@bastani/atomic 0.6.8 → 0.7.0-2

package/src/sdk/providers/claude.ts

@@ -1,1450 +0,0 @@
-/**
- * Claude Code query abstraction.
- *
- * Sends a prompt to an interactive Claude Code session running in a tmux pane
- * using `tmux send-keys -l --` (literal text) + `C-m` (raw carriage return).
- * Verifies delivery by polling `capture-pane` and retries if needed.
- *
- * This is NOT headless — Claude runs as a full interactive TUI in the pane.
- * We're automating keyboard input and reading pane output.
- *
- * Reliability hardened from oh-my-codex's sendToWorker implementation:
- * - Pre-send readiness wait with exponential backoff
- * - CLI-specific submit plan (Claude: 1 C-m per round)
- * - Per-round capture verification (6 rounds)
- * - Adaptive retry with C-u clear + retype
- * - Post-submit active-task detection
- * - File-based idle detection via session JSONL watching
- */
-
-import {
-  getSessionMessages,
-  query as sdkQuery,
-  type SessionMessage,
-  type SDKUserMessage,
-  type Options as SDKOptions,
-} from "@anthropic-ai/claude-agent-sdk";
-import { respawnPane } from "../runtime/tmux.ts";
-import { escBash } from "../runtime/executor.ts";
-import { watch, unlink, mkdir, writeFile } from "node:fs/promises";
-import { existsSync, writeFileSync } from "node:fs";
-import { join } from "node:path";
-import { randomUUID } from "node:crypto";
-import { claudeHookDirs } from "../../commands/cli/claude-stop-hook.ts";
-import {
-  clearInflightTracking,
-  waitForInflightDrained,
-} from "../../commands/cli/claude-inflight-hook.ts";
-import { resolveAdditionalInstructionsContent } from "../../services/config/additional-instructions.ts";
-import {
-  atomicContentTempPath,
-  atomicTempPath,
-  withAtomicTempEnv,
-} from "../../lib/atomic-temp.ts";
-
-// ---------------------------------------------------------------------------
-// Session tracking — ensures createClaudeSession is called before claudeQuery
-// ---------------------------------------------------------------------------
-
-/** Per-pane state for Claude sessions. */
-interface PaneState {
-  /**
-   * Claude Code's session ID. Pre-generated via `crypto.randomUUID()` in
-   * `createClaudeSession` and passed to `claude --session-id <UUID>` on the
-   * first query, so we know the JSONL filename without polling.
-   */
-  claudeSessionId: string;
-  /** Whether the `claude` CLI has been spawned in this pane yet. */
-  claudeStarted: boolean;
-  /** CLI flags to pass to `claude` when it is spawned on the first query. */
-  chatFlags: string[];
-}
-
-const initializedPanes = new Map<string, PaneState>();
-
-/**
- * Remove a pane from the initialized map and signal the currently-blocked
- * Stop hook that the session is over, so Claude stops promptly instead of
- * waiting out the hook's safety timeout.
- *
- * Called by the runtime when a Claude stage is being torn down. Idempotent.
- *
- * After writing the release marker, this waits for the per-session in-flight
- * marker dir (`~/.atomic/claude-inflight/<session_id>/`) to drain. The
- * marker dir is populated by the SubagentStart/Stop and TaskCreated/Completed
- * hooks registered in {@link WORKFLOW_HOOK_SETTINGS}. This wait is the
- * synchronization barrier that prevents the executor from advancing to the
- * next stage while the previous stage's backgrounded subagents/tasks still
- * hold FDs/PTYs on the atomic tmux server — the failure mode that surfaced
- * intermittently as `tmux respawn-pane: fork failed: Device not configured`.
- *
- * The wait has its own bounded timeout (default 30 minutes) so a wedged
- * subagent can't permanently block the workflow; the in-hook stale-sweep
- * (~2 hours TTL) is the ultimate safety net.
- */
-export async function clearClaudeSession(paneId: string): Promise<void> {
-  const state = initializedPanes.get(paneId);
-  if (state) {
-    try {
-      await releaseClaudeSession(state.claudeSessionId);
-    } catch {
-      // Best-effort — if release fails the hook will still exit on its
-      // own safety timeout.
-    }
-    // Wait for in-flight subagents/tasks to finish before letting the
-    // executor advance. Resolves immediately when the dir is empty/missing
-    // (the common case, including any stage that didn't spawn subagents).
-    try {
-      await waitForInflightDrained(state.claudeSessionId);
-    } catch {
-      // Best-effort — the wait swallows internal errors and resolves on
-      // timeout. A throw here would only happen on a path bug.
-    }
-    try {
-      await unlinkAtomicPidFile(state.claudeSessionId);
-    } catch {
-      // Best-effort — stale pid file is inert; the next session writes a
-      // fresh one under its own UUID.
-    }
-    try {
-      await clearStaleReadyMarker(state.claudeSessionId);
-    } catch {
-      // Best-effort — stale ready marker is inert; the next session writes
-      // a fresh one under its own UUID and clears any prior leftover in
-      // `claudeQuery` before respawn.
-    }
-    try {
-      await clearInflightTracking(state.claudeSessionId);
-    } catch {
-      // Best-effort — leftover marker files are reaped by the next session's
-      // stale-sweep, and the .session-roots/ entries are tiny.
-    }
-  }
-  initializedPanes.delete(paneId);
-}
-
-/** Default CLI flags passed to the `claude` command. */
-const DEFAULT_CHAT_FLAGS = [
-  "--allow-dangerously-skip-permissions",
-  "--dangerously-skip-permissions",
-];
-
-/**
- * Build the shell command Claude Code runs from an injected workflow hook.
- *
- * - **Published install** (`import.meta.dir` under `node_modules`): resolve
- *   `atomic` via the user's PATH. That's the binary they installed, and
- *   relying on PATH is robust across shells and platforms.
- * - **Dev** (source checkout): re-invoke THIS repo's `src/cli.ts` using the
- *   same Bun runtime that's executing us, so edits to the hook logic are
- *   picked up without rebuilding or re-linking. Mirrors the
- *   `spawnAttachedFooter` pattern in `src/sdk/runtime/executor.ts:293-303`.
- *
- * The dev-detection heuristic (`node_modules` in `import.meta.dir`) is the
- * same one used by `src/services/system/auto-sync.ts:50`.
- */
-function buildWorkflowHookCommand(subcommand: string, extraArgs: readonly string[] = []): string {
-  if (import.meta.dir.includes("node_modules")) {
-    return ["atomic", subcommand, ...extraArgs].join(" ");
-  }
-  const runtime = process.execPath;
-  const cliPath = join(import.meta.dir, "..", "..", "cli.ts");
-  if (process.platform === "win32") {
-    const script = [
-      quotePwshLiteral(runtime),
-      quotePwshLiteral(cliPath),
-      quotePwshLiteral(subcommand),
-      ...extraArgs.map(quotePwshLiteral),
-    ].join(" ");
-    const encoded = Buffer.from(`& ${script}`, "utf16le").toString("base64");
-    return `pwsh -NoProfile -EncodedCommand ${encoded}`;
-  }
-  return [
-    `"${escBash(runtime)}"`,
-    `"${escBash(cliPath)}"`,
-    subcommand,
-    ...extraArgs,
-  ].join(" ");
-}
-
-function quotePwshLiteral(s: string): string {
-  return `'${s
-    .replace(/\x00/g, "")
-    .replace(/[\n\r]+/g, " ")
-    .replace(/'/g, "''")}'`;
-}
-
-/**
- * Effectively-unbounded timeout (in seconds) for the Stop hook command.
