", { outcome: "stuck", pad: "y".repeat(20_000) }),
+    );
+    vi.mocked(NodeFS.readFileSync).mockClear();
+    expect(listUnfinishedRuns(dataDir)).toEqual([]);
+    expect(fullReads()).toContain(long);
+  });
+
+  it("lists a run file of only blank lines as unfinished, as readRun reads it", () => {
+    const dataDir = tempDir();
+    writeRun(dataDir, rid(1), "\n");
+    writeRun(dataDir, rid(2), "\n\n");
+    writeRun(dataDir, rid(3), "\n\r\n");
+    const unfinished = { state: null, unreadable: false, confirmed: false, session: false };
+    expect(listUnfinishedRuns(dataDir)).toEqual([
+      { id: rid(1), ...unfinished },
+      { id: rid(2), ...unfinished },
+      { id: rid(3), ...unfinished },
+    ]);
+  });
+
+  it("lists a large run file whose whole tail is blank lines, with its state", () => {
+    const dataDir = tempDir();
+    writeRun(dataDir, rid(1), stateLine("building") + "\n".repeat(9000));
+    expect(listUnfinishedRuns(dataDir)).toEqual([
+      { id: rid(1), state: "building", unreadable: false, confirmed: false, session: false },
+    ]);
+  });
+});
+
 describe("holdsSlot", () => {
   const run = (state: State | null, confirmed: boolean, unreadable = false) => ({
     id: rid(1),
diff --git a/apps/server/src/factory/ledger/listRuns.ts b/apps/server/src/factory/ledger/listRuns.ts
index cf260579b2e6..3a0aa09124bc 100644
--- a/apps/server/src/factory/ledger/listRuns.ts
+++ b/apps/server/src/factory/ledger/listRuns.ts
@@ -5,8 +5,8 @@ import * as NodePath from "node:path";
 import type { State } from "../domain/lifecycle.ts";
 import { readProgress } from "../runner/progress.ts";
 import { errnoCode, LedgerIoError } from "./errors.ts";
-import { readRun } from "./read.ts";
-import { asciiLower, RUNS_DIR, runIdFromFileName, type RunId } from "./runId.ts";
+import { lastLineIsFinished, readRun } from "./read.ts";
+import { asciiLower, RUNS_DIR, runIdFromFileName, runPath, type RunId } from "./runId.ts";
 
 /** One unfinished run as the pass sees it: its id, its last state, and whether its file could be read. */
 export interface RunInfo {
@@ -37,7 +37,8 @@ export const holdsSlot = (run: RunInfo): boolean =>
   run.unreadable || (run.confirmed && run.state !== "pr-open");
 
 /**
- * Every run under `<dir>/runs/` that is not finished, oldest first by the first line's time; with
+ * Every run under `<dir>/runs/` that is not finished (a file whose last line is `finished` is
+ * skipped by reading only its tail), oldest first by the first line's time; with
  * `only`, the runs of that one issue, and no other file is read. Files that are not named like a
  * run are ignored, and a missing runs directory is an empty list. A run file that cannot be read
  * is still listed, flagged `unreadable`, so one bad file does not hide the others. A run with no
@@ -68,6 +69,7 @@ export const listUnfinishedRuns = (
     let firstMs: number | null = null;
     let info: RunInfo;
     try {
+      if (lastLineIsFinished(runPath(dataDir, id))) continue;
       const record = readRun(dataDir, id);
       firstMs = record.entries[0]?.ms ?? null;
       const seen = readProgress(record.entries);
diff --git a/apps/server/src/factory/ledger/read.ts b/apps/server/src/factory/ledger/read.ts
index d92e8a0221b0..8dd0c574745b 100644
--- a/apps/server/src/factory/ledger/read.ts
+++ b/apps/server/src/factory/ledger/read.ts
@@ -152,3 +152,60 @@ export const readRun = (dataDir: string, id: RunId): RunRecord => {
   const bytes = readRunBytes(runPath(dataDir, id));
   return { id, ...parseLines(bytes) };
 };
+
+/** How much of a file's end `lastLineIsFinished` reads; a `finished` line is a few hundred bytes. */
+const TAIL_BYTES = 8192;
+
+/**
+ * Whether the last complete line of a run file is a valid `finished` event, read from at most
+ * `TAIL_BYTES` at the end of the file. `finish` writes that line last, so a run for which this is
+ * true is over without parsing the rest. Answers false for anything it cannot be sure of: a cut-off
+ * last line, a last line that is not a valid entry, or one longer than the tail. Never fails on
+ * content; IO errors map as `readRun` maps them.
+ */
+export const lastLineIsFinished = (path: string): boolean => {
+  requireRegularFile(path);
+  let tail: Buffer;
+  let fromStart: boolean;
+  try {
+    const fd = NodeFS.openSync(path, NodeFS.constants.O_RDONLY | NodeFS.constants.O_NOFOLLOW);
+    try {
+      const size = NodeFS.fstatSync(fd).size;
+      const offset = Math.max(0, size - TAIL_BYTES);
+      tail = Buffer.alloc(size - offset);
+      let filled = 0;
+      while (filled < tail.length) {
+        const n = NodeFS.readSync(fd, tail, filled, tail.length - filled, offset + filled);
+        if (n === 0) break;
+        filled += n;
+      }
+      tail = tail.subarray(0, filled);
+      fromStart = offset === 0;
+    } finally {
+      NodeFS.closeSync(fd);
+    }
+  } catch (error) {
+    if (errnoCode(error) === "ENOENT") throw new LedgerNotFoundError({ path });
+    throw new LedgerIoError({ path, cause: error });
+  }
+
+  // Bytes after the last newline are a write that was cut off, as `parseLines` reads them.
+  let end = tail.lastIndexOf(NEWLINE) + 1;
+  if (end === 0 || !isBlankLine(tail.subarray(end))) return false;
+  // Step back over blank lines to the last line that holds something.
+  for (;;) {
+    const lineEnd = end - 1;
+    // A negative offset would count from the end of the buffer, so the first line has no earlier one.
+    const start = lineEnd === 0 ? 0 : tail.lastIndexOf(NEWLINE, lineEnd - 1) + 1;
+    const line = tail.subarray(start, lineEnd);
+    if (isBlankLine(line)) {
+      if (start === 0) return false;
+      end = start;
+      continue;
+    }
+    // A line that starts the chunk of a longer file may have begun before it.
+    if (start === 0 && !fromStart) return false;
+    const object = decodeLine(line);
+    return object !== null && lineHead(object)?.event === "finished";
+  }
+};