AgentHub/server/lib/pty-run.js

/**
 * @file pty-run.js
 * @description Owns the tmux-backed run lifecycle for the dashboard's
 * terminal-run feature: Start (create-or-adopt), Resume (`--resume`), Kill,
 * and List. Unlike the old run-spawner.js, there is no in-memory handle Map —
 * `listRuns`/`getRun` are computed fresh from `tmux list-sessions` on every
 * call, the same "computed fact, never a stored one" principle this repo
 * already applies to lane runtime up/down (see CLAUDE.md). A session killed
 * out-of-band (crash, manual `tmux kill-session`, host reboot) self-corrects
 * on the next read instead of leaving a ghost "running" row.
 *
 * Start/Resume is create-or-reuse: when the lane's tmux session already
 * exists but its pane sits at a shell prompt, the argv is typed into that
 * pane instead of being dropped on the floor by a silent adopt.
 *
 * Every session is named `agenthub-lane-<laneId>` so a real terminal can attach
 * to the exact same session (`tmux attach -t agenthub-lane-<id>`, or
 * `agenthub lanes shell`) — that's the whole point: the dashboard both creates
 * the session (one-click Start/Resume) and is just one of possibly several
 * attached clients tmux already keeps in sync.
 *
 * Spawns `claude`, `codex` or `agy` (Antigravity) in the tmux pane based on
 * the tool parameter.
 *
 * A lane can hold up to four sessions, one per Workspace split pane: pane 0 is
 * `agenthub-lane-<laneId>` (the lane's own `run_id`), panes 1–3 are
 * `agenthub-lane-<laneId>-<slot>`.
 *
 * @author Nguyễn Ngọc Trí Vĩ <vinnt@smartgift.vn>
 */

const tmux = require("./tmux");
const { resolveRouter, routerForSpawn } = require("./router-apply");
const { desktopEnv } = require("./desktop-env");
const { getCodexHome } = require("./codex-home");
const aiTools = require("./ai-tools");

let dashboardRuns = null;
try {
  dashboardRuns = require("./dashboard-runs");
} catch {
  /* db-less environment, skip persistence */
}

const RUN_ID_RE = /^agenthub-lane-(\d+)(?:-([1-3]))?$/;
/** Highest pane slot a lane may use (the Workspace's 4-pane grid is 0–3). */
const MAX_PANE_SLOT = 3;
const EFFORT_LEVELS = new Set(["low", "medium", "high", "xhigh", "max"]);
// `auto` is the CLI's own default-ish mode (`claude --permission-mode auto`):
// it asks only for what its classifier reads as risky. `default` is kept for
// callers that stored it, though this CLI now spells that one `manual`.
const ALLOWED_PERMISSION_MODES = new Set([
  "auto",
  "acceptEdits",
  "default",
  "plan",
  "bypassPermissions",
]);
// Pane commands that mean "idle shell prompt, safe to type a command into".
const SHELL_COMMANDS = new Set(["sh", "bash", "zsh", "fish", "dash", "ksh", "csh", "tcsh"]);
const CODEX_SANDBOXES = new Set(["read-only", "workspace-write", "danger-full-access"]);
const CODEX_APPROVALS = new Set(["on-request", "never"]);
// Codex has no `--effort` flag; reasoning effort is a config override.
const CODEX_EFFORTS = new Set(["low", "medium", "high", "xhigh"]);

function runIdForLane(laneId, slot = 0) {
  return slot ? `agenthub-lane-${laneId}-${slot}` : `agenthub-lane-${laneId}`;
}

function slotFromRunId(id) {
  const m = typeof id === "string" ? id.match(RUN_ID_RE) : null;
  return m ? Number(m[2] || 0) : null;
}

function laneIdFromRunId(id) {
  const m = typeof id === "string" ? id.match(RUN_ID_RE) : null;
  return m ? Number(m[1]) : null;
}

function makeErr(code, message) {
  const err = new Error(message);
  err.code = code;
  return err;
}

/**
 * Build the pane's command argv for Claude. Unlike the old stream-json spawner
 * there is no headless/conversation split — every run is a live interactive
 * pane, so an initial prompt (when given) is a trailing POSITIONAL argument:
 * `claude` treats a bare positional as the first turn's message and stays
 * interactive afterward (unlike `-p`, which forces print-and-exit and closes
 * stdin).
 */
/**
 * A prompt as an argv element no CLI can read as an option: one that starts
 * with "-" (`-c fix it`) would otherwise be parsed as flags — `claude -c`
 * continues the last conversation, `-p` turns the pane into print mode. A
 * leading space keeps it a positional and changes nothing the model reads.
 */
function promptArg(p) {
  return p.startsWith("-") ? ` ${p}` : p;
}