- *
- * Claude Code's Stop hook process runs three phases sequentially — the
- * initial Stop hooks, then TaskCompleted hooks (per in-progress task owned
- * by the teammate), then TeammateIdle hooks — and the per-hook `timeout`
- * applies to the entire lifecycle. Under Claude Code's default (10 min),
- * a turn that leaves tasks in-progress (e.g. via the TaskList/TodoWrite
- * tool) can blow the budget and get killed, which also severs our
- * `_claude-stop-hook`'s queue/release poll and strands the workflow.
- *
- * ~24 days — the max safe `setTimeout` value (2^31 - 1 ms) expressed in
- * seconds — removes the timeout in practical terms. `waitForIdle`'s
- * marker-file watch still fires as soon as our initial hook writes the
- * marker, so the workflow proceeds on real hook completion, not on timer
- * expiry.
- */
-const STOP_HOOK_TIMEOUT_SECONDS = 2_147_483;
-
-/**
- * Effectively-unbounded ms ceiling for `waitForReadyMarker`. Mirrors
- * {@link STOP_HOOK_TIMEOUT_SECONDS} but expressed in ms for `setTimeout`.
- *
- * The SessionStart hook fires well under a second on a working spawn, so in
- * practice this timer never expires. It only protects against failure modes
- * where the hook will never fire at all (claude binary missing, hook
- * command not resolvable, settings JSON rejected), where a clear error
- * beats a hung pane.
- */
-const READY_HOOK_TIMEOUT_MS = 2_147_483_000;
-
-/**
- * Inline settings injected via `claude --settings <json>` on every workflow
- * spawn. Registers the workflow-owned hooks without relying on
- * `.claude/settings.json` — so the hooks fire only for workflow-spawned
- * Claude sessions, not when a user runs `claude` manually.
- *
- * Registered hooks:
- *   - `Stop`: deliver queued follow-up prompts via `{decision:"block"}` and
- *     write an idle-marker file that `waitForIdle` watches. `timeout` is
- *     set to {@link STOP_HOOK_TIMEOUT_SECONDS} so the hook survives long
- *     TaskCompleted/TeammateIdle phases — see the constant's docstring.
- *   - `PreToolUse` matched on `AskUserQuestion`: write
- *     `~/.atomic/claude-hil/<session_id>` so `watchHILMarker` can fire
- *     `onHIL(true)` — the node card flips to the blue "awaiting_input" pulse.
- *   - `PostToolUse` / `PostToolUseFailure` matched on `AskUserQuestion`:
- *     remove the HIL marker. Claude Code fires exactly one of these per
- *     tool invocation (PostToolUse on success, PostToolUseFailure in the
- *     catch path — see `src/services/tools/toolExecution.ts` in the CLI
- *     source), so registering the same command on both guarantees the
- *     marker clears regardless of which completion path the tool takes.
- *   - `SubagentStart` / `SubagentStop`: maintain a per-root-session marker
- *     dir under `~/.atomic/claude-inflight/<root>/` so the Stop hook and
- *     `clearClaudeSession` can both gate on subagent completion before
- *     letting the stage advance. Without this gate, a stage that spawned
- *     `run_in_background: true` subagents would tear down its pane while
- *     children still hold FDs/PTYs on the atomic tmux server, intermittently
- *     surfacing as `tmux respawn-pane: fork failed: Device not configured`
- *     when the next stage tried to spawn.
- *   - `TeammateIdle`: same gating applied at agent-team teammate idle.
- *     Unlike Stop, this fires when a teammate (potentially a different
- *     `session_id` from the stage's root) goes idle, so we route it to a
- *     focused `_claude-inflight-hook wait` mode that only awaits in-flight
- *     drain — no claude-stop marker write (that would confuse `waitForIdle`)
- *     and no queue/release polling (those are keyed on the stage's root).
- *
- * Built once at module load. Contains no single quotes (JSON syntax doesn't
- * produce them and paths rarely do), so POSIX single-quoting at the spawn
- * site is sufficient shell escaping.
- */
-const WORKFLOW_HOOK_SETTINGS = JSON.stringify({
-  hooks: {
-    SessionStart: [
-      {
-        matcher: "startup",
-        hooks: [
-          {
-            type: "command",
-            command: buildWorkflowHookCommand("_claude-session-start-hook"),
-          },
-        ],
-      },
-    ],
-    Stop: [
-      {
-        hooks: [
-          {
-            type: "command",
-            command: buildWorkflowHookCommand("_claude-stop-hook"),
-            timeout: STOP_HOOK_TIMEOUT_SECONDS,
-          },
-        ],
-      },
-    ],
-    PreToolUse: [
-      {
-        matcher: "AskUserQuestion",
-        hooks: [
-          {
-            type: "command",
-            command: buildWorkflowHookCommand("_claude-ask-hook", ["enter"]),
-          },
-        ],
-      },
-    ],
-    PostToolUse: [
-      {
-        matcher: "AskUserQuestion",
-        hooks: [
-          {
-            type: "command",
-            command: buildWorkflowHookCommand("_claude-ask-hook", ["exit"]),
-          },
-        ],
-      },
-    ],
-    PostToolUseFailure: [
-      {
-        matcher: "AskUserQuestion",
-        hooks: [
-          {
-            type: "command",
-            command: buildWorkflowHookCommand("_claude-ask-hook", ["exit"]),
-          },
-        ],
-      },
-    ],
-    // SubagentStart/SubagentStop fire per Agent-tool dispatch (no matcher)
-    // and route to a single subcommand that touches/removes one marker file
-    // per `agent_id`. The handler is bulletproof — any error exits 0
-    // silently — so a hook failure can't kill the stage.
-    SubagentStart: [
-      {
-        hooks: [
-          {
-            type: "command",
-            command: buildWorkflowHookCommand("_claude-inflight-hook", ["start"]),
-          },
-        ],
-      },
-    ],
-    SubagentStop: [
-      {
-        hooks: [
-          {
-            type: "command",
-            command: buildWorkflowHookCommand("_claude-inflight-hook", ["stop"]),
-          },
-        ],
-      },
-    ],
-    // TeammateIdle gets a focused `wait` mode (gates on in-flight drain,
-    // nothing else) — see the WORKFLOW_HOOK_SETTINGS docstring for why this
-    // doesn't reuse the Stop hook handler. Timeout matches Stop's so the
-    // wait can run for as long as the workflow holds onto teammates.
-    TeammateIdle: [
-      {
-        hooks: [
-          {
-            type: "command",
-            command: buildWorkflowHookCommand("_claude-inflight-hook", ["wait"]),
-            timeout: STOP_HOOK_TIMEOUT_SECONDS,
-          },
-        ],
-      },
-    ],
-  },
-});
-
-// ---------------------------------------------------------------------------
-// createClaudeSession
-// ---------------------------------------------------------------------------
-
-export interface ClaudeSessionOptions {
-  /** tmux pane ID where Claude should be started */
-  paneId: string;
-  /** CLI flags to pass to the `claude` command (default: ["--allow-dangerously-skip-permissions", "--dangerously-skip-permissions"]) */
-  chatFlags?: string[];
-}
-
-/**
- * Initialize per-pane Claude state. Does NOT spawn the `claude` CLI — the
- * pane is left as a bare shell. The CLI is spawned lazily on the first
- * `claudeQuery()` call, with the prompt baked into the spawn command:
- *
- *     claude [chatFlags] --session-id <UUID> 'Read the prompt in <tmpfile>'
- *
- * Pre-generating the session UUID here lets the first query pass it to the
- * CLI, so we know the JSONL filename up front and can skip discovery polling.
- *
- * Must be called before any `claudeQuery()` calls targeting the same pane.
- *
- * @example
- * ```typescript
- * import { createClaudeSession, claudeQuery } from "@bastani/atomic/workflows";
- *
- * await createClaudeSession({ paneId: ctx.paneId });
- * await claudeQuery({ paneId: ctx.paneId, prompt: "Describe this project" });
- * ```
- *
- * @example
- * ```typescript
- * // With custom flags
- * await createClaudeSession({
- *   paneId: ctx.paneId,
- *   chatFlags: ["--model", "opus", "--dangerously-skip-permissions"],
- * });
- * ```
- */
-export async function createClaudeSession(options: ClaudeSessionOptions): Promise<string> {
-  const { paneId, chatFlags = DEFAULT_CHAT_FLAGS } = options;
-
-  const claudeSessionId = randomUUID();
-  initializedPanes.set(paneId, {
-    claudeSessionId,
-    claudeStarted: false,
-    chatFlags,
-  });
-
-  // Write our PID so the Stop hook can detect an orphaned session if we
-  // crash/get SIGKILL'd without running teardown. Best-effort; failures just
-  // mean the hook falls back to waiting out Claude's own hook timeout.
-  await writeAtomicPidFile(claudeSessionId);
-
-  return claudeSessionId;
-}
-
-/**
- * Build the short, single-line natural-language prompt we send to Claude
- * (either as spawn argv or as a follow-up message). Claude's first action
- * is then a Read tool call against `promptFile` — which sidesteps shell
- * escaping, ARG_MAX, and tmux paste-buffer flakiness for large prompts.
- *
- * The session dir and filename are slug-based (`prompt-<N>.txt` under
- * `~/.atomic/sessions/...`), so they never contain shell-special characters.
- */
-function readPromptInstruction(promptFile: string): string {
-  return `Read ${promptFile} and follow the instructions inside.`;
-}
-
-/**
- * Spawn `claude` in the pane with the prompt baked in via the Read tool.
- *
- * The prompt is already written to `promptFile` by the caller. The spawn
- * argv is `'Read the prompt in <path>'`, so Claude's first action is a Read
- * tool call against that file.
- */
-async function spawnClaudeWithPrompt(
-  paneId: string,
-  promptFile: string,
-  chatFlags: string[],
-  sessionId: string,
-): Promise<void> {
-  const settingsPath = workflowHookSettingsPath();
-  const argvPrompt = `"${escBash(readPromptInstruction(promptFile))}"`;
-  const cmd = [
-    "claude",
-    ...chatFlags,
-    // Workflow-owned hooks. Placed AFTER chatFlags so commander's last-wins
-    // semantics shadow any user-provided --settings, making this
-    // non-overridable by `.atomic/settings.json` chatFlags overrides. Passing
-    // a path avoids Claude Code's content-hashed /tmp/claude-settings*.json.
-    "--settings",
-    `"${escBash(settingsPath)}"`,
-    "--session-id",
-    sessionId,
-    argvPrompt,
-  ].join(" ");
-
-  // Replace the pane's shell with `claude` directly. tmux execs the command
-  // itself, so there's no shell line editor to race with — the previous
-  // approach keystroked into a zsh that hadn't finished ZLE init yet, and
-  // zsh's TCSAFLUSH during startup would discard the buffered `\r`, leaving
-  // the command typed at the prompt but never submitted.
-  respawnPane(paneId, cmd);
-
-  // Positive readiness signal: wait for Claude's SessionStart hook (matcher
-  // `startup`) to write `~/.atomic/claude-ready/<session_id>`. This fires
-  // before Claude writes the JSONL transcript, so it beats the old
-  // transcript-file race and is deterministic.
-  await waitForReadyMarker(sessionId);
-}
-
-function workflowHookSettingsPath(): string {
-  const path = atomicContentTempPath(
-    "claude-settings-atomic",
-    ".json",
-    WORKFLOW_HOOK_SETTINGS,
-  );
-  writeFileSync(path, WORKFLOW_HOOK_SETTINGS, {
-    encoding: "utf-8",
-    mode: 0o600,
-  });
-  return path;
-}
-
-/**
- * Wait for the SessionStart hook's ready marker at
- * `~/.atomic/claude-ready/<session_id>`.
- *
- * `atomic _claude-session-start-hook` is registered in
- * {@link WORKFLOW_HOOK_SETTINGS} with matcher `startup`; the Claude CLI
- * dispatches it during spawn, before the first API call and before the JSONL
- * transcript is created. Waiting on the resulting marker file gives us a
- * positive "Claude is alive" signal instead of racing the transcript writer.
- *
- * The timeout only fires on catastrophic startup failure (bad binary, exec
- * error) — under load, Claude's own session bootstrap runs well under the
- * limit because SessionStart is dispatched early in the startup sequence.
- */
-async function waitForReadyMarker(sessionId: string): Promise<void> {
-  const { ready: readyDir } = claudeHookDirs();
-  await mkdir(readyDir, { recursive: true });
-  const target = join(readyDir, sessionId);
-
-  if (existsSync(target)) return;
-
-  const ac = new AbortController();
-  const timeout = setTimeout(() => ac.abort(), READY_HOOK_TIMEOUT_MS);
-
-  try {
-    await Promise.race([
-      // fs.watch — instant OS-native notification when the hook writes the file
-      (async (): Promise<void> => {
-        try {
-          for await (const _event of watch(readyDir, { signal: ac.signal })) {
-            // Trust disk state, not event.filename (Linux can deliver
-            // unexpected basenames under tmp+rename writes).
-            if (existsSync(target)) return;
-          }
-        } catch (e: unknown) {
-          if (e instanceof Error && e.name === "AbortError") throw e;
-        }
-        return new Promise<void>(() => {});
-      })(),
-
-      // Polling fallback — catches dropped inotify/FSEvent notifications
-      (async (): Promise<void> => {
-        while (!ac.signal.aborted) {
-          if (existsSync(target)) return;
-          await Bun.sleep(250);
-        }
-        throw new DOMException("Aborted", "AbortError");
-      })(),
-    ]);
-  } catch (e: unknown) {
-    if (e instanceof DOMException && e.name === "AbortError") {
-      throw new Error(
-        `Timed out waiting for Claude SessionStart hook to signal readiness ` +
-        `at ${target}. Verify the \`claude\` command started successfully.`,
-      );
-    }
-    throw e;
-  } finally {
-    clearTimeout(timeout);
-    ac.abort();
-  }
-}
-
-// ---------------------------------------------------------------------------
-// HIL detection helpers
-// ---------------------------------------------------------------------------
-
-/**
- * Returns true when the most recent assistant message in the transcript
- * ended with `stop_reason: "tool_use"` — i.e. the agent stopped the current
- * API response to call a tool but has not yet produced its post-tool answer.
- *
- * Claude Code's Stop hook fires each time Claude "finishes responding",
- * which includes intermediate tool-use responses in a multi-step agent
- * loop (not just the final `end_turn`). If we return from `waitForIdle`
- * on the first Stop event, we capture the transcript mid-loop — the
- * final assistant text block is still being generated and won't be on
- * disk yet, so `inbox.md` drops the actual answer.
- *
- * We keep watching until we see an assistant message with a terminal
- * stop_reason (`end_turn`, `max_tokens`, `stop_sequence`, `refusal`),
- * which is the real end of the turn.
- *
- * Exported as `_isMidAgentLoop` for unit testing.
- */
-export function _isMidAgentLoop(messages: SessionMessage[]): boolean {
-  for (let i = messages.length - 1; i >= 0; i--) {
-    const msg = messages[i];
-    if (msg?.type !== "assistant") continue;
-    const inner = msg.message as { stop_reason?: unknown } | null;
-    const stopReason = inner?.stop_reason;
-    return stopReason === "tool_use";
-  }
-  // No assistant message yet — treat as mid-loop so we wait for one.
-  return true;
-}
-
-/**
- * Watch `~/.atomic/claude-hil/` for this session's marker file and fire
- * `onHIL(true|false)` on create/unlink. Returns when `signal` is aborted.
- *
- * The marker is written by the `_claude-ask-hook enter` subcommand from
- * Claude Code's `PreToolUse` hook (matched on `AskUserQuestion`) and removed
- * by `_claude-ask-hook exit` from `PostToolUse` / `PostToolUseFailure`. That
- * makes the signal deterministic and independent of Claude Code's batched
- * JSONL flush timing, which used to hide the HIL window entirely when
- * tool_use and tool_result landed in the same file write.
- *
- * @internal Exported for tests.
- */
-export async function watchHILMarker(
-  claudeSessionId: string,
-  onHIL: (waiting: boolean) => void,
-  signal: AbortSignal,
-): Promise<void> {
-  const { hil: dir } = claudeHookDirs();
-  const target = join(dir, claudeSessionId);
-
-  await mkdir(dir, { recursive: true });
-
-  let wasHIL = false;
-  const emit = (isHIL: boolean): void => {
-    if (isHIL !== wasHIL) {
-      onHIL(isHIL);
-      wasHIL = isHIL;
-    }
-  };
-
-  // Attach the watcher BEFORE the initial existsSync so any event that fires
-  // during the check is buffered by the iterator instead of being dropped.
-  const watcher = watch(dir, { signal });
-
-  // Polling fallback: Bun/inotify can drop events under heavy fs load, which
-  // would leave the UI stuck on (or off) the blue "awaiting_input" pulse.
-  // A cheap periodic existsSync guarantees eventual consistency. `emit` is
-  // guarded by `wasHIL` so the interval is idempotent w.r.t. the watcher.
-  const poll = setInterval(() => emit(existsSync(target)), 250);
-
-  // Initial existsSync: handles resumed sessions whose PreToolUse marker was
-  // already on disk before the watcher attached.
-  if (existsSync(target)) emit(true);
-
-  try {
-    for await (const _event of watcher) {
-      // Don't trust event.filename — Bun/Linux deliver inconsistent basenames
-      // across OSes and write patterns. Disk existence is authoritative.
-      emit(existsSync(target));
-    }
-  } catch (e: unknown) {
-    if (!(e instanceof Error && e.name === "AbortError")) {
-      throw e;
-    }
-  } finally {
-    clearInterval(poll);
-  }
-}
-
-// ---------------------------------------------------------------------------
-// Helpers
-// ---------------------------------------------------------------------------
-
-/**
- * Path of the directory where the claude-stop-hook writes marker files.
- * Each Claude turn creates `~/.atomic/claude-stop/<session_id>` which
- * triggers the `fs.watch` event in `waitForIdle`.
- *
- * @internal Exported for unit tests.
- */
-export function markerDir(): string {
-  return claudeHookDirs().marker;
-}
-
-/**
- * Return the marker file path for a given Claude session ID.
- *
- * @internal Exported for unit tests.
- */
-export function markerPath(claudeSessionId: string): string {
-  return join(markerDir(), claudeSessionId);
-}
-
-/**
- * Directory where the workflow runtime writes queued follow-up prompts that
- * `atomic _claude-stop-hook` picks up and feeds back to Claude as
- * `{decision:"block", reason:<prompt>}`. @internal Exported for unit tests.
- */
-export function queueDir(): string {
-  return claudeHookDirs().queue;
-}
-
-/** Return the queue file path for a given Claude session ID. @internal */
-export function queuePath(claudeSessionId: string): string {
-  return join(queueDir(), claudeSessionId);
-}
-
-/**
- * Directory where the runtime writes session-release signals. When the Stop
- * hook sees `~/.atomic/claude-release/<session_id>` it exits 0 without
- * emitting a block decision — the signal used by `clearClaudeSession` to
- * tell Claude it's safe to actually stop. @internal Exported for unit tests.
- */
-export function releaseDir(): string {
-  return claudeHookDirs().release;
-}
-
-/** Return the release file path for a given Claude session ID. @internal */
-export function releasePath(claudeSessionId: string): string {
-  return join(releaseDir(), claudeSessionId);
-}
-
-/**
- * Ensure the marker directory exists and remove any stale marker left from a
- * previous turn of this session. Call this BEFORE submitting the prompt so
- * the subsequent `waitForIdle` watch loop doesn't fire on a stale file.
- *
- * Ignores ENOENT on `unlink` — the file simply doesn't exist yet.
- */
-async function clearStaleMarker(claudeSessionId: string): Promise<void> {
-  await mkdir(markerDir(), { recursive: true });
-  try {
-    await unlink(markerPath(claudeSessionId));
-  } catch (e: unknown) {
-    // ENOENT is expected — ignore it; rethrow anything else
-    if (!(e instanceof Error && "code" in e && (e as NodeJS.ErrnoException).code === "ENOENT")) {
-      throw e;
-    }
-  }
-}
-
-/**
- * Ensure the queue directory exists and remove any stale entry from a prior
- * turn so the Stop hook doesn't race on it. Ignores ENOENT.
- */
-async function clearStaleQueue(claudeSessionId: string): Promise<void> {
-  await mkdir(queueDir(), { recursive: true });
-  try {
-    await unlink(queuePath(claudeSessionId));
-  } catch (e: unknown) {
-    if (!(e instanceof Error && "code" in e && (e as NodeJS.ErrnoException).code === "ENOENT")) {
-      throw e;
-    }
-  }
-}
-
-/**
- * Remove a stale HIL marker left over from a prior turn (e.g. the ask-hook
- * process was SIGKILL'd between PreToolUse and PostToolUse). Without this,
- * `watchHILMarker`'s initial `existsSync` would spuriously fire `onHIL(true)`
- * at the start of a fresh turn. Ignores ENOENT.
- */
-async function clearStaleHILMarker(claudeSessionId: string): Promise<void> {
-  const { hil } = claudeHookDirs();
-  await mkdir(hil, { recursive: true });
-  try {
-    await unlink(join(hil, claudeSessionId));
-  } catch (e: unknown) {
-    if (!(e instanceof Error && "code" in e && (e as NodeJS.ErrnoException).code === "ENOENT")) {
-      throw e;
-    }
-  }
-}
-
-/**
- * Remove a stale ready marker from a prior session that reused this UUID (in
- * practice impossible — UUIDs are fresh per session — but cheap insurance so
- * `waitForReadyMarker`'s initial existsSync can't false-positive on anything
- * we left behind). Ignores ENOENT.
- */
-async function clearStaleReadyMarker(claudeSessionId: string): Promise<void> {
-  const { ready } = claudeHookDirs();
-  await mkdir(ready, { recursive: true });
-  try {
-    await unlink(join(ready, claudeSessionId));
-  } catch (e: unknown) {
-    if (!(e instanceof Error && "code" in e && (e as NodeJS.ErrnoException).code === "ENOENT")) {
-      throw e;
-    }
-  }
-}
-
-/**
- * Write the next prompt to the session queue file. The currently-running
- * Stop hook process (blocked on poll from the previous turn) picks it up,
- * emits `{decision:"block", reason:<prompt>}` on stdout, and Claude feeds
- * it back as the next user message — no tmux keystrokes required.
- */
-async function enqueuePrompt(claudeSessionId: string, prompt: string): Promise<void> {
-  await mkdir(queueDir(), { recursive: true });
-  await writeFile(queuePath(claudeSessionId), prompt, "utf-8");
-}
-
-/**
- * Signal the Stop hook's blocking wait that this session is done. Called
- * during session teardown so the final hook invocation exits 0 promptly.
- * Safe to call more than once.
- */
-export async function releaseClaudeSession(claudeSessionId: string): Promise<void> {
-  await mkdir(releaseDir(), { recursive: true });
-  await writeFile(releasePath(claudeSessionId), "");
-}
-
-/** @internal */
-function pidDir(): string {
-  return claudeHookDirs().pid;
-}
-
-/** @internal */
-function pidFilePath(claudeSessionId: string): string {
-  return join(pidDir(), claudeSessionId);
-}
-
-/**
- * Write `process.pid` to `~/.atomic/claude-pid/<session_id>` so the Stop hook
- * can use it as a liveness signal. If atomic is SIGKILL'd (no chance to run
- * `clearClaudeSession`), the hook detects the dead PID via `process.kill(..,0)`
- * and self-exits instead of parking Claude for the full 24-day timeout.
- */
-async function writeAtomicPidFile(claudeSessionId: string): Promise<void> {
-  await mkdir(pidDir(), { recursive: true });
-  await writeFile(pidFilePath(claudeSessionId), String(process.pid), "utf-8");
-}
-
-/** Remove the pid file for a session. Idempotent — ENOENT is swallowed. */
-async function unlinkAtomicPidFile(claudeSessionId: string): Promise<void> {
-  try {
-    await unlink(pidFilePath(claudeSessionId));
-  } catch (e: unknown) {
-    if (!(e instanceof Error && "code" in e && (e as NodeJS.ErrnoException).code === "ENOENT")) {
-      throw e;
-    }
-  }
-}
-
-// ---------------------------------------------------------------------------
-// Idle detection via marker file watch
-// ---------------------------------------------------------------------------
-
-/**
- * Wait for the Claude session to become idle using `fs.watch` on the
- * `~/.atomic/claude-stop/` marker directory.
- *
- * When Claude finishes a turn, the `atomic _claude-stop-hook` Stop hook writes
- * `~/.atomic/claude-stop/<session_id>`. The write triggers an OS-native
- * `fs.watch` event on the parent directory — far more reliable than polling
- * tmux pane glyphs, which vary between Claude Code versions.
- *
- * This function is strictly about *idle detection*. HIL is detected separately
- * by {@link watchHILMarker}; the Stop hook does not fire while
- * `AskUserQuestion` is pending (the agent loop blocks on deferred tools), so
- * mixing the two would silently miss the HIL window.
- *
- * Algorithm:
- * 1. Attach the directory watcher, then check for the marker file on disk —
- *    this closes the race where the Stop hook fires between prompt submission
- *    and watcher attach.
- * 2. On any event, re-check the marker file on disk (we intentionally do NOT
- *    filter by `event.filename`, because on Linux a write can deliver multiple
- *    events with varying filenames and editor tools may race us).
- * 3. Read the session transcript via `getSessionMessages` and slice messages
- *    from `transcriptBeforeCount`.
- * 4. Clean up the `fs.watch` watcher on any exit path via AbortController.
- *
- * @param claudeSessionId       - Claude's session UUID (used to identify marker file)
- * @param transcriptBeforeCount - number of messages in transcript before this turn
- */
-/**
- * @internal Exported for unit tests.
- */
-export async function waitForIdle(
-  claudeSessionId: string,
-  transcriptBeforeCount: number,
-): Promise<SessionMessage[]> {
-
-  const dir = markerDir();
-  const sessionId = claudeSessionId;
-  const target = markerPath(sessionId);
-  const ac = new AbortController();
-
-  // Process a marker that has appeared on disk. Returns a tuple:
-  //   [resolved, result] — when resolved=true, waitForIdle should return.
-  const readMessages = async (): Promise<SessionMessage[] | null> => {
-    try {
-      return await getSessionMessages(sessionId, {
-        dir: process.cwd(),
-        includeSystemMessages: true,
-      });
-    } catch {
-      return null;
-    }
-  };
-
-  const handleMarker = async (): Promise<[boolean, SessionMessage[]]> => {
-    let msgs = await readMessages();
-    if (msgs === null) {
-      // Transcript read failed — keep watching; the next event will retry.
-      return [false, []];
-    }
-
-    // The Stop hook fires only once per agent loop completion (when there
-    // are no more tool_use blocks to resolve — see Claude Code's
-    // `src/query/stopHooks.ts` / `query.ts`: `if (!needsFollowUp)`). But
-    // Claude Code writes to the JSONL transcript asynchronously via
-    // `enqueueWrite()` with a batched ~100ms flush, so the final
-    // `assistant[text]` message can still be in the page-cache when our
-    // marker watcher fires. Reading the transcript at that moment races
-    // the writer and returns a prefix ending at `user[tool_result]`.
-    //
-    // Because no further marker events are coming, we can't just "keep
-    // watching the marker dir". Instead, poll the transcript file directly
-    // until it either settles on a terminal stop_reason or the poll budget
-    // expires. The budget covers Claude Code's flush interval plus headroom
-    // for slow disks and buffered `fs/promises` writes.
-    if (_isMidAgentLoop(msgs)) {
-      const pollIntervalMs = 50;
-      const pollBudgetMs = 3_000;
-      const start = Date.now();
-      while (_isMidAgentLoop(msgs) && Date.now() - start < pollBudgetMs) {
-        await Bun.sleep(pollIntervalMs);
-        const next = await readMessages();
-        if (next) msgs = next;
-      }
-      // Whether we recovered or ran out of budget, fall through — returning
-      // what we have beats hanging forever if the writer really did drop a
-      // message (e.g. max-tokens collapse, abort mid-stream).
-    }
-
-    const sliced = msgs.length > transcriptBeforeCount
-      ? msgs.slice(transcriptBeforeCount)
-      : [];
-    return [true, sliced];
-  };
-
-  try {
-    // Attach the watcher FIRST; fs.watch returns an iterable whose underlying
-    // inotify/FSEvent subscription is live from this point on.
-    const watcher = watch(dir, { signal: ac.signal });
-
-    // Close the race: if the Stop hook fired between clearStaleMarker() and
-    // the watcher attach above, the marker is already on disk and no further
-    // events will be emitted. Handle it synchronously.
-    if (existsSync(target)) {
-      const [done, result] = await handleMarker();
-      if (done) {
-        ac.abort();
-        return result;
-      }
-    }
-
-    for await (const _event of watcher) {
-      // We don't trust event.filename — on Linux, a tmp+rename write emits
-      // events with the `.tmp` basename, and other files in the marker dir
-      // can race us. The marker file's existence on disk is authoritative.
-      if (!existsSync(target)) continue;
-
-      const [done, result] = await handleMarker();
-      if (done) {
-        ac.abort();
-        return result;
-      }
-    }
-  } catch (e: unknown) {
-    // AbortError is expected when we call ac.abort() to stop watching.
-    if (!(e instanceof Error && e.name === "AbortError")) {
-      throw e;
-    }
-  }
-
-  return [];
-}
-
-// ---------------------------------------------------------------------------
-// claudeQuery
-// ---------------------------------------------------------------------------
-
-export interface ClaudeQueryOptions {
-  /** tmux pane ID where Claude is running */
-  paneId: string;
-  /** The prompt to send */
-  prompt: string;
-  /**
-   * Called when the agent's human-in-the-loop state changes.
-   * `waiting=true`  → AskUserQuestion is pending (agent blocked on user input).
-   * `waiting=false` → AskUserQuestion was resolved (agent resumed processing).
-   */
-  onHIL?: (waiting: boolean) => void;
-}
-
-/**
- * Extract text content from assistant messages in a transcript slice.
- *
- * Walks messages from `afterIndex` forward, pulls `TextBlock.text` from each
- * assistant message's content array, and joins them. The `message` payload is
- * `unknown` in the SDK type so we do runtime narrowing.
- *
- * Exported so workflow authors can extract text from `SessionMessage[]`
- * returned by `s.session.query()`.
- */
-export function extractAssistantText(
-  msgs: ReadonlyArray<{ type: string; message: unknown }>,
-  afterIndex: number,
-): string {
-  const parts: string[] = [];
-  for (let i = afterIndex; i < msgs.length; i++) {
-    const msg = msgs[i];
-    if (!msg || msg.type !== "assistant") continue;
-    const m = msg.message;
-    if (!m || typeof m !== "object") continue;
-    const content = (m as Record<string, unknown>).content;
-    if (!Array.isArray(content)) continue;
-    for (const block of content) {
-      if (
-        block &&
-        typeof block === "object" &&
-        (block as Record<string, unknown>).type === "text"
-      ) {
-        parts.push(String((block as Record<string, unknown>).text ?? ""));
-      }
-    }
-  }
-  return parts.join("\n");
-}
-
-/**
- * Send a prompt to a Claude Code interactive session running in a tmux pane.
- *
- * First query and follow-up queries use different delivery channels:
- *
- *   - **First query**: stages the prompt in a tmp file and spawns
- *     `claude --session-id <UUID> 'Read the prompt in <path>'` into the
- *     empty pane. Claude's first action is a Read tool call, which
- *     sidesteps ARG_MAX on the spawn argv.
- *
- *   - **Follow-up query**: writes the prompt to
- *     `~/.atomic/claude-queue/<session_id>`. The Stop hook from the
- *     previous turn is blocked in a poll loop there; it reads the queue
- *     entry and emits `{"decision":"block","reason":<prompt>}` on stdout,
- *     which Claude Code feeds back as the next user message. No tmux
- *     keystrokes, no paste-buffer dance, no pane-state polling — the
- *     whole delivery rides Claude's own continuation API.
- *
- * Both paths converge on `waitForIdle`, which watches the Stop-hook marker
- * file for this session and returns the transcript slice for the turn.
- *
- * @example
- * ```typescript
- * import { claudeQuery } from "@bastani/atomic/workflows";
- *
- * const result = await claudeQuery({
- *   paneId: ctx.paneId,
- *   prompt: "Describe this project",
- * });
- * ctx.log(result.output);
- * ```
- */
-export async function claudeQuery(options: ClaudeQueryOptions): Promise<SessionMessage[]> {
-  const { paneId, prompt, onHIL } = options;
-
-  const paneState = initializedPanes.get(paneId);
-  if (!paneState) {
-    throw new Error(
-      "claudeQuery() called without a prior createClaudeSession() for this pane. " +
-      "Call createClaudeSession({ paneId }) first to start the Claude CLI.",
-    );
-  }
-
-  const dir = process.cwd();
-  const claudeSessionId = paneState.claudeSessionId;
-
-  // Clear stale marker AND stale queue entry before submitting so the
-  // Stop-hook for the previous turn (if any) cannot race this one. The HIL
-  // marker is cleared too so a crashed ask-hook process from turn N-1 can't
-  // make `watchHILMarker`'s initial existsSync spuriously fire onHIL(true).
-  await clearStaleMarker(claudeSessionId);
-  await clearStaleQueue(claudeSessionId);
-  await clearStaleHILMarker(claudeSessionId);
-  await clearStaleReadyMarker(claudeSessionId);
-
-  let transcriptBeforeCount = 0;
-  let spawnPromptFile: string | undefined;
-
-  try {
-    if (paneState.claudeStarted) {
-      // Follow-up query: snapshot the transcript length so waitForIdle can
-      // slice out the messages produced by THIS turn, then enqueue the
-      // prompt for the Stop hook to pick up.
-      try {
-        const msgs = await getSessionMessages(claudeSessionId, {
-          dir,
-          includeSystemMessages: true,
-        });
-        transcriptBeforeCount = msgs.length;
-      } catch {
-        // Best-effort — 0 means we scan all messages (correct, slightly less efficient)
-      }
-
-      await enqueuePrompt(claudeSessionId, prompt);
-    } else {
-      // First query: spawn claude with the prompt baked into argv via the
-      // Read-tool indirection. The tmp file only has to live long enough
-      // for Claude's first Read tool call, so we delete it once waitForIdle
-      // returns (the turn is complete by then).
-      spawnPromptFile = atomicTempPath(
-        "atomic-claude-prompt",
-        ".txt",
-        `${claudeSessionId}-${randomUUID()}`,
-      );
-      writeFileSync(spawnPromptFile, prompt, {
-        encoding: "utf-8",
-        mode: 0o600,
-      });
-
-      await spawnClaudeWithPrompt(
-        paneId,
-        spawnPromptFile,
-        paneState.chatFlags,
-        claudeSessionId,
-      );
-      paneState.claudeStarted = true;
-    }
-
-    // HIL detection runs in parallel with idle detection. The
-    // PreToolUse/PostToolUse/PostToolUseFailure hooks on `AskUserQuestion`
-    // write/remove `~/.atomic/claude-hil/<session_id>`; we watch that dir
-    // for create/unlink events so HIL state is deterministic and immune to
-    // Claude Code's batched JSONL flush timing.
-    const hilAc = new AbortController();
-    if (onHIL) {
-      void watchHILMarker(claudeSessionId, onHIL, hilAc.signal).catch(() => {
-        // Best-effort — never fail the query over HIL detection.
-      });
-    }
-
-    try {
-      return await waitForIdle(claudeSessionId, transcriptBeforeCount);
-    } finally {
-      hilAc.abort();
-      // Safety: waitForIdle only returns at true turn-idle. If the ask-hook
-      // process crashed mid-turn and left the marker on disk, the UI could
-      // be stuck on awaiting_input. `resumeSession` in the panel store is
-      // idempotent (no-op when the session isn't in awaiting_input), so
-      // this is always safe.
-      onHIL?.(false);
-    }
-  } finally {
-    if (spawnPromptFile) {
-      try {
-        await unlink(spawnPromptFile);
-      } catch {
-        // ENOENT / already removed is fine.
-      }
-    }
-  }
-}
-
-// ---------------------------------------------------------------------------
-// Synthetic wrappers — uniform s.client / s.session API for Claude stages
-// ---------------------------------------------------------------------------
-
-/**
- * Merge two `disallowedTools` lists, preserving caller entries and appending
- * any extras that aren't already present. Exported for unit testing.
- */
-export function mergeDisallowedTools(
-  existing: string[] | undefined,
-  extras: string[],
-): string[] {
-  const merged = [...(existing ?? [])];
-  for (const tool of extras) {
-    if (!merged.includes(tool)) merged.push(tool);
-  }
-  return merged;
-}
-
-/**
- * Fold the atomic-managed additional instructions into a caller's
- * `systemPrompt` value. Behavior, in order of precedence:
- *
- *   - **No caller value** → return a `claude_code` preset with our content
- *     in `append`. Preserves the SDK's full Claude Code persona.
- *   - **Caller passed a preset object** → concatenate our content onto the
- *     existing `append` (newline-separated when both are present).
- *   - **Caller passed a custom string or array** → leave it alone. The
- *     caller has explicitly opted into a custom prompt, and silently
- *     prepending the persona-style preset text would break that contract.
- *
- * Exported for unit testing.
- */
-export function mergeSystemPromptAppend(
-  existing: SDKOptions["systemPrompt"],
-  extra: string,
-): SDKOptions["systemPrompt"] {
-  if (!extra) return existing;
-  if (existing === undefined) {
-    return { type: "preset", preset: "claude_code", append: extra };
-  }
-  if (typeof existing === "object" && !Array.isArray(existing) && existing.type === "preset") {
-    const prevAppend = existing.append ?? "";
-    const merged = prevAppend ? `${prevAppend}\n\n${extra}` : extra;
-    return { ...existing, append: merged };
-  }
-  return existing;
-}
-
-/**
- * Synthetic client wrapper for Claude stages.
- * Auto-starts the Claude CLI in the tmux pane during `start()`.
- */
-export class ClaudeClientWrapper {
-  readonly paneId: string;
-  private readonly opts: { chatFlags?: string[] };
-
-  constructor(
-    paneId: string,
-    opts: { chatFlags?: string[] } = {},
-  ) {
-    this.paneId = paneId;
-    this.opts = opts;
-  }
-
-  /**
-   * Start the Claude CLI in the tmux pane. Returns the Claude session UUID
-   * so the caller can pass it to `ClaudeSessionWrapper` (and thus expose it
-   * as `s.sessionId` to workflows). This is the UUID used by Claude Code to
-   * name its JSONL transcript file and to key the Stop-hook marker — workflows
-   * pass it to `s.save(s.sessionId)` so the save path reads the correct
-   * transcript even when many Claude sessions run in parallel.
-   */
-  async start(): Promise<string> {
-    return await createClaudeSession({
-      paneId: this.paneId,
-      chatFlags: this.opts.chatFlags,
-    });
-  }
-
-  /** Noop — cleanup is handled by the runtime via `clearClaudeSession`. */
-  async stop(): Promise<void> {}
-}
-
-/**
- * Synthetic session wrapper for Claude stages.
- * Wraps `claudeQuery()` so users call `s.session.query(prompt)`.
- */
-export class ClaudeSessionWrapper {
-  readonly paneId: string;
-  readonly sessionId: string;
-  private readonly onHIL: ((waiting: boolean) => void) | undefined;
-
-  constructor(
-    paneId: string,
-    sessionId: string,
-    onHIL?: (waiting: boolean) => void,
-  ) {
-    this.paneId = paneId;
-    this.sessionId = sessionId;
-    this.onHIL = onHIL;
-  }
-
-  /**
-   * Send a prompt to Claude and wait for the response.
-   *
-   * The `_options` parameter exists for signature compatibility with
-   * {@link HeadlessClaudeSessionWrapper#query} (which forwards SDK options
-   * like `agent`, `permissionMode`, etc. to the Agent SDK). In the
-   * interactive pane path these options don't apply — we're driving the
-   * `claude` CLI binary, not the SDK — so they are silently ignored.
-   */
-  async query(
-    prompt: string,
-    _options?: Partial<SDKOptions>,
-  ): Promise<SessionMessage[]> {
-    return claudeQuery({
-      paneId: this.paneId,
-      prompt,
-      onHIL: this.onHIL,
-    });
-  }
-
-  /**
-   * Structured output is only produced by the Agent SDK's `result` message,
-   * which interactive stages don't consume (they drive the `claude` CLI via
-   * tmux, not the SDK). Always `undefined` here — pair `outputFormat` with a
-   * headless stage to read {@link HeadlessClaudeSessionWrapper#lastStructuredOutput}.
-   */
-  get lastStructuredOutput(): unknown {
-    return undefined;
-  }
-
-  /** Noop — for API symmetry with CopilotSession.disconnect(). */
-  async disconnect(): Promise<void> {}
-}
-
-// ---------------------------------------------------------------------------
-// Headless wrappers — use the Agent SDK directly (no tmux pane)
-// ---------------------------------------------------------------------------
-
-/**
- * Headless client wrapper for Claude stages. No tmux pane — noop start/stop.
- * Used when `options.headless` is true in `ctx.stage()`.
- */
-export class HeadlessClaudeClientWrapper {
-  /**
-   * Headless Claude stages don't pre-allocate a session — each `query()` call
-   * to {@link HeadlessClaudeSessionWrapper} spawns a fresh Agent SDK run that
-   * emits its own `session_id`. We still return an empty string here so the
-   * method signature matches {@link ClaudeClientWrapper#start}.
-   */
-  async start(): Promise<string> {
-    return "";
-  }
-  async stop(): Promise<void> {}
-}
-
-/**
- * Resolve the `claude` CLI binary for headless SDK queries.
- *
- * Pins the SDK to the same binary interactive stages already spawn via tmux
- * (`AGENT_CONFIG.claude.cmd` on PATH), bypassing
- * `@anthropic-ai/claude-agent-sdk`'s built-in resolver. That resolver probes
- * optional native packages in a fixed order — on Linux it tries
- * `linux-${arch}-musl` before `linux-${arch}` and returns whichever
- * `require.resolve` finds first — so on a glibc host where both optional
- * packages got installed (Bun installs every optionalDependency by default)
- * it picks the musl binary, which can't exec because its dynamic linker
- * (`/lib/ld-musl-*.so.1`) is absent. The SDK surfaces the resulting ENOENT
- * as a misleading "Claude Code native binary not found" error.
- *
- * `chatCommand` and `workflowCommand` already fail fast when `claude` isn't
- * on PATH (see `isCommandInstalled` in each), so in practice this lookup
- * always succeeds. The throw here is a belt-and-suspenders guard that
- * prefers a clear failure over silently falling back to the SDK's resolver.
- */
-export function resolveHeadlessClaudeBin(): string {
-  // Pass PATH explicitly — the 1-arg form of Bun.which caches the value
-  // captured at process start, which makes the lookup insensitive to later
-  // env mutations (and un-exercisable from tests that tweak `process.env.PATH`).
-  const onPath = Bun.which("claude", { PATH: process.env.PATH ?? "" });
-  if (!onPath) {
-    throw new Error(
-      "`claude` CLI not found on PATH. Install Claude Code via the native " +
-        "installer (https://docs.claude.com/en/docs/claude-code/overview) " +
-        "and retry.",
-    );
-  }
-  return onPath;
-}
-
-/**
- * Headless session wrapper for Claude stages. Uses the Agent SDK's `query()`
- * directly instead of tmux pane operations. Implements the same `query()`
- * interface as {@link ClaudeSessionWrapper} so workflow callbacks work
- * identically for headless and interactive stages.
- *
- * The `query()` method accepts the full Agent SDK parameter types —
- * `prompt` can be a plain string or an `AsyncIterable<SDKUserMessage>`
- * for multi-turn streaming, and `options` passes through SDK-level
- * configuration (abort controllers, allowed tools, agents, etc.).
- */
-export class HeadlessClaudeSessionWrapper {
-  readonly paneId = "";
-  /**
-   * Project root the workflow is operating against. Used to resolve
-   * project-scoped config (e.g. `additional-instructions`) against the
-   * workflow's actual root rather than `process.cwd()`, which can drift
-   * when workflows are invoked programmatically or from a subdirectory.
-   */
-  private readonly _projectRoot: string;
-  /**
-   * The Claude session UUID of the most recently completed `query()`. Exposed
-   * via `s.sessionId` so workflows can pass it to `s.save(s.sessionId)` and
-   * have the save path read the correct transcript, even when several headless
-   * Claude stages run in parallel (each call gets its own SDK-assigned UUID).
-   */
-  private _lastSessionId: string = "";
-
-  constructor(projectRoot: string) {
-    this._projectRoot = projectRoot;
-  }
-  /**
-   * Validated structured output captured from the most recent `query()`'s
-   * `result` message. Populated only when callers pass
-   * `options.outputFormat = { type: "json_schema", schema }` and the SDK
-   * produced a `subtype: "success"` result with `structured_output` attached.
-   * Remains `undefined` on plain text runs or when the SDK fails validation
-   * (`error_max_structured_output_retries`).
-   */
-  private _lastStructuredOutput: unknown = undefined;
-
-  get sessionId(): string {
-    return this._lastSessionId;
-  }
-
-  get lastStructuredOutput(): unknown {
-    return this._lastStructuredOutput;
-  }
-
-  async query(
-    prompt: string | AsyncIterable<SDKUserMessage>,
-    options?: Partial<SDKOptions>,
-  ): Promise<SessionMessage[]> {
-    // Auto-deny the `AskUserQuestion` tool in headless runs. Without this, the
-    // agent can call it and the SDK query will sit blocked forever since no
-    // human is attached to answer.
-    const sdkOpts = options ?? {};
-    const additional = await resolveAdditionalInstructionsContent(this._projectRoot);
-    const headlessSdkOpts: Partial<SDKOptions> = {
-      ...sdkOpts,
-      pathToClaudeCodeExecutable:
-        sdkOpts.pathToClaudeCodeExecutable ?? resolveHeadlessClaudeBin(),
-      disallowedTools: mergeDisallowedTools(sdkOpts.disallowedTools, [
-        "AskUserQuestion",
-      ]),
-      ...(additional
-        ? { systemPrompt: mergeSystemPromptAppend(sdkOpts.systemPrompt, additional) }
-        : {}),
-    };
-
-    let sdkSessionId = "";
-    let structuredOutput: unknown = undefined;
-    try {
-      await withAtomicTempEnv(async () => {
-        for await (const msg of sdkQuery({ prompt, options: headlessSdkOpts })) {
-          if (msg.type === "result") {
-            const record = msg as Record<string, unknown>;
-            sdkSessionId = String(record.session_id ?? "");
-            if (record.subtype === "success" && "structured_output" in record) {
-              structuredOutput = record.structured_output;
-            }
-          }
-        }
-      });
-    } catch (err) {
-      const detail = err instanceof Error ? err.message : String(err);
-      throw new Error(`Claude SDK query failed: ${detail}`);
-    }
-    if (!sdkSessionId) {
-      throw new Error(
-        "Claude SDK query completed without a `result` message — " +
-          "likely a stream idle timeout, aborted request, or upstream API error. " +
-          "Set CLAUDE_ENABLE_STREAM_WATCHDOG=1 (and tune CLAUDE_STREAM_IDLE_TIMEOUT_MS / " +
-          "API_TIMEOUT_MS) so the CLI surfaces a concrete failure instead of exiting silently.",
-      );
-    }
-    this._lastSessionId = sdkSessionId;
-    this._lastStructuredOutput = structuredOutput;
-    return getSessionMessages(sdkSessionId, { dir: process.cwd() });
-  }
-
-  async disconnect(): Promise<void> {}
-}
-
-// ---------------------------------------------------------------------------
-// Static source validation
-// ---------------------------------------------------------------------------
-
-import { createProviderValidator } from "../types.ts";
-
-/**
- * Validate a Claude workflow source file for common mistakes.
- *
- * Warns on direct usage of createClaudeSession/claudeQuery — the runtime
- * now handles init/cleanup automatically via s.client and s.session.
- */
-export const validateClaudeWorkflow = createProviderValidator([
-  {
-    pattern: /\bcreateClaudeSession\b/,
-    rule: "claude/manual-session",
-    message:
-      "Manual createClaudeSession() call detected. The runtime auto-starts the Claude CLI — " +
-      "use s.session.query() instead of claudeQuery(). Pass chatFlags via the second arg to ctx.stage().",
-  },
-  {
-    pattern: /\bclaudeQuery\b/,
-    rule: "claude/manual-query",
-    message:
-      "Direct claudeQuery() call detected. Use s.session.query(prompt) instead — " +
-      "it wraps claudeQuery with the correct paneId.",
-  },
-]);