fix(mirror): capture shell + launcher output to the session page

apps/pwa/src/routes/sessions.mjs

@@ -43,6 +43,26 @@ const LONG_POLL_MS = readLongPollMs();
 // enough that a mis-paste of a whole file can't queue thousands of lines
 // against a prompt that runs them one at a time.
 const MAX_PASTED_LINES = 50;
+// The most output one scrollback row may hold. The CLI flushes at ~16k, so a
+// normal chunk is one row; a burst bigger than this becomes several rows in
+// order rather than one truncated one.
+export const MAX_CHUNK = 20000;
+
+/**
+ * One posted chunk as the rows it will be stored in, in order.
+ *
+ * Every byte survives: nothing is dropped for being past the cap, which is
+ * what the old `slice(0, MAX_CHUNK)` did — quietly, and only to the copy the
+ * scrollback kept, so it showed up as a session that read differently after a
+ * reload than it had live.
+ */
+export function splitChunk(text, max = MAX_CHUNK) {
+  const s = String(text ?? "");
+  if (!s) return [];
+  const parts = [];
+  for (let i = 0; i < s.length; i += max) parts.push(s.slice(i, i + max));
+  return parts;
+}
 
 // Arrow keys pressed on the session page. The command queue carries opaque
 // strings, so a key rides it as a sentinel the CLI decodes — ESC leads it
@@ -172,7 +192,13 @@ sessionsRouter.post("/api/sessions/:id/output", cliAuth, async (req, res) => {
   // at the old column and stays wrong in the scrollback for good.
   if (resized) publish(session.id, { type: "size", cols, rows });
 
-  if (chunk) {
+  // Split rather than truncate. A row is capped so one flush can't put an
+  // unbounded string in the table, but a chunk that overran the cap used to be
+  // stored short while the live watchers got all of it — so the same session
+  // read two different ways (watching, or reloading the page) disagreed, and
+  // the bytes past the cap were gone for good. A build that prints a wall of
+  // output in one 150ms window is exactly when that happens.
+  for (const part of splitChunk(chunk)) {
     // seq is per-session and monotonic so a reconnecting browser can ask for
     // "everything after N" instead of replaying the whole scrollback.
     //
@@ -187,14 +213,14 @@ sessionsRouter.post("/api/sessions/:id/output", cliAuth, async (req, res) => {
       `INSERT INTO session_output (session_id,seq,chunk,created_at)
        SELECT ?, COALESCE(MAX(seq), 0) + 1, ?, ? FROM session_output WHERE session_id = ?
        RETURNING seq`,
-      [session.id, chunk.slice(0, 20000), now, session.id]
+      [session.id, part, now, session.id]
     );
     const seq = Number(inserted.seq);
     await run(
       `DELETE FROM session_output WHERE session_id = ? AND seq <= ?`,
       [session.id, seq - SCROLLBACK]
     );
-    publish(session.id, { type: "out", seq, chunk });
+    publish(session.id, { type: "out", seq, chunk: part });
   }
   if (engine !== session.engine) publish(session.id, { type: "engine", engine });
   res.json({ ok: true });

apps/pwa/test/sessions-output-seq.test.mjs

@@ -216,20 +216,41 @@ test("sessions: an empty chunk writes no output row", skip, async () => {
   assert.deepEqual(await rowsFor(sid), [], "no row for an empty chunk");
 });
 
-test("sessions: a long chunk is still truncated to the stored limit", skip, async () => {
+test("sessions: an ordinary chunk is one row, at the stored limit", skip, async () => {
   const { one, rowsFor } = await app();
 
-  const reg = await one("/api/sessions", { name: "truncate" });
+  const reg = await one("/api/sessions", { name: "at-cap" });
   const sid = reg.body.id;
 
-  await one(`/api/sessions/${sid}/output`, { chunk: "x".repeat(20050) });
+  await one(`/api/sessions/${sid}/output`, { chunk: "x".repeat(20000) });
 
   const rows = await rowsFor(sid);
   assert.equal(rows.length, 1);
-  assert.equal(rows[0].chunk.length, 20000, "chunk is capped at 20000 chars");
+  assert.equal(rows[0].chunk.length, 20000, "a row is still capped at 20000 chars");
   assert.equal(Number(rows[0].seq), 1);
 });
 