function buildClaudeArgv({
  model,
  permissionMode,
  effort,
  resumeSessionId,
  forkSession = false,
  newSessionId = null,
  worktreeName = null,
  initialPrompt,
  addDirs = [],
  systemPromptFile = null,
  routerArgv = [],
}) {
  const argv = ["claude"];
  // `--add-dir` is VARIADIC (`<directories...>`): it swallows every following
  // argument up to the next option, so it goes first, where `--permission-mode`
  // always ends it — after the prompt it would eat the prompt as a directory.
  if (addDirs.length) argv.push("--add-dir", ...addDirs);
  argv.push("--permission-mode", permissionMode || "auto");
  // The session's AI Router: `--settings=<file>` (router-apply.js).
  argv.push(...routerArgv);
  // Claude Code makes the worktree itself (`.claude/worktrees/<name>`, branch
  // `worktree-<name>`, `.worktreeinclude`, the user's WorktreeCreate hook).
  // `=` form: `-w [name]` takes an OPTIONAL value and would otherwise swallow
  // the prompt as the worktree's name.
  if (worktreeName) argv.push(`--worktree=${worktreeName}`);
  // From a file: the instructions can be long, and the whole argv has to fit
  // in one tmux message. `=` form, so the path is never read as an option.
  if (systemPromptFile) argv.push(`--append-system-prompt-file=${systemPromptFile}`);
  // Puts "bypass permissions" INTO the pane's shift+tab cycle without turning
  // it on: `claude --help` — "Enable bypassing all permission checks as an
  // option, without it being enabled by default." Without it the cycle is only
  // plan/auto/manual/acceptEdits, so a run started in any non-bypass mode could
  // never reach bypass from inside the pane, while a real terminal could.
  // The starting mode is still whatever --permission-mode says above.
  argv.push("--allow-dangerously-skip-permissions");
  if (model) argv.push("--model", model);
  if (effort && EFFORT_LEVELS.has(effort)) argv.push("--effort", effort);
  if (resumeSessionId) argv.push("--resume", resumeSessionId);
  else if (newSessionId) argv.push("--session-id", newSessionId);
  // A fork resumes the history into a NEW session id; the original is left as
  // it was.
  if (resumeSessionId && forkSession) argv.push("--fork-session");
  if (initialPrompt) argv.push(promptArg(initialPrompt));
  return argv;
}

/**
 * Codex's CLI shape differs from Claude's in two ways that matter here:
 * `resume` is a SUBCOMMAND (so it precedes every flag) and reasoning effort has
 * no flag of its own — it rides on a `-c` config override. The prompt stays a
 * trailing positional, same as Claude, so the pane stays interactive.
 */
function buildCodexArgv({
  model,
  effort,
  sandbox,
  approval,
  resumeSessionId,
  resumeLast,
  bypass,
  initialPrompt,
  routerArgv = [],
  addDirs = [],
  systemPrompt = null,
  systemPromptFile = null,
}) {
  const argv = ["codex"];
  // Before the subcommand and before the prompt: `-p` layers
  // `$CODEX_HOME/<name>.config.toml` over the base config, and an option that
  // lands after the positional prompt is not an option any more.
  argv.push(...routerArgv);
  if (resumeSessionId) argv.push("resume", resumeSessionId);
  else if (resumeLast) argv.push("resume", "--last");
  if (model) argv.push("-m", model);
  // The bypass flag turns off both sandbox and approvals, so the two flags it
  // replaces are not sent beside it.
  if (bypass) argv.push(...(aiTools.get("codex")?.bypass || []));
  else {
    if (sandbox) argv.push("-s", sandbox);
    if (approval) argv.push("-a", approval);
  }
  if (effort && CODEX_EFFORTS.has(effort)) argv.push("-c", `model_reasoning_effort=${effort}`);
  for (const dir of addDirs) argv.push("--add-dir", dir);
  // `-c` parses its value as TOML; a JSON string literal is a valid TOML basic
  // string, so any text — quotes, newlines — arrives as one string.
  if (systemPrompt) {
    const { codexInstructions, tomlString } = require("./lane-sources");
    const text = codexInstructions(systemPrompt, systemPromptFile);
    if (text) argv.push("-c", `developer_instructions=${tomlString(text)}`);
  }
  if (initialPrompt) argv.push(promptArg(initialPrompt));
  return argv;
}

/**
 * Any catalog tool without a hand-written launcher: `[bin, ...args, ...resume?,
 * ...bypass?, ...prompt]` (`[flag, model]` right after the fixed args when the
 * catalog declares a model flag). Every piece but the prompt and the
 * validated model comes from the catalog
 * file, and the prompt is one argv element — or, for `after-start`, nothing
 * here: spawnRun pastes it into the pane once the CLI is up.
 */
function buildCatalogArgv(entry, bin, { model, resumeLast, bypass, initialPrompt }) {
  const argv = [bin, ...entry.args];
  if (model && entry.model) argv.push(entry.model.flag, model);
  if (resumeLast && entry.resume) argv.push(...entry.resume);
  if (bypass && entry.bypass) argv.push(...entry.bypass);
  const p = entry.prompt;
  if (initialPrompt && p.mode === "flag") argv.push(p.flag, promptArg(initialPrompt));
  if (initialPrompt && p.mode === "argv")
    // A `--` separator already ends the options; without one, guard the prompt.
    argv.push(...(p.separator ? [p.separator, initialPrompt] : [promptArg(initialPrompt)]));
  return argv;
}

function buildArgv(args, entry) {
  if (entry.launcher === "codex") return buildCodexArgv(args);
  if (entry.launcher === "claude") return buildClaudeArgv(args);
  const d = aiTools.detect(entry.id);
  // By absolute path when found; the bare name otherwise, so a CLI missing
  // from PATH dies in the pane with the shell's own "not found".
  return buildCatalogArgv(entry, d.path || entry.detect[0], args);
}

/**
 * @param {object} args
 * @param {number} args.laneId
 * @param {string} args.cwd
 * @param {string} [args.tool="claude"]
 * @param {string} [args.model]
 * @param {string} [args.permissionMode]
 * @param {string} [args.effort]
 * @param {string} [args.sandbox]
 * @param {string} [args.approval]
 * @param {string} [args.resumeSessionId]
 * @param {boolean} [args.forkSession] resume `resumeSessionId` into a new
 *   session id (`claude --fork-session`); Claude only.
 * @param {boolean} [args.worktree] start a NEW Claude session in its own git
 *   worktree (`claude --worktree=<name>`, name made here); Claude only.
 * @param {boolean} [args.inWorktree] `cwd` is a session's worktree (a resume
 *   into it, resolved by the caller from the session row) — recorded only.
 * @param {boolean} [args.resumeLast]
 * @param {boolean} [args.bypass] the user ticked "bypass permissions" for THIS
 *   run; appends the catalog's bypass args/env. Claude has no catalog bypass.
 * @param {string} [args.initialPrompt]
 * @param {string} [args.router] AI Router profile THIS session runs on — a
 *   name only, its key and endpoint come from the store; null/"native" = the
 *   vendor's own API. Claude and Codex only (`EUNSUPPORTED`).
 */
function spawnRun(args) {
  const {
    laneId,
    slot = 0,
    cwd,
    tool = "claude",
    model,
    permissionMode,
    effort,
    sandbox,
    approval,
    resumeSessionId,
    forkSession,
    worktree,
    inWorktree,
    resumeLast,
    bypass,
    initialPrompt,
    router,
  } = args || {};
  if (typeof laneId !== "number" || !Number.isInteger(laneId)) {
    throw makeErr("EBADLANE", "laneId must be an integer");
  }
  if (!Number.isInteger(slot) || slot < 0 || slot > MAX_PANE_SLOT) {
    throw makeErr("EBADSLOT", `slot must be an integer 0–${MAX_PANE_SLOT}`);
  }
  if (typeof cwd !== "string" || !cwd) {
    throw makeErr("EBADCWD", "cwd is required");
  }
  // A lane whose tool has since left the catalog keeps its row; it just
  // cannot start until the catalog knows the tool again.
  const entry = aiTools.get(tool);
  if (!entry) throw makeErr("ENOTOOL", `unknown AI tool: ${tool}`);
  if (tool === "codex") {
    if (sandbox != null && sandbox !== "" && !CODEX_SANDBOXES.has(sandbox)) {
      throw makeErr(
        "EBADSANDBOX",
        `sandbox must be one of: ${Array.from(CODEX_SANDBOXES).join(", ")}`
      );
    }
    if (approval != null && approval !== "" && !CODEX_APPROVALS.has(approval)) {
      throw makeErr(
        "EBADAPPROVAL",
        `approval must be one of: ${Array.from(CODEX_APPROVALS).join(", ")}`
      );
    }
  }
  if (permissionMode != null && !ALLOWED_PERMISSION_MODES.has(permissionMode)) {
    throw makeErr(
      "EBADMODE",
      `permissionMode must be one of: ${Array.from(ALLOWED_PERMISSION_MODES).join(", ")}`
    );
  }
  // A model is one argv element after a model flag; one starting with `-`
  // would be read as a flag of its own.
  if (model != null && model !== "" && !aiTools.validModel(model)) {
    throw makeErr("EBADMODEL", "model is not a valid model id");
  }
  if (effort != null && effort !== "" && !EFFORT_LEVELS.has(effort)) {
    throw makeErr("EBADEFFORT", `effort must be one of: ${Array.from(EFFORT_LEVELS).join(", ")}`);
  }
  if (
    resumeSessionId != null &&
    (typeof resumeSessionId !== "string" || !/^[A-Za-z0-9-]{8,}$/.test(resumeSessionId))
  ) {
    throw makeErr("EBADSESSION", "resumeSessionId is not a valid session id");
  }
  if (worktree === true && entry.launcher !== "claude") {
    throw makeErr("EUNSUPPORTED", `${tool} cannot start a session in its own worktree`);
  }
  const routerProfile = resolveRouter(tool, router); // ENOTFOUND, EUNSUPPORTED
  if (forkSession === true) {
    if (!resumeSessionId) throw makeErr("EBADSESSION", "a fork needs resumeSessionId");
    if (entry.launcher !== "claude") {
      throw makeErr("EUNSUPPