@asterxsk/kiln 0.1.0 → 0.2.0

package/agent/extensions/background-terminals/src/manager.ts

@@ -1,907 +1,907 @@
-/**
- * TerminalManager — owns the registry of running/settled background
- * terminals.
- *
- * Each terminal is a raw `node:child_process` spawn (own process group on
- * POSIX, stdin ignored) whose stdout/stderr 'data' callbacks fold into two
- * bounded OutputBuffers. Closing a terminal's scope kills the whole process
- * tree (SIGTERM → SIGKILL escalation).
- *
- * The manager also exposes a synchronous `TerminalReadModel` so the
- * imperative TUI components (which render synchronously) can read snapshots
- * and issue fire-and-forget kills without touching the Effect runtime.
- */
-
-import { spawn, type ChildProcess } from "node:child_process";
-import * as fs from "node:fs";
-import * as os from "node:os";
-import * as path from "node:path";
-import {
-  Context,
-  Deferred,
-  Effect,
-  Exit,
-  FiberSet,
-  Layer,
-  Scope,
-} from "effect";
-import {
-  ConcurrencyLimitError,
-  formatExit,
-  SpawnError,
-  UnknownTerminalError,
-  type TerminalSnapshot,
-  type TerminalStatus,
-} from "./domain.ts";
-import { OutputBuffer } from "./output.ts";
-
-export const MAX_RUNNING = 8;
-export const MAX_TRACKED = 32;
-const MAX_SETTLED_HISTORY = MAX_TRACKED * 4;
-/** In-memory retained cap per stream. */
-export const RETAINED_PER_STREAM = 2 * 1024 * 1024;
-/** Private full-log spills are bounded so a firehose cannot fill the temp disk. */
-export const MAX_SPILL_BYTES_PER_STREAM = 256 * 1024 * 1024;
-const STOP_TIMEOUT_MS = 5_000;
-/** SIGTERM is normally enough; the second deadline covers a wedged process. */
-const FORCE_KILL_AFTER_MS = 2_000;
-/** After termination, how long to wait for the natural close→flush→settle
- * path before force-settling (a grandchild can hold the stdio pipes open). */
-const SETTLE_GRACE_MS = 1_000;
-/** Bound on waiting for spill WriteStreams to flush before settling; a hung
- * filesystem must not leave an exited entry "running" (and kill() waiting).
- * Terminate (≤2.5s) + settle grace (1s) + flush (1.5s) stays inside the 5s
- * scope-close bound, so teardown remains bounded end to end. */
-const SPILL_FLUSH_TIMEOUT_MS = 1_500;
-const ERROR_TEXT_MAX_LENGTH = 4_096;
-
-function bounded(text: string) {
-  return text.slice(0, ERROR_TEXT_MAX_LENGTH);
-}
-
-function boundedError(error: unknown) {
-  return bounded(error instanceof Error ? error.message : String(error));
-}
-
-// --- Internal state -----------------------------------------------------------
-
-/** Mutable snapshot; exposed to readers via the readonly TerminalSnapshot type.
- * stdout/stderr are getters over the live OutputBuffers. */
-interface MutableSnapshot extends TerminalSnapshot {
-  status: TerminalStatus;
-  pid?: number;
-  settledAt?: number;
-  exitCode?: number;
-  signal?: string;
-  errorText?: string;
-}
-
-interface Entry {
-  snapshot: MutableSnapshot;
-  child: ChildProcess;
-  scope: Scope.Closeable;
-  stdoutBuf: OutputBuffer;
-  stderrBuf: OutputBuffer;
-  spillStreams: fs.WriteStream[];
-  /** Set in the same synchronous effect that sends SIGTERM so a natural exit
-   * before signaling keeps its truthful status. */
-  killSignaled: boolean;
-  /** The child emitted 'error' (spawn failure etc.); settles as "failed".
-   * Kept separate from errorText, which also carries non-fatal notes
-   * (spill failures) that must not flip a clean exit to "failed". */
-  processErrored: boolean;
-  /** 'exit' event observed (code/signal recorded). */
-  exited: boolean;
-  /** 'close' event observed (stdio flushed; the settle trigger). */
-  stdioClosed: boolean;
-  /** A settle-after-spill-flush is in flight; don't start a second one. */
-  settling: boolean;
-  /** The shell exited without stdio closing; a bounded scope close is queued
-   * to reap descendants that still hold the inherited pipes open. */
-  exitCleanupStarted: boolean;
-  /** Completed exactly once when the entry settles. Kill callers and the scope
-   * finalizer can all await the same result without missing a notification. */
-  settled: Deferred.Deferred<void>;
-}
-
-export interface StartOptions {
-  readonly command: string;
-  readonly title: string;
-  readonly cwd: string;
-}
-
-export interface KillResult {
-  readonly id: string;
-  readonly title: string;
-  readonly status: TerminalStatus;
-  /** True when the entry was still running when this kill began. */
-  readonly wasRunning: boolean;
-  /** True when this call initiated the termination AND the entry settled as
-   * killed (a natural exit that won the race reports killed: false). */
-  readonly killed: boolean;
-  /** Final exit rendering ("exit 0", "SIGTERM", ...) captured at settle time,
-   * so reports stay accurate even if the entry is pruned afterwards. */
-  readonly exit: string;
-}
-
-// --- Read model ----------------------------------------------------------------
-
-/** Synchronous bridge for the TUI. Snapshots are live objects; do not mutate. */
-export interface TerminalReadModel {
-  list(): ReadonlyArray<TerminalSnapshot>;
-  get(id: string): TerminalSnapshot | undefined;
-  size(): number;
-  /** Any-change notification (widget, /ps list). */
-  subscribe(listener: () => void): () => void;
-  /** Per-terminal notification (/ps detail view). */
-  subscribeTo(id: string, listener: () => void): () => void;
-  /** Fire-and-forget kill (dashboard/detail `x`). Not marked consumed: the
-   * settle still flows back to the model as a follow-up message. */
-  requestKill(id: string): void;
-  /**
-   * Register the settle hook. `consumed` is true when an active bg_kill is
-   * collecting the result (so it must not also be delivered as a follow-up).
-   */
-  setOnSettled(
-    hook: ((snap: TerminalSnapshot, consumed: boolean) => void) | undefined,
-  ): void;
-}
-
-// --- Service --------------------------------------------------------------------
-
-export interface TerminalManagerShape {
-  start(
-    options: StartOptions,
-  ): Effect.Effect<TerminalSnapshot, SpawnError | ConcurrencyLimitError>;
-  status(id: string): Effect.Effect<TerminalSnapshot, UnknownTerminalError>;
-  /** Kill running terminals; resolves only after they have settled. */
-  kill(ids: ReadonlyArray<string>): Effect.Effect<ReadonlyArray<KillResult>>;
-  readonly list: Effect.Effect<ReadonlyArray<TerminalSnapshot>>;
-  readonly disposeAll: Effect.Effect<void>;
-  readonly view: TerminalReadModel;
-}
-
-export class TerminalManager extends Context.Service<
-  TerminalManager,
-  TerminalManagerShape
->()("background-terminals/TerminalManager") {}
-
-// --- Process helpers ------------------------------------------------------------
-
-function shellInvocation(command: string) {
-  if (process.platform === "win32") {
-    const shell = process.env.ComSpec ?? "cmd.exe";
-    return { shell, args: ["/d", "/s", "/c", command] };
-  }
-  return { shell: "/bin/sh", args: ["-c", command] };
-}
-
-/** Signal the whole process group on POSIX so descendants (servers a shell
- * command spawned) die with it; a wedged child must not orphan its tree. */
-function killTree(child: ChildProcess, signal: NodeJS.Signals) {
-  if (process.platform === "win32" && child.pid) {
-    try {
-      const killer = spawn(
-        "taskkill",
-        [
-          "/pid",
-          String(child.pid),
-          "/T",
-          ...(signal === "SIGKILL" ? ["/F"] : []),
-        ],
-        { stdio: "ignore", windowsHide: true },
-      );
-      killer.once("error", () => {
-        try {
-          child.kill(signal);
-        } catch {
-          // Process may already be gone.