+test("sessions: a chunk past the row limit is split, not truncated", skip, async () => {
+  const { one, rowsFor } = await app();
+
+  // The row cap keeps one flush from writing an unbounded string. It used to be
+  // applied with slice(), so the overflow was published live to every watcher
+  // and then thrown away — the same session read one way while you watched it
+  // and another way after a reload, with no sign anything was missing. A build
+  // that dumps a wall of output inside one 150ms flush is exactly that case.
+  const chunk = `${"x".repeat(20000)}TAIL-MUST-SURVIVE\n`;
+  const reg = await one("/api/sessions", { name: "oversized" });
+  const sid = reg.body.id;
+
+  await one(`/api/sessions/${sid}/output`, { chunk });
+
+  const rows = await rowsFor(sid);
+  assert.equal(rows.length, 2, "the overflow becomes another row");
+  assert.ok(rows.every((r) => r.chunk.length <= 20000), "each row still respects the cap");
+  assert.equal(rows.map((r) => r.chunk).join(""), chunk, "every byte posted must be stored");
+  assert.deepEqual(rows.map((r) => Number(r.seq)), [1, 2], "and in order, so a replay reassembles it");
+});
+
 test("sessions: a foreign key still cannot append output", skip, async () => {
   const { one, two, rowsFor } = await app();
 

src/billing.mjs

@@ -19,6 +19,7 @@ import { spawnSync } from "node:child_process";
 
 import { loadBusiness, loadTimers, newId, updateBusiness, updateTimers } from "./business-store.mjs";
 import { clientLabel, parseFields, resolveClient } from "./clients.mjs";
+import { captureSpec } from "./pty.mjs";
 import { GATEWAYS, defaultGateway, gatewayState } from "./payments.mjs";
 import { chargeFor, describeRate, formatMoney, isDollarPegged, isFiat, rateFor } from "./rates.mjs";
 import { humanDuration, selectEntries, windowFrom } from "./timer.mjs";
@@ -289,7 +290,12 @@ function handOff(record, invoice, business, fields, write, run) {
     return 0;
   }
 
-  const result = run("coinpay", args, { stdio: "inherit" });
+  // Mirrored like every other hand-off: sending an invoice is exactly the kind
+  // of thing you want to read back from the session page afterwards.
+  const launch = captureSpec({ cmd: "coinpay", args });
+  let result;
+  try { result = run(launch.cmd, launch.args, { stdio: "inherit" }); }
+  finally { launch.stop(); }
   if (result?.error) { write(err(String(result.error.message || result.error))); return 1; }
   if (result?.status) { write(err(`coinpay exited ${result.status} — invoice ${record.id} is still a local draft`)); return result.status; }
   updateBusiness((data) => {

src/commands.mjs

@@ -22,6 +22,7 @@ import { capture, killSession, remoteStatus, sendPrompt } from "./herd.mjs";
 import { herdStart, isRemoteMember, roster, waitForMany, waitMember } from "./herd-cli.mjs";
 import { endTask, findTask, readTasks, startTask } from "./herd-tasks.mjs";
 import { shellInvocation } from "./shell.mjs";
+import { captureSpec } from "./pty.mjs";
 import { identity, loginAuto, logout as forgetCreds } from "./auth.mjs";
 import { expandAlias, getAlias, loadAliases, removeAlias, setAlias } from "./aliases.mjs";
 import { CORE_CLI_COMMAND_NAMES, PIT_COMMANDS } from "./cli-schema.mjs";
@@ -116,7 +117,15 @@ const SHELL = {
     // has the reasoning, including why a headless run stays non-interactive.
     const { shell: sh, args: shArgs } = shellInvocation(cmd);
     ctx.out(`  ▶ shell: ${cmd}`);
-    const res = spawnSync(sh, shArgs, { stdio: "inherit" });
+    // Captured for the session mirror like the pit's own `!cmd`. A blocking
+    // spawn holds the event loop, so the follower's poll never runs and the
+    // whole command arrives in the drain stop() does — batched rather than
+    // live, which is still the difference between reading it from a phone and
+    // not.
+    const launch = captureSpec({ cmd: sh, args: shArgs });
+    let res;
+    try { res = spawnSync(launch.cmd, launch.args, { stdio: "inherit" }); }
+    finally { launch.stop(); }
     if (res.error) throw res.error;
     const code = res.status ?? 1;
     if (code !== 0) {

src/engines.mjs

@@ -32,11 +32,11 @@
 // a session that starts fresh is a small disappointment, and one that starts
 // with a flag the engine does not have is a crash.
 import { spawn } from "node:child_process";
-import { existsSync, mkdtempSync, readFileSync, rmSync, statSync, writeFileSync } from "node:fs";
-import { homedir, tmpdir } from "node:os";
+import { existsSync, readFileSync, statSync } from "node:fs";
+import { homedir } from "node:os";
 import path from "node:path";
 