-        }
-      });
-      killer.once("exit", (code) => {
-        if (code === 0) return;
-        try {
-          child.kill(signal);
-        } catch {
-          // Process may already be gone.
-        }
-      });
-      killer.unref();
-      return;
-    } catch {
-      // Fall through to the direct signal when taskkill cannot be launched.
-    }
-  }
-  if (process.platform !== "win32" && child.pid) {
-    try {
-      process.kill(-child.pid, signal);
-      return;
-    } catch {
-      // Group may already be gone; fall through to the direct signal.
-    }
-  }
-  try {
-    child.kill(signal);
-  } catch {
-    // Process may already be gone.
-  }
-}
-
-/** Await stdio closure without retaining a listener after interruption. */
-function awaitChildClose(child: ChildProcess, closed: () => boolean) {
-  return Effect.callback<void>((resume) => {
-    if (closed()) {
-      resume(Effect.void);
-      return;
-    }
-    const onClose = () => resume(Effect.void);
-    child.once("close", onClose);
-    return Effect.sync(() => child.off("close", onClose));
-  });
-}
-
-/** SIGTERM → deadline → SIGKILL; waits for stdio closure rather than only the
- * shell's exit because descendants can keep the inherited pipes and process
- * group alive after the shell itself is gone. */
-function terminateChild(
-  child: ChildProcess,
-  closed: () => boolean,
-  onSignal: () => void,
-) {
-  return Effect.suspend(() => {
-    if (closed()) return Effect.void;
-    return Effect.gen(function* () {
-      yield* Effect.sync(() => {
-        onSignal();
-        killTree(child, "SIGTERM");
-      });
-      yield* awaitChildClose(child, closed).pipe(
-        Effect.timeout(FORCE_KILL_AFTER_MS),
-        Effect.ignore,
-      );
-      if (closed()) return;
-      yield* Effect.sync(() => killTree(child, "SIGKILL"));
-      yield* awaitChildClose(child, closed).pipe(
-        Effect.timeout(500),
-        Effect.ignore,
-      );
-    });
-  });
-}
-
-// --- Implementation --------------------------------------------------------------
-
-const makeManager = Effect.gen(function* () {
-  // Scoped detached forker for sync contexts (read-model kills, process-event
-  // settlement, pruning). Completed fibers remove themselves; manager scope
-  // close interrupts any work that outlives the bounded disposeAll wait.
-  const cleanupFibers = yield* FiberSet.make();
-  const runCleanup = yield* FiberSet.runtime(cleanupFibers)();
-
-  const entries = new Map<string, Entry>();
-  /** Small immutable tombstones preserve truthful kill reports if pruning
-   * races the tool boundary after an id was validated. */
-  const settledHistory = new Map<
-    string,
-    Pick<KillResult, "title" | "status" | "exit">
-  >();
-  /** ids with an in-flight kill() collecting the result (settle → consumed). */
-  const killInterest = new Map<string, number>();
-  const listeners = new Set<() => void>();
-  const idListeners = new Map<string, Set<() => void>>();
-  let counter = 0;
-  let reserved = 0;
-  let disposed = false;
-  let spillDir: string | undefined | null;
-  let onSettled:
-    ((snap: TerminalSnapshot, consumed: boolean) => void) | undefined;
-
-  const notify = (id?: string) => {
-    for (const listener of [...listeners]) {
-      try {
-        listener();
-      } catch {
-        // A failed widget/render listener must not corrupt lifecycle state.
-      }
-    }
-    if (id) {
-      for (const listener of idListeners.get(id) ?? []) {
-        try {
-          listener();
-        } catch {
-          // Same.
-        }
-      }
-    }
-  };
-
-  const runningCount = () =>
-    [...entries.values()].filter((e) => e.snapshot.status === "running").length;
-
-  const addKillInterest = (ids: ReadonlyArray<string>) => {
-    for (const id of ids) killInterest.set(id, (killInterest.get(id) ?? 0) + 1);
-  };
-  const releaseKillInterest = (ids: ReadonlyArray<string>) => {
-    for (const id of ids) {
-      const count = (killInterest.get(id) ?? 1) - 1;
-      if (count <= 0) killInterest.delete(id);
-      else killInterest.set(id, count);
-    }
-  };
-
-  const closeEntryScope = (entry: Entry) =>
-    Scope.close(entry.scope, Exit.void).pipe(Effect.ignore);
-
-  const pruneSettled = () => {
-    if (entries.size <= MAX_TRACKED) return;
-    const candidates = [...entries.values()]
-      .filter(
-        (e) =>
-          e.snapshot.status !== "running" && !killInterest.has(e.snapshot.id),
-      )
-      .sort(
-        (a, b) =>
-          (a.snapshot.settledAt ?? a.snapshot.createdAt) -
-          (b.snapshot.settledAt ?? b.snapshot.createdAt),
-      );
-    for (const entry of candidates) {
-      if (entries.size <= MAX_TRACKED) break;
-      entries.delete(entry.snapshot.id);
-      runCleanup(closeEntryScope(entry));
-    }
-  };
-
-  /** End all spill streams; resolves when their buffers are flushed to disk
-   * (bounded), so a settle notification never points at a partial file. */
-  const flushSpillStreams = (entry: Entry) => {
-    const streams = entry.spillStreams;
-    entry.spillStreams = [];
-    return Effect.forEach(
-      streams,
-      (stream) =>
-        Effect.callback<void>((resume) => {
-          const done = () => resume(Effect.void);
-          try {
-            stream.end(done);
-          } catch {
-            // Best effort; tmpdir contents are disposable.
-            done();
-          }
-        }),
-      { concurrency: "unbounded", discard: true },
-    ).pipe(
-      Effect.timeoutOrElse({
-        duration: SPILL_FLUSH_TIMEOUT_MS,
-        orElse: () =>
-          Effect.sync(() => {
-            entry.stdoutBuf.spillPath = undefined;
-            entry.stderrBuf.spillPath = undefined;
-            entry.snapshot.errorText ??=
-              "Full-log spill flush timed out; full output may be incomplete";
-          }),
-      }),
-    );
-  };
-
-  /** Single settle path — idempotent; kill vs natural exit vs error races are
-   * resolved by whichever lands first (the second call is a no-op). */
-  const settle = (entry: Entry) => {
-    const s = entry.snapshot;
-    if (s.status !== "running") return;
-    s.settledAt = Date.now();
-    s.status = entry.killSignaled
-      ? "killed"
-      : entry.processErrored
-        ? "failed"
-        : s.exitCode === 0
-          ? "done"
-          : "failed";
-    settledHistory.set(s.id, {
-      title: s.title,
-      status: s.status,
-      exit: formatExit(s),
-    });
-    while (settledHistory.size > MAX_SETTLED_HISTORY) {
-      const oldest = settledHistory.keys().next().value;
-      if (oldest === undefined) break;
-      settledHistory.delete(oldest);
-    }
-    // Completing the Deferred can immediately resume kill waiters, whose
-    // ensuring blocks release interest. Snapshot consumption first so the
-    // settle hook observes the interest that existed when settlement won.
-    const consumed = (killInterest.get(s.id) ?? 0) > 0;
-    Deferred.doneUnsafe(entry.settled, Effect.void);
-    notify(s.id);
-    try {
-      // During teardown, don't queue results into a shutting-down session.
-      if (!disposed) onSettled?.(s, consumed);
-    } catch {
-      // The parent session may be unavailable; settlement stays final.
-    }
-    pruneSettled();
-  };
-
-  /** Flush the spill files, then settle: the completion follow-up (and the
-   * kill() resolution) reference the spill path, so the full capture must be
-   * on disk before anyone is told about it. Idempotent via `settling`. */
-  const settleAfterFlush = (entry: Entry) => {
-    if (entry.settling || entry.snapshot.status !== "running") return;
-    entry.settling = true;
-    runCleanup(
-      flushSpillStreams(entry).pipe(
-        Effect.andThen(Effect.sync(() => settle(entry))),
-      ),
-    );
-  };
-
-  const scheduleExitCleanup = (entry: Entry) => {
-    if (entry.exitCleanupStarted) return;
-    entry.exitCleanupStarted = true;
-    runCleanup(
-      Effect.sleep(SETTLE_GRACE_MS).pipe(
-        Effect.andThen(
-          Effect.suspend(() =>
-            entry.snapshot.status === "running" && !entry.stdioClosed
-              ? closeEntryScope(entry).pipe(
-                  Effect.timeout(STOP_TIMEOUT_MS),
-                  Effect.ignore,
-                )
-              : Effect.void,
-          ),
-        ),
-      ),
-    );
-  };
-
-  const resolveSpillDir = () => {
-    if (spillDir !== undefined) return spillDir ?? undefined;
-    try {
-      const base = path.join(os.tmpdir(), "pi-background-terminals");
-      fs.mkdirSync(base, { recursive: true, mode: 0o700 });
-      fs.chmodSync(base, 0o700);
-      spillDir = fs.mkdtempSync(path.join(base, "session-"));
-      fs.chmodSync(spillDir, 0o700);
-    } catch {
-      spillDir = null;
-    }
-    return spillDir ?? undefined;
-  };
-
-  const makeSpill = (
-    entry: () => Entry | undefined,
-    id: string,
-    stream: "stdout" | "stderr",
-    resumeSource: () => void,
-  ) => {
-    const dir = resolveSpillDir();
-    if (!dir) return undefined;
-    const spillPath = path.join(dir, `${id}.${stream}.log`);
-    try {
-      const file = fs.createWriteStream(spillPath, {
-        flags: "a",
-        mode: 0o600,
-      });
-      let broken = false;
-      let capped = false;
-      let writtenBytes = 0;
-      file.on("error", (error) => {
-        broken = true;
-        resumeSource();
-        const current = entry();
-        if (current) {
-          const buf =
-            stream === "stdout" ? current.stdoutBuf : current.stderrBuf;
-          buf.spillPath = undefined;
-          current.snapshot.errorText ??= bounded(
-            `Full-log spill to ${spillPath} failed: ${boundedError(error)}`,
-          );
-        }
-      });
-      return {
-        spillPath,
-        file,
-        write: (chunk: string) => {
-          // writableEnded guard: late 'data' after the settle flush must not
-          // error the ended stream (and falsely report the spill as broken).
-          if (broken || capped || file.writableEnded) return true;
-          const chunkBytes = Buffer.byteLength(chunk, "utf8");
-          if (writtenBytes + chunkBytes > MAX_SPILL_BYTES_PER_STREAM) {
-            capped = true;
-            const current = entry();
-            if (current) {
-              const buf =
-                stream === "stdout" ? current.stdoutBuf : current.stderrBuf;
-              buf.spillPath = undefined;
-              current.snapshot.errorText ??= bounded(
-                `${stream} full-log spill reached the ${MAX_SPILL_BYTES_PER_STREAM}-byte safety limit`,
-              );
-            }
-            return true;
-          }
-          writtenBytes += chunkBytes;
-          const accepted = file.write(chunk);
-          if (!accepted) file.once("drain", resumeSource);
-          return accepted;
-        },
-      };
-    } catch {
-      return undefined;
-    }
-  };
-
-  const start = (options: StartOptions) =>
-    Effect.gen(function* () {
-      // Reserve synchronously (before the first yield inside doStart) so
-      // parallel tool calls cannot race past the cap.
-      yield* Effect.suspend(
-        (): Effect.Effect<void, SpawnError | ConcurrencyLimitError> => {
-          if (disposed) {
-            return new SpawnError({
-              message: "Background terminal manager is shutting down.",
-            });
-          }
-          if (runningCount() + reserved >= MAX_RUNNING) {
-            return new ConcurrencyLimitError({
-              message: `Max ${MAX_RUNNING} background terminals can run concurrently. Stop one with bg_kill before starting another.`,
-            });
-          }
-          reserved++;
-          return Effect.void;
-        },
-      );
-
-      const doStart = Effect.gen(function* () {
-        const { shell, args } = shellInvocation(options.command);
-        const child = yield* Effect.try({
-          try: () =>
-            spawn(shell, args, {
-              cwd: options.cwd,
-              env: process.env,
-              // stdin IGNORED: there is no input surface, ever. A process
-              // that reads stdin sees EOF immediately.
-              stdio: ["ignore", "pipe", "pipe"],
-              // Own process group on POSIX → group kill takes the whole tree.
-              detached: process.platform !== "win32",
-            }),
-          catch: (error) => new SpawnError({ message: boundedError(error) }),
-        });
-
-        const id = `bt-${++counter}`;
-        const entryRef = () => entries.get(id);
-        const stdoutSpill = makeSpill(entryRef, id, "stdout", () =>
-          child.stdout?.resume(),
-        );
-        const stderrSpill = makeSpill(entryRef, id, "stderr", () =>
-          child.stderr?.resume(),
-        );
-        const stdoutBuf = new OutputBuffer(
-          RETAINED_PER_STREAM,
-          stdoutSpill?.write,
-        );
-        const stderrBuf = new OutputBuffer(
-          RETAINED_PER_STREAM,
-          stderrSpill?.write,
-        );
-        stdoutBuf.spillPath = stdoutSpill?.spillPath;
-        stderrBuf.spillPath = stderrSpill?.spillPath;
-
-        const snapshot: MutableSnapshot = {
-          id,
-          command: options.command,
-          title: options.title,
-          cwd: options.cwd,
-          pid: child.pid,
-          status: "running",
-          createdAt: Date.now(),
-          get stdout() {
-            return stdoutBuf.view();
-          },
-          get stderr() {
-            return stderrBuf.view();
-          },
-        };
-
-        const scope = yield* Scope.make();
-        const settled = yield* Deferred.make<void>();
-        const entry: Entry = {
-          snapshot,
-          child,
-          scope,
-          stdoutBuf,
-          stderrBuf,
-          spillStreams: [stdoutSpill?.file, stderrSpill?.file].filter(
-            (file): file is fs.WriteStream => file !== undefined,
-          ),
-          killSignaled: false,
-          processErrored: false,
-          exited: false,
-          stdioClosed: false,
-          settling: false,
-          exitCleanupStarted: false,
-          settled,
-        };
-
-        // Plain-callback stream plumbing (the codex-backend precedent):
-        // setEncoding's internal StringDecoder is multibyte-safe across
-        // chunk boundaries.
-        child.stdout?.setEncoding("utf8");
-        child.stdout?.on("data", (chunk: string) => {
-          if (!stdoutBuf.push(chunk)) child.stdout?.pause();
-          notify(id);
-        });
-        child.stderr?.setEncoding("utf8");
-        child.stderr?.on("data", (chunk: string) => {
-          if (!stderrBuf.push(chunk)) child.stderr?.pause();
-          notify(id);
-        });
-        // Spawn failures (ENOENT etc.) arrive via 'error', not a throw. Node
-        // still emits 'close' afterwards (with a bogus errno as code), so
-        // record the failure here and let the close path do the one settle.
-        child.once("error", (error) => {
-          entry.processErrored = true;
-          snapshot.errorText ??= boundedError(error);
-          entry.exited = true;
-          settleAfterFlush(entry);
-        });
-        // Record code/signal on 'exit'; settle on 'close' so the completion
-        // notification always carries the final flushed output.
-        child.once("exit", (code, signal) => {
-          entry.exited = true;
-          snapshot.exitCode = code ?? undefined;
-          snapshot.signal = signal ?? undefined;
-          // A descendant can keep the pipes open after the shell exits. Give
-          // close a short natural grace, then close the scope to terminate
-          // the surviving process group and force a bounded settlement.
-          scheduleExitCleanup(entry);
-        });
-        child.once("close", (code, signal) => {
-          entry.exited = true;
-          entry.stdioClosed = true;
-          // Only trust close's code/signal when 'exit' never fired (a spawn
-          // 'error' close reports the errno, e.g. -2, as its code).
-          if (!entry.processErrored) {
-            snapshot.exitCode ??= code ?? undefined;
-            snapshot.signal ??= signal ?? undefined;
-          }
-          settleAfterFlush(entry);
-        });
-
-        // One teardown path: kill(), requestKill, pruning, disposeAll, and
-        // runtime.dispose() all converge on closing this scope.
-        yield* Scope.provide(
-          Effect.addFinalizer(() =>
-            Effect.gen(function* () {
-              // Only claim "killed" when we are actually about to signal a
-              // live process; a natural exit that already happened (still
-              // waiting on 'close') keeps its truthful done/failed status.
-              yield* terminateChild(
-                child,
-                () => entry.stdioClosed,
-                () => {
-                  entry.killSignaled ||=
-                    !entry.exited && entry.snapshot.status === "running";
-                },
-              );
-              // Give the natural close→flush→settle path a bounded grace,
-              // then force the settle: a grandchild holding the pipe open
-              // (detached into a new group) must not leave the entry
-              // "running" forever.
-              if (entry.snapshot.status === "running") {
-                yield* Deferred.await(entry.settled).pipe(
-                  Effect.timeout(SETTLE_GRACE_MS),
-                  Effect.ignore,
-                );
-              }
-              if (entry.snapshot.status === "running" && !entry.settling) {
-                // Force the settle ourselves. When `settling` is set, the
-                // close path's flush→settle is already in flight (bounded by
-                // SPILL_FLUSH_TIMEOUT_MS) — settling here first would cite a
-                // spill file that is still being flushed.
-                if (!entry.stdioClosed) {
-                  entry.snapshot.errorText ??=
-                    "stdio did not close after termination; output may be incomplete";
-                }
-                entry.settling = true;
-                yield* flushSpillStreams(entry);
-                settle(entry);
-              }
-            }),
-          ),
-          scope,
-        );
-
-        // disposeAll may have swept the entries map while we were setting up;
-        // an entry added after the sweep would never be torn down. Close our
-        // own scope (kills the child) and fail instead (subagents precedent).
-        if (disposed) {
-          yield* closeEntryScope(entry);
-          return yield* new SpawnError({
-            message: "Background terminal manager shut down while starting.",
-          });
-        }
-        entries.set(id, entry);
-        notify(id);
-        return snapshot as TerminalSnapshot;
-      });
-
-      // Uninterruptible: between spawn() and entries.set there must be no
-      // window where an interrupt (tool abort, runtime dispose) leaves a
-      // live child that no scope/registry knows about. All steps are sync.
-      return yield* doStart.pipe(
-        Effect.uninterruptible,
-        Effect.ensuring(
-          Effect.sync(() => {
-            reserved--;
-            notify();
-          }),
-        ),
-      );
-    });
-
-  const status = (id: string) =>
-    Effect.suspend(
-      (): Effect.Effect<TerminalSnapshot, UnknownTerminalError> => {
-        const entry = entries.get(id);
-        if (!entry) {
-          const known = [...entries.keys()];
-          return new UnknownTerminalError({
-            message: `Unknown terminal id "${id}". Known: ${known.join(", ") || "none"}.`,
-          });
-        }
-        return Effect.succeed(entry.snapshot as TerminalSnapshot);
-      },
-    );
-
-  /** Kill one running entry: close the scope — whose finalizer marks the kill
-   * at the signal point, terminates the tree, and force-settles —
-   * in a DETACHED fiber. Once the flag is set the termination must actually
-   * happen; a tool abort interrupting the caller cannot cancel it (this is
-   * what makes "termination continues in the background" truthful). */
-  const killEntry = (entry: Entry) =>
-    Effect.sync(() => {
-      if (entry.snapshot.status !== "running") return;
-      runCleanup(
-        closeEntryScope(entry).pipe(
-          Effect.timeout(STOP_TIMEOUT_MS),
-          Effect.ignore,
-        ),
-      );
-    });
-
-  const kill = (ids: ReadonlyArray<string>) =>
-    Effect.suspend(() => {
-      const unique = [...new Set(ids)];
-      const byId = new Map(
-        unique
-          .map((id) => entries.get(id))
-          .filter((entry): entry is Entry => entry !== undefined)
-          .map((entry) => [entry.snapshot.id, entry]),
-      );
-      const running = [...byId.values()].filter(
-        (entry) => entry.snapshot.status === "running",
-      );
-      const runningIds = running.map((entry) => entry.snapshot.id);
-      // Mark consumed before signaling so this kill's settlements are not
-      // ALSO queued as automatic follow-up messages to the model.
-      addKillInterest(runningIds);
-      const work = Effect.gen(function* () {
-        yield* Effect.forEach(running, killEntry, {
-          concurrency: "unbounded",
-        });
-        // Every caller waits on the entries that were running when its kill
-        // began. Deferred completion cannot be missed and supports concurrent
-        // overlapping/multi-id kill calls.
-        yield* Effect.forEach(
-          running,
-          (entry) => Deferred.await(entry.settled),
-          { concurrency: "unbounded", discard: true },
-        );
-        // Capture the report BEFORE the ensuring below releases interest and
-        // prunes — a just-settled entry must not vanish out from under it.
-        return unique.map((id): KillResult => {
-          const snapshot = byId.get(id)?.snapshot;
-          const history = settledHistory.get(id);
-          const status = snapshot?.status ?? history?.status ?? "killed";
-          const wasRunning = runningIds.includes(id);
-          return {
-            id,
-            title: snapshot?.title ?? history?.title ?? "?",
-            status,
-            wasRunning,
-            // A natural exit can win the race with our SIGTERM; report what
-            // actually happened rather than claiming the kill did it.
-            killed: wasRunning && status === "killed",
-            exit: snapshot
-              ? formatExit(snapshot)
-              : (history?.exit ?? "unknown"),
-          };
-        });
-      });
-      return work.pipe(
-        Effect.ensuring(
-          Effect.sync(() => {
-            releaseKillInterest(runningIds);
-            pruneSettled();
-          }),
-        ),
-      );
-    });
-
-  const disposeAll = Effect.gen(function* () {
-    disposed = true;
-    const all = [...entries.values()];
-    entries.clear();
-    yield* Effect.forEach(
-      all,
-      (entry) =>
-        closeEntryScope(entry).pipe(
-          Effect.timeout(STOP_TIMEOUT_MS),
-          Effect.ignore,
-        ),
-      { concurrency: "unbounded" },
-    );
-    // Detached kill/prune/flush work is scoped to the manager. Wait for it
-    // within the shutdown bound; the FiberSet finalizer interrupts anything
-    // still live when the manager scope closes, so cleanup cannot leak.
-    yield* FiberSet.awaitEmpty(cleanupFibers).pipe(
-      Effect.timeout(STOP_TIMEOUT_MS),
-      Effect.ignore,
-    );
-    yield* Effect.sync(() => {
-      const dir = spillDir;
-      spillDir = null;
-      if (dir) fs.rmSync(dir, { recursive: true, force: true });
-    });
-    yield* Effect.sync(() => notify());
-  });
-
-  const view: TerminalReadModel = {
-    list: () => [...entries.values()].map((entry) => entry.snapshot),
-    get: (id) => entries.get(id)?.snapshot,
-    size: () => entries.size,
-    subscribe: (listener) => {
-      listeners.add(listener);
-      return () => listeners.delete(listener);
-    },
-    subscribeTo: (id, listener) => {
-      let set = idListeners.get(id);
-      if (!set) {
-        set = new Set();
-        idListeners.set(id, set);
-      }
-      set.add(listener);
-      return () => {
-        set.delete(listener);
-        if (set.size === 0) idListeners.delete(id);
-      };
-    },
-    requestKill: (id) => {
-      const entry = entries.get(id);
-      if (!entry) return;
-      // UI-initiated kills are not "consumed": the killed result still flows
-      // back to the model as a follow-up message (subagents precedent).
-      runCleanup(killEntry(entry).pipe(Effect.ignore));
-    },
-    setOnSettled: (hook) => {
-      onSettled = hook;
-    },
-  };
-
-  // Safety net: disposing the ManagedRuntime tears everything down even if
-  // the extension forgot to call disposeAll explicitly.
-  yield* Effect.addFinalizer(() => disposeAll);
-
-  return TerminalManager.of({
-    start,
-    status,
-    kill,
-    list: Effect.sync(() => [...entries.values()].map((e) => e.snapshot)),
-    disposeAll,
-    view,
-  });
-});
-
-export const TerminalManagerLive: Layer.Layer<TerminalManager> = Layer.effect(
-  TerminalManager,
-  makeManager,
-);
+/**
+ * TerminalManager — owns the registry of running/settled background
+ * terminals.
+ *
+ * Each terminal is a raw `node:child_process` spawn (own process group on
+ * POSIX, stdin ignored) whose stdout/stderr 'data' callbacks fold into two
+ * bounded OutputBuffers. Closing a terminal's scope kills the whole process
+ * tree (SIGTERM → SIGKILL escalation).
+ *
+ * The manager also exposes a synchronous `TerminalReadModel` so the
+ * imperative TUI components (which render synchronously) can read snapshots
+ * and issue fire-and-forget kills without touching the Effect runtime.
+ */
+
+import { spawn, type ChildProcess } from "node:child_process";
+import * as fs from "node:fs";
+import * as os from "node:os";
+import * as path from "node:path";
+import {
+  Context,
+  Deferred,
+  Effect,
+  Exit,
+  FiberSet,
+  Layer,
+  Scope,
+} from "effect";
+import {
+  ConcurrencyLimitError,
+  formatExit,
+  SpawnError,
+  UnknownTerminalError,
+  type TerminalSnapshot,
+  type TerminalStatus,
+} from "./domain.ts";
+import { OutputBuffer } from "./output.ts";
+
+export const MAX_RUNNING = 8;
+export const MAX_TRACKED = 32;
+const MAX_SETTLED_HISTORY = MAX_TRACKED * 4;
+/** In-memory retained cap per stream. */
+export const RETAINED_PER_STREAM = 2 * 1024 * 1024;
+/** Private full-log spills are bounded so a firehose cannot fill the temp disk. */
+export const MAX_SPILL_BYTES_PER_STREAM = 256 * 1024 * 1024;
+const STOP_TIMEOUT_MS = 5_000;
+/** SIGTERM is normally enough; the second deadline covers a wedged process. */
+const FORCE_KILL_AFTER_MS = 2_000;
+/** After termination, how long to wait for the natural close→flush→settle
+ * path before force-settling (a grandchild can hold the stdio pipes open). */
+const SETTLE_GRACE_MS = 1_000;
+/** Bound on waiting for spill WriteStreams to flush before settling; a hung
+ * filesystem must not leave an exited entry "running" (and kill() waiting).
+ * Terminate (≤2.5s) + settle grace (1s) + flush (1.5s) stays inside the 5s
+ * scope-close bound, so teardown remains bounded end to end. */
+const SPILL_FLUSH_TIMEOUT_MS = 1_500;
+const ERROR_TEXT_MAX_LENGTH = 4_096;
+
+function bounded(text: string) {
+  return text.slice(0, ERROR_TEXT_MAX_LENGTH);
+}
+
+function boundedError(error: unknown) {
+  return bounded(error instanceof Error ? error.message : String(error));
+}
+
+// --- Internal state -----------------------------------------------------------
+
+/** Mutable snapshot; exposed to readers via the readonly TerminalSnapshot type.
+ * stdout/stderr are getters over the live OutputBuffers. */
+interface MutableSnapshot extends TerminalSnapshot {
+  status: TerminalStatus;
+  pid?: number;
+  settledAt?: number;
+  exitCode?: number;
+  signal?: string;
+  errorText?: string;
+}
+
+interface Entry {
+  snapshot: MutableSnapshot;
+  child: ChildProcess;
+  scope: Scope.Closeable;
+  stdoutBuf: OutputBuffer;
+  stderrBuf: OutputBuffer;
+  spillStreams: fs.WriteStream[];
+  /** Set in the same synchronous effect that sends SIGTERM so a natural exit
+   * before signaling keeps its truthful status. */
+  killSignaled: boolean;
+  /** The child emitted 'error' (spawn failure etc.); settles as "failed".
+   * Kept separate from errorText, which also carries non-fatal notes
+   * (spill failures) that must not flip a clean exit to "failed". */
+  processErrored: boolean;
+  /** 'exit' event observed (code/signal recorded). */
+  exited: boolean;
+  /** 'close' event observed (stdio flushed; the settle trigger). */
+  stdioClosed: boolean;
+  /** A settle-after-spill-flush is in flight; don't start a second one. */
+  settling: boolean;
+  /** The shell exited without stdio closing; a bounded scope close is queued
+   * to reap descendants that still hold the inherited pipes open. */
+  exitCleanupStarted: boolean;
+  /** Completed exactly once when the entry settles. Kill callers and the scope
+   * finalizer can all await the same result without missing a notification. */
+  settled: Deferred.Deferred<void>;
+}
+
+export interface StartOptions {
+  readonly command: string;
+  readonly title: string;
+  readonly cwd: string;
+}
+
+export interface KillResult {
+  readonly id: string;
+  readonly title: string;
+  readonly status: TerminalStatus;
+  /** True when the entry was still running when this kill began. */
+  readonly wasRunning: boolean;
+  /** True when this call initiated the termination AND the entry settled as
+   * killed (a natural exit that won the race reports killed: false). */
+  readonly killed: boolean;
+  /** Final exit rendering ("exit 0", "SIGTERM", ...) captured at settle time,
+   * so reports stay accurate even if the entry is pruned afterwards. */
+  readonly exit: string;
+}
+
+// --- Read model ----------------------------------------------------------------
+
+/** Synchronous bridge for the TUI. Snapshots are live objects; do not mutate. */
+export interface TerminalReadModel {
+  list(): ReadonlyArray<TerminalSnapshot>;
+  get(id: string): TerminalSnapshot | undefined;
+  size(): number;
+  /** Any-change notification (widget, /ps list). */
+  subscribe(listener: () => void): () => void;
+  /** Per-terminal notification (/ps detail view). */
+  subscribeTo(id: string, listener: () => void): () => void;
+  /** Fire-and-forget kill (dashboard/detail `x`). Not marked consumed: the
+   * settle still flows back to the model as a follow-up message. */
+  requestKill(id: string): void;
+  /**
+   * Register the settle hook. `consumed` is true when an active bg_kill is
+   * collecting the result (so it must not also be delivered as a follow-up).
+   */
+  setOnSettled(
+    hook: ((snap: TerminalSnapshot, consumed: boolean) => void) | undefined,
+  ): void;
+}
+
+// --- Service --------------------------------------------------------------------
+
+export interface TerminalManagerShape {
+  start(
+    options: StartOptions,
+  ): Effect.Effect<TerminalSnapshot, SpawnError | ConcurrencyLimitError>;
+  status(id: string): Effect.Effect<TerminalSnapshot, UnknownTerminalError>;
+  /** Kill running terminals; resolves only after they have settled. */
+  kill(ids: ReadonlyArray<string>): Effect.Effect<ReadonlyArray<KillResult>>;
+  readonly list: Effect.Effect<ReadonlyArray<TerminalSnapshot>>;
+  readonly disposeAll: Effect.Effect<void>;
+  readonly view: TerminalReadModel;
+}
+
+export class TerminalManager extends Context.Service<
+  TerminalManager,
+  TerminalManagerShape
+>()("background-terminals/TerminalManager") {}
+
+// --- Process helpers ------------------------------------------------------------
+
+function shellInvocation(command: string) {
+  if (process.platform === "win32") {
+    const shell = process.env.ComSpec ?? "cmd.exe";
+    return { shell, args: ["/d", "/s", "/c", command] };
+  }
+  return { shell: "/bin/sh", args: ["-c", command] };
+}
+
+/** Signal the whole process group on POSIX so descendants (servers a shell
+ * command spawned) die with it; a wedged child must not orphan its tree. */
+function killTree(child: ChildProcess, signal: NodeJS.Signals) {
+  if (process.platform === "win32" && child.pid) {
+    try {
+      const killer = spawn(
+        "taskkill",
+        [
+          "/pid",
+          String(child.pid),
+          "/T",
+          ...(signal === "SIGKILL" ? ["/F"] : []),
+        ],
+        { stdio: "ignore", windowsHide: true },
+      );
+      killer.once("error", () => {
+        try {
+          child.kill(signal);
+        } catch {
+          // Process may already be gone.
+        }
+      });
+      killer.once("exit", (code) => {
+        if (code === 0) return;
+        try {
+          child.kill(signal);
+        } catch {
+          // Process may already be gone.
+        }
+      });
+      killer.unref();
+      return;
+    } catch {
+      // Fall through to the direct signal when taskkill cannot be launched.
+    }
+  }
+  if (process.platform !== "win32" && child.pid) {
+    try {
+      process.kill(-child.pid, signal);
+      return;
+    } catch {
+      // Group may already be gone; fall through to the direct signal.
+    }
+  }
+  try {
+    child.kill(signal);
+  } catch {
+    // Process may already be gone.
+  }
+}
+
+/** Await stdio closure without retaining a listener after interruption. */
+function awaitChildClose(child: ChildProcess, closed: () => boolean) {
+  return Effect.callback<void>((resume) => {
+    if (closed()) {
+      resume(Effect.void);
+      return;
+    }
+    const onClose = () => resume(Effect.void);
+    child.once("close", onClose);
+    return Effect.sync(() => child.off("close", onClose));
+  });
+}
+
+/** SIGTERM → deadline → SIGKILL; waits for stdio closure rather than only the
+ * shell's exit because descendants can keep the inherited pipes and process
+ * group alive after the shell itself is gone. */
+function terminateChild(
+  child: ChildProcess,
+  closed: () => boolean,
+  onSignal: () => void,
+) {
+  return Effect.suspend(() => {
+    if (closed()) return Effect.void;
+    return Effect.gen(function* () {
+      yield* Effect.sync(() => {
+        onSignal();
+        killTree(child, "SIGTERM");
+      });
+      yield* awaitChildClose(child, closed).pipe(
+        Effect.timeout(FORCE_KILL_AFTER_MS),
+        Effect.ignore,
+      );
+      if (closed()) return;
+      yield* Effect.sync(() => killTree(child, "SIGKILL"));
+      yield* awaitChildClose(child, closed).pipe(
+        Effect.timeout(500),
+        Effect.ignore,
+      );
+    });
+  });
+}
+
+// --- Implementation --------------------------------------------------------------
+
+const makeManager = Effect.gen(function* () {
+  // Scoped detached forker for sync contexts (read-model kills, process-event
+  // settlement, pruning). Completed fibers remove themselves; manager scope
+  // close interrupts any work that outlives the bounded disposeAll wait.
+  const cleanupFibers = yield* FiberSet.make();
+  const runCleanup = yield* FiberSet.runtime(cleanupFibers)();
+
+  const entries = new Map<string, Entry>();
+  /** Small immutable tombstones preserve truthful kill reports if pruning
+   * races the tool boundary after an id was validated. */
+  const settledHistory = new Map<
+    string,
+    Pick<KillResult, "title" | "status" | "exit">
+  >();
+  /** ids with an in-flight kill() collecting the result (settle → consumed). */
+  const killInterest = new Map<string, number>();
+  const listeners = new Set<() => void>();
+  const idListeners = new Map<string, Set<() => void>>();
+  let counter = 0;
+  let reserved = 0;
+  let disposed = false;
+  let spillDir: string | undefined | null;
+  let onSettled:
+    ((snap: TerminalSnapshot, consumed: boolean) => void) | undefined;
+
+  const notify = (id?: string) => {
+    for (const listener of [...listeners]) {
+      try {
+        listener();
+      } catch {
+        // A failed widget/render listener must not corrupt lifecycle state.
+      }
+    }
+    if (id) {
+      for (const listener of idListeners.get(id) ?? []) {
+        try {
+          listener();
+        } catch {
+          // Same.
+        }
+      }
+    }
+  };
+
+  const runningCount = () =>
+    [...entries.values()].filter((e) => e.snapshot.status === "running").length;
+
+  const addKillInterest = (ids: ReadonlyArray<string>) => {
+    for (const id of ids) killInterest.set(id, (killInterest.get(id) ?? 0) + 1);
+  };
+  const releaseKillInterest = (ids: ReadonlyArray<string>) => {
+    for (const id of ids) {
+      const count = (killInterest.get(id) ?? 1) - 1;
+      if (count <= 0) killInterest.delete(id);
+      else killInterest.set(id, count);
+    }
+  };
+
+  const closeEntryScope = (entry: Entry) =>
+    Scope.close(entry.scope, Exit.void).pipe(Effect.ignore);
+
+  const pruneSettled = () => {
+    if (entries.size <= MAX_TRACKED) return;
+    const candidates = [...entries.values()]
+      .filter(
+        (e) =>
+          e.snapshot.status !== "running" && !killInterest.has(e.snapshot.id),
+      )
+      .sort(
+        (a, b) =>
+          (a.snapshot.settledAt ?? a.snapshot.createdAt) -
+          (b.snapshot.settledAt ?? b.snapshot.createdAt),
+      );
+    for (const entry of candidates) {
+      if (entries.size <= MAX_TRACKED) break;
+      entries.delete(entry.snapshot.id);
+      runCleanup(closeEntryScope(entry));
+    }
+  };
+
+  /** End all spill streams; resolves when their buffers are flushed to disk
+   * (bounded), so a settle notification never points at a partial file. */
+  const flushSpillStreams = (entry: Entry) => {
+    const streams = entry.spillStreams;
+    entry.spillStreams = [];
+    return Effect.forEach(
+      streams,
+      (stream) =>
+        Effect.callback<void>((resume) => {
+          const done = () => resume(Effect.void);
+          try {
+            stream.end(done);
+          } catch {
+            // Best effort; tmpdir contents are disposable.
+            done();
+          }
+        }),
+      { concurrency: "unbounded", discard: true },
+    ).pipe(
+      Effect.timeoutOrElse({
+        duration: SPILL_FLUSH_TIMEOUT_MS,
+        orElse: () =>
+          Effect.sync(() => {
+            entry.stdoutBuf.spillPath = undefined;
+            entry.stderrBuf.spillPath = undefined;
+            entry.snapshot.errorText ??=
+              "Full-log spill flush timed out; full output may be incomplete";
+          }),
+      }),
+    );
+  };
+
+  /** Single settle path — idempotent; kill vs natural exit vs error races are
+   * resolved by whichever lands first (the second call is a no-op). */
+  const settle = (entry: Entry) => {
+    const s = entry.snapshot;
+    if (s.status !== "running") return;
+    s.settledAt = Date.now();
+    s.status = entry.killSignaled
+      ? "killed"
+      : entry.processErrored
+        ? "failed"
+        : s.exitCode === 0
+          ? "done"
+          : "failed";
+    settledHistory.set(s.id, {
+      title: s.title,
+      status: s.status,
+      exit: formatExit(s),
+    });
+    while (settledHistory.size > MAX_SETTLED_HISTORY) {
+      const oldest = settledHistory.keys().next().value;
+      if (oldest === undefined) break;
+      settledHistory.delete(oldest);
+    }
+    // Completing the Deferred can immediately resume kill waiters, whose
+    // ensuring blocks release interest. Snapshot consumption first so the
+    // settle hook observes the interest that existed when settlement won.
+    const consumed = (killInterest.get(s.id) ?? 0) > 0;
+    Deferred.doneUnsafe(entry.settled, Effect.void);
+    notify(s.id);
+    try {
+      // During teardown, don't queue results into a shutting-down session.
+      if (!disposed) onSettled?.(s, consumed);
+    } catch {
+      // The parent session may be unavailable; settlement stays final.
+    }
+    pruneSettled();
+  };
+
+  /** Flush the spill files, then settle: the completion follow-up (and the
+   * kill() resolution) reference the spill path, so the full capture must be
+   * on disk before anyone is told about it. Idempotent via `settling`. */
+  const settleAfterFlush = (entry: Entry) => {
+    if (entry.settling || entry.snapshot.status !== "running") return;
+    entry.settling = true;
+    runCleanup(
+      flushSpillStreams(entry).pipe(
+        Effect.andThen(Effect.sync(() => settle(entry))),
+      ),
+    );
+  };
+
+  const scheduleExitCleanup = (entry: Entry) => {
+    if (entry.exitCleanupStarted) return;
+    entry.exitCleanupStarted = true;
+    runCleanup(
+      Effect.sleep(SETTLE_GRACE_MS).pipe(
+        Effect.andThen(
+          Effect.suspend(() =>
+            entry.snapshot.status === "running" && !entry.stdioClosed
+              ? closeEntryScope(entry).pipe(
+                  Effect.timeout(STOP_TIMEOUT_MS),
+                  Effect.ignore,
+                )
+              : Effect.void,
+          ),
+        ),
+      ),
+    );
+  };
+
+  const resolveSpillDir = () => {
+    if (spillDir !== undefined) return spillDir ?? undefined;
+    try {
+      const base = path.join(os.tmpdir(), "pi-background-terminals");
+      fs.mkdirSync(base, { recursive: true, mode: 0o700 });
+      fs.chmodSync(base, 0o700);
+      spillDir = fs.mkdtempSync(path.join(base, "session-"));
+      fs.chmodSync(spillDir, 0o700);
+    } catch {
+      spillDir = null;
+    }
+    return spillDir ?? undefined;
+  };
+
+  const makeSpill = (
+    entry: () => Entry | undefined,
+    id: string,
+    stream: "stdout" | "stderr",
+    resumeSource: () => void,
+  ) => {
+    const dir = resolveSpillDir();
+    if (!dir) return undefined;
+    const spillPath = path.join(dir, `${id}.${stream}.log`);
+    try {
+      const file = fs.createWriteStream(spillPath, {
+        flags: "a",
+        mode: 0o600,
+      });
+      let broken = false;
+      let capped = false;
+      let writtenBytes = 0;
+      file.on("error", (error) => {
+        broken = true;
+        resumeSource();
+        const current = entry();
+        if (current) {
+          const buf =
+            stream === "stdout" ? current.stdoutBuf : current.stderrBuf;
+          buf.spillPath = undefined;
+          current.snapshot.errorText ??= bounded(
+            `Full-log spill to ${spillPath} failed: ${boundedError(error)}`,
+          );
+        }
+      });
+      return {
+        spillPath,
+        file,
+        write: (chunk: string) => {
+          // writableEnded guard: late 'data' after the settle flush must not
+          // error the ended stream (and falsely report the spill as broken).
+          if (broken || capped || file.writableEnded) return true;
+          const chunkBytes = Buffer.byteLength(chunk, "utf8");
+          if (writtenBytes + chunkBytes > MAX_SPILL_BYTES_PER_STREAM) {
+            capped = true;
+            const current = entry();
+            if (current) {
+              const buf =
+                stream === "stdout" ? current.stdoutBuf : current.stderrBuf;
+              buf.spillPath = undefined;
+              current.snapshot.errorText ??= bounded(
+                `${stream} full-log spill reached the ${MAX_SPILL_BYTES_PER_STREAM}-byte safety limit`,
+              );
+            }
+            return true;
+          }
+          writtenBytes += chunkBytes;
+          const accepted = file.write(chunk);
+          if (!accepted) file.once("drain", resumeSource);
+          return accepted;
+        },
+      };
+    } catch {
+      return undefined;
+    }
+  };
+
+  const start = (options: StartOptions) =>
+    Effect.gen(function* () {
+      // Reserve synchronously (before the first yield inside doStart) so
+      // parallel tool calls cannot race past the cap.
+      yield* Effect.suspend(
+        (): Effect.Effect<void, SpawnError | ConcurrencyLimitError> => {
+          if (disposed) {
+            return new SpawnError({
+              message: "Background terminal manager is shutting down.",
+            });
+          }
+          if (runningCount() + reserved >= MAX_RUNNING) {
+            return new ConcurrencyLimitError({
+              message: `Max ${MAX_RUNNING} background terminals can run concurrently. Stop one with bg_kill before starting another.`,
+            });
+          }
+          reserved++;
+          return Effect.void;
+        },
+      );
+
+      const doStart = Effect.gen(function* () {
+        const { shell, args } = shellInvocation(options.command);
+        const child = yield* Effect.try({
+          try: () =>
+            spawn(shell, args, {
+              cwd: options.cwd,
+              env: process.env,
+              // stdin IGNORED: there is no input surface, ever. A process
+              // that reads stdin sees EOF immediately.
+              stdio: ["ignore", "pipe", "pipe"],
+              // Own process group on POSIX → group kill takes the whole tree.
+              detached: process.platform !== "win32",
+            }),
+          catch: (error) => new SpawnError({ message: boundedError(error) }),
+        });
+
+        const id = `bt-${++counter}`;
+        const entryRef = () => entries.get(id);
+        const stdoutSpill = makeSpill(entryRef, id, "stdout", () =>
+          child.stdout?.resume(),
+        );
+        const stderrSpill = makeSpill(entryRef, id, "stderr", () =>
+          child.stderr?.resume(),
+        );
+        const stdoutBuf = new OutputBuffer(
+          RETAINED_PER_STREAM,
+          stdoutSpill?.write,
+        );
+        const stderrBuf = new OutputBuffer(
+          RETAINED_PER_STREAM,
+          stderrSpill?.write,
+        );
+        stdoutBuf.spillPath = stdoutSpill?.spillPath;
+        stderrBuf.spillPath = stderrSpill?.spillPath;
+
+        const snapshot: MutableSnapshot = {
+          id,
+          command: options.command,
+          title: options.title,
+          cwd: options.cwd,
+          pid: child.pid,
+          status: "running",
+          createdAt: Date.now(),
+          get stdout() {
+            return stdoutBuf.view();
+          },
+          get stderr() {
+            return stderrBuf.view();
+          },
+        };
+
+        const scope = yield* Scope.make();
+        const settled = yield* Deferred.make<void>();
+        const entry: Entry = {
+          snapshot,
+          child,
+          scope,
+          stdoutBuf,
+          stderrBuf,
+          spillStreams: [stdoutSpill?.file, stderrSpill?.file].filter(
+            (file): file is fs.WriteStream => file !== undefined,
+          ),
+          killSignaled: false,
+          processErrored: false,
+          exited: false,
+          stdioClosed: false,
+          settling: false,
+          exitCleanupStarted: false,
+          settled,
+        };
+
+        // Plain-callback stream plumbing (the codex-backend precedent):
+        // setEncoding's internal StringDecoder is multibyte-safe across
+        // chunk boundaries.
+        child.stdout?.setEncoding("utf8");
+        child.stdout?.on("data", (chunk: string) => {
+          if (!stdoutBuf.push(chunk)) child.stdout?.pause();
+          notify(id);
+        });
+        child.stderr?.setEncoding("utf8");
+        child.stderr?.on("data", (chunk: string) => {
+          if (!stderrBuf.push(chunk)) child.stderr?.pause();
+          notify(id);
+        });
+        // Spawn failures (ENOENT etc.) arrive via 'error', not a throw. Node
+        // still emits 'close' afterwards (with a bogus errno as code), so
+        // record the failure here and let the close path do the one settle.
+        child.once("error", (error) => {
+          entry.processErrored = true;
+          snapshot.errorText ??= boundedError(error);
+          entry.exited = true;
+          settleAfterFlush(entry);
+        });
+        // Record code/signal on 'exit'; settle on 'close' so the completion
+        // notification always carries the final flushed output.
+        child.once("exit", (code, signal) => {
+          entry.exited = true;
+          snapshot.exitCode = code ?? undefined;
+          snapshot.signal = signal ?? undefined;
+          // A descendant can keep the pipes open after the shell exits. Give
+          // close a short natural grace, then close the scope to terminate
+          // the surviving process group and force a bounded settlement.
+          scheduleExitCleanup(entry);
+        });
+        child.once("close", (code, signal) => {
+          entry.exited = true;
+          entry.stdioClosed = true;
+          // Only trust close's code/signal when 'exit' never fired (a spawn
+          // 'error' close reports the errno, e.g. -2, as its code).
+          if (!entry.processErrored) {
+            snapshot.exitCode ??= code ?? undefined;
+            snapshot.signal ??= signal ?? undefined;
+          }
+          settleAfterFlush(entry);
+        });
+
+        // One teardown path: kill(), requestKill, pruning, disposeAll, and
+        // runtime.dispose() all converge on closing this scope.
+        yield* Scope.provide(
+          Effect.addFinalizer(() =>
+            Effect.gen(function* () {
+              // Only claim "killed" when we are actually about to signal a
+              // live process; a natural exit that already happened (still
+              // waiting on 'close') keeps its truthful done/failed status.
+              yield* terminateChild(
+                child,
+                () => entry.stdioClosed,
+                () => {
+                  entry.killSignaled ||=
+                    !entry.exited && entry.snapshot.status === "running";
+                },
+              );
+              // Give the natural close→flush→settle path a bounded grace,
+              // then force the settle: a grandchild holding the pipe open
+              // (detached into a new group) must not leave the entry
+              // "running" forever.
+              if (entry.snapshot.status === "running") {
+                yield* Deferred.await(entry.settled).pipe(
+                  Effect.timeout(SETTLE_GRACE_MS),
+                  Effect.ignore,
+                );
+              }
+              if (entry.snapshot.status === "running" && !entry.settling) {
+                // Force the settle ourselves. When `settling` is set, the
+                // close path's flush→settle is already in flight (bounded by
+                // SPILL_FLUSH_TIMEOUT_MS) — settling here first would cite a
+                // spill file that is still being flushed.
+                if (!entry.stdioClosed) {
+                  entry.snapshot.errorText ??=
+                    "stdio did not close after termination; output may be incomplete";
+                }
+                entry.settling = true;
+                yield* flushSpillStreams(entry);
+                settle(entry);
+              }
+            }),
+          ),
+          scope,
+        );
+
+        // disposeAll may have swept the entries map while we were setting up;
+        // an entry added after the sweep would never be torn down. Close our
+        // own scope (kills the child) and fail instead (subagents precedent).
+        if (disposed) {
+          yield* closeEntryScope(entry);
+          return yield* new SpawnError({
+            message: "Background terminal manager shut down while starting.",
+          });
+        }
+        entries.set(id, entry);
+        notify(id);
+        return snapshot as TerminalSnapshot;
+      });
+
+      // Uninterruptible: between spawn() and entries.set there must be no
+      // window where an interrupt (tool abort, runtime dispose) leaves a
+      // live child that no scope/registry knows about. All steps are sync.
+      return yield* doStart.pipe(
+        Effect.uninterruptible,
+        Effect.ensuring(
+          Effect.sync(() => {
+            reserved--;
+            notify();
+          }),
+        ),
+      );
+    });
+
+  const status = (id: string) =>
+    Effect.suspend(
+      (): Effect.Effect<TerminalSnapshot, UnknownTerminalError> => {
+        const entry = entries.get(id);
+        if (!entry) {
+          const known = [...entries.keys()];
+          return new UnknownTerminalError({
+            message: `Unknown terminal id "${id}". Known: ${known.join(", ") || "none"}.`,
+          });
+        }
+        return Effect.succeed(entry.snapshot as TerminalSnapshot);
+      },
+    );
+
+  /** Kill one running entry: close the scope — whose finalizer marks the kill
+   * at the signal point, terminates the tree, and force-settles —
+   * in a DETACHED fiber. Once the flag is set the termination must actually
+   * happen; a tool abort interrupting the caller cannot cancel it (this is
+   * what makes "termination continues in the background" truthful). */
+  const killEntry = (entry: Entry) =>
+    Effect.sync(() => {
+      if (entry.snapshot.status !== "running") return;
+      runCleanup(
+        closeEntryScope(entry).pipe(
+          Effect.timeout(STOP_TIMEOUT_MS),
+          Effect.ignore,
+        ),
+      );
+    });
+
+  const kill = (ids: ReadonlyArray<string>) =>
+    Effect.suspend(() => {
+      const unique = [...new Set(ids)];
+      const byId = new Map(
+        unique
+          .map((id) => entries.get(id))
+          .filter((entry): entry is Entry => entry !== undefined)
+          .map((entry) => [entry.snapshot.id, entry]),
+      );
+      const running = [...byId.values()].filter(
+        (entry) => entry.snapshot.status === "running",
+      );
+      const runningIds = running.map((entry) => entry.snapshot.id);
+      // Mark consumed before signaling so this kill's settlements are not
+      // ALSO queued as automatic follow-up messages to the model.
+      addKillInterest(runningIds);
+      const work = Effect.gen(function* () {
+        yield* Effect.forEach(running, killEntry, {
+          concurrency: "unbounded",
+        });
+        // Every caller waits on the entries that were running when its kill
+        // began. Deferred completion cannot be missed and supports concurrent
+        // overlapping/multi-id kill calls.
+        yield* Effect.forEach(
+          running,
+          (entry) => Deferred.await(entry.settled),
+          { concurrency: "unbounded", discard: true },
+        );
+        // Capture the report BEFORE the ensuring below releases interest and
+        // prunes — a just-settled entry must not vanish out from under it.
+        return unique.map((id): KillResult => {
+          const snapshot = byId.get(id)?.snapshot;
+          const history = settledHistory.get(id);
+          const status = snapshot?.status ?? history?.status ?? "killed";
+          const wasRunning = runningIds.includes(id);
+          return {
+            id,
+            title: snapshot?.title ?? history?.title ?? "?",
+            status,
+            wasRunning,
+            // A natural exit can win the race with our SIGTERM; report what
+            // actually happened rather than claiming the kill did it.
+            killed: wasRunning && status === "killed",
+            exit: snapshot
+              ? formatExit(snapshot)
+              : (history?.exit ?? "unknown"),
+          };
+        });
+      });
+      return work.pipe(
+        Effect.ensuring(
+          Effect.sync(() => {
+            releaseKillInterest(runningIds);
+            pruneSettled();
+          }),
+        ),
+      );
+    });
+
+  const disposeAll = Effect.gen(function* () {
+    disposed = true;
+    const all = [...entries.values()];
+    entries.clear();
+    yield* Effect.forEach(
+      all,
+      (entry) =>
+        closeEntryScope(entry).pipe(
+          Effect.timeout(STOP_TIMEOUT_MS),
+          Effect.ignore,
+        ),
+      { concurrency: "unbounded" },
+    );
+    // Detached kill/prune/flush work is scoped to the manager. Wait for it
+    // within the shutdown bound; the FiberSet finalizer interrupts anything
+    // still live when the manager scope closes, so cleanup cannot leak.
+    yield* FiberSet.awaitEmpty(cleanupFibers).pipe(
+      Effect.timeout(STOP_TIMEOUT_MS),
+      Effect.ignore,
+    );
+    yield* Effect.sync(() => {
+      const dir = spillDir;
+      spillDir = null;
+      if (dir) fs.rmSync(dir, { recursive: true, force: true });
+    });
+    yield* Effect.sync(() => notify());
+  });
+
+  const view: TerminalReadModel = {
+    list: () => [...entries.values()].map((entry) => entry.snapshot),
+    get: (id) => entries.get(id)?.snapshot,
+    size: () => entries.size,
+    subscribe: (listener) => {
+      listeners.add(listener);
+      return () => listeners.delete(listener);
+    },
+    subscribeTo: (id, listener) => {
+      let set = idListeners.get(id);
+      if (!set) {
+        set = new Set();
+        idListeners.set(id, set);
+      }
+      set.add(listener);
+      return () => {
+        set.delete(listener);
+        if (set.size === 0) idListeners.delete(id);
+      };
+    },
+    requestKill: (id) => {
+      const entry = entries.get(id);
+      if (!entry) return;
+      // UI-initiated kills are not "consumed": the killed result still flows
+      // back to the model as a follow-up message (subagents precedent).
+      runCleanup(killEntry(entry).pipe(Effect.ignore));
+    },
+    setOnSettled: (hook) => {
+      onSettled = hook;
+    },
+  };
+
+  // Safety net: disposing the ManagedRuntime tears everything down even if
+  // the extension forgot to call disposeAll explicitly.
+  yield* Effect.addFinalizer(() => disposeAll);
+
+  return TerminalManager.of({
+    start,
+    status,
+    kill,
+    list: Effect.sync(() => [...entries.values()].map((e) => e.snapshot)),
+    disposeAll,
+    view,
+  });
+});
+
+export const TerminalManagerLive: Layer.Layer<TerminalManager> = Layer.effect(
+  TerminalManager,
+  makeManager,
+);