-import { followFile, ptyEnabled, ptySpec, scriptFlavor, stripScriptBanner } from "./pty.mjs";
+import { captureSpec } from "./pty.mjs";
 
 export const ENGINES = {
   opencode: {
@@ -439,21 +439,29 @@ export function runCmd(cmd, args = [], { capture = false } = {}) {
     let child;
     const spec = spawnSpec(cmd, args);
     const stdio = capture ? ["inherit", "pipe", "pipe"] : "inherit";
-    try { child = spawn(spec.cmd, spec.args, { stdio }); }
-    catch (e) { resolve({ ok: false, error: e }); return; }
+    // The `capture` branch already reaches a watching browser: it re-writes
+    // every byte through this process's own stdout/stderr, which the mirror
+    // tees. The inherited branch does not — those bytes go to the tty and
+    // nowhere else — so it goes under a pty when a mirror is live. This is what
+    // an upgrade, a plugin install and an `mcp add` all run through, and all
+    // three used to be a rule, a blank stretch, and a result line.
+    const launch = capture ? { ...spec, stop: () => {} } : captureSpec(spec);
+    const finish = (result) => { try { launch.stop(); } catch { /* already drained */ } resolve(result); };
+    try { child = spawn(launch.cmd, launch.args, { stdio }); }
+    catch (e) { finish({ ok: false, error: e }); return; }
     let output = "";
     if (capture) {
       for (const [stream, sink] of [[child.stdout, process.stdout], [child.stderr, process.stderr]]) {
         stream?.on("data", (chunk) => { output += chunk.toString(); sink.write(chunk); });
       }
     }
-    child.on("error", (e) => resolve({ ok: false, error: e, output }));
+    child.on("error", (e) => finish({ ok: false, error: e, output }));
     // "exit" fires as soon as the process is gone, which with pipes can leave
     // the last chunk still queued — the one line we are trying to read. "close"
     // waits for the streams too. With stdio inherited there are no streams, so
     // the two are the same moment and existing callers are unaffected; the
     // distinction is kept explicit so neither branch changes by accident.
-    child.on(capture ? "close" : "exit", (code, signal) => resolve({ ok: true, code, signal, output }));
+    child.on(capture ? "close" : "exit", (code, signal) => finish({ ok: true, code, signal, output }));
   });
 }
 
@@ -501,35 +509,8 @@ export function openPassthrough(target, args = [], { onOutput } = {}) {
     // the child the tty's own file descriptors, so none of its bytes ever pass
     // through this process. See src/pty.mjs for why this is script(1) and not
     // a pipe or node-pty.
-    let transcript = null;
-    let workDir = null;
-    let stopFollow = null;
-    let launch = { ...spec, stdio: "inherit" };
-    if (ptyEnabled(onOutput)) {
-      try {
-        workDir = mkdtempSync(path.join(tmpdir(), "moshcode-pty-"));
-        transcript = path.join(workDir, "transcript");
-        writeFileSync(transcript, "");
-        const wrapped = ptySpec(spec.cmd, spec.args, transcript, scriptFlavor());
-        if (wrapped) {
-          launch = { ...wrapped, stdio: "inherit" };
-          let first = true;
-          stopFollow = followFile(transcript, (chunk) => {
-            const clean = stripScriptBanner(chunk, first);
-            first = false;
-            if (clean) onOutput(clean);
-          });
-        }
-      } catch {
-        // Capture is a nicety; never let it stop the session from opening.
-        transcript = null;
-      }
-    }
-
-    const cleanup = () => {
-      try { stopFollow?.(); } catch { /* nothing left to drain */ }
-      if (workDir) { try { rmSync(workDir, { recursive: true, force: true }); } catch { /* temp dir */ } }
-    };
+    const launch = captureSpec(spec, onOutput);
+    const cleanup = () => launch.stop();
 
     let child;
     try { child = spawn(launch.cmd, launch.args, { stdio: "inherit", env }); }

src/mirror.mjs

@@ -56,6 +56,28 @@ export function pressKey(name, rl = null, stdin = process.stdin) {
 const FLUSH_MS = 150;      // batch writes so a busy render is one request, not fifty
 const MAX_BUFFER = 16000;  // flush early once a batch gets big
 
+// Where a child process's output should be copied while a mirror is watching.
+//
+// Module-level rather than threaded through every call, because "is anyone
+// watching this pit" is one fact about the process and the launchers that need
+// it are scattered: the shell, the installers, the upgrader, the plugin/skill/
+// MCP hand-offs. Passing it down by hand is what left most of them writing
+// straight to the tty with the session page showing nothing — each new launcher
+// had to remember, and none of them did. src/pty.mjs reads this as its default,
+// so capture is what a launcher gets for free and opting out is the deliberate
+// act.
+let activeSink = null;
+
+/** Point child capture at this mirror (or null when the pit stops mirroring). */
+export function setActiveSink(sink) {
+  activeSink = typeof sink === "function" ? sink : null;
+}
+
+/** The sink a child's output should be copied to, or null when unmirrored. */
+export function activeChildSink() {
+  return activeSink;
+}
+
 export function createMirror({
   version = "",
   cwd = process.cwd(),

src/payments.mjs

@@ -22,6 +22,7 @@
 // `/payments connect stripe` records a *reference* — vault and key name — and
 // says out loud where the secret should go.
 import { spawnSync } from "node:child_process";
+import { captureSpec } from "./pty.mjs";
 
 import { loadBusiness, updateBusiness } from "./business-store.mjs";
 import { parseFields } from "./clients.mjs";
@@ -178,7 +179,10 @@ function connectGateway(args, write, run) {
     return 1;
   }
   write(info(`handing you to ${bone(gateway.bin)} — it owns its own session`));
-  const result = run(gateway.bin, gateway.connect, { stdio: "inherit" });
+  const launch = captureSpec({ cmd: gateway.bin, args: gateway.connect });
+  let result;
+  try { result = run(launch.cmd, launch.args, { stdio: "inherit" }); }
+  finally { launch.stop(); }
   if (result?.error) { write(err(String(result.error.message || result.error))); return 1; }
   if (result?.status) {
     write(err(`${gateway.bin} ${gateway.connect.join(" ")} exited ${result.status} — nothing recorded`));

src/pty.mjs

@@ -19,8 +19,11 @@
 // `script` disagree on both flag names and argument order, and anything we
 // cannot positively identify falls back to today's plain `inherit`.
 import { spawnSync } from "node:child_process";
-import { closeSync, existsSync, openSync, readSync, statSync } from "node:fs";
+import { closeSync, existsSync, mkdtempSync, openSync, readSync, rmSync, statSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import path from "node:path";
 import { StringDecoder } from "node:string_decoder";
+import { activeChildSink } from "./mirror.mjs";
 
 /**
  * POSIX single-quote escaping, for argv that has to survive being flattened
@@ -178,3 +181,57 @@ export function ptyEnabled(sink, flavor = scriptFlavor()) {
   if (process.env.MOSHCODE_MIRROR_PTY === "0") return false;
   return Boolean(flavor);
 }
+
+/**
+ * Wrap a spawn spec so a copy of everything the child prints reaches `onOutput`
+ * while the child still owns the real terminal.
+ *
+ * The whole capture dance in one place — temp transcript, the flavour-specific
+ * `script` argv, the follower, the banner strip, the cleanup — because every
+ * launcher in the pit needs it, and each one growing its own copy is how a
+ * shell command ended up invisible in the mirror while `/agents claude` was
+ * captured: both spawn `inherit`, and only one of them had been taught this.
+ *
+ * `onOutput` defaults to whatever the live mirror is (src/mirror.mjs), so a
+ * launcher gets capture without having to know the mirror exists — the reverse
+ * of how this started, where each launcher had to be taught separately and only
+ * two ever were. Pass `null` to opt a launch out.
+ *
+ * Returns `{ cmd, args, stop }`. With nothing watching, or on a box with no
+ * `script(1)` we can drive, `cmd`/`args` come back exactly as passed in and
+ * `stop` is a no-op — the caller spawns what it always spawned. `stop()` must
+ * be called once the child exits: it drains the tail of the transcript (the
+ * last lines of a command are usually the ones you were waiting for) and
+ * removes the temp dir.
+ */
+export function captureSpec({ cmd, args = [] }, onOutput = activeChildSink(), { flavor = scriptFlavor() } = {}) {
+  const plain = { cmd, args, stop: () => {} };
+  if (!ptyEnabled(onOutput, flavor)) return plain;
+  let workDir = null;
+  try {
+    workDir = mkdtempSync(path.join(tmpdir(), "moshcode-pty-"));
+    const transcript = path.join(workDir, "transcript");
+    writeFileSync(transcript, "");
+    const wrapped = ptySpec(cmd, args, transcript, flavor);
+    if (!wrapped) throw new Error("no script(1) spec for this flavour");
+    let first = true;
+    const stopFollow = followFile(transcript, (chunk) => {
+      const clean = stripScriptBanner(chunk, first);
+      first = false;
+      if (clean) onOutput(clean);
+    });
+    const dir = workDir;
+    return {
+      cmd: wrapped.cmd,
+      args: wrapped.args,
+      stop() {
+        try { stopFollow(); } catch { /* nothing left to drain */ }
+        try { rmSync(dir, { recursive: true, force: true }); } catch { /* temp dir */ }
+      },
+    };
+  } catch {
+    // Capture is a nicety; never let it stop a command from running.
+    if (workDir) { try { rmSync(workDir, { recursive: true, force: true }); } catch { /* temp dir */ } }
+    return plain;
+  }
+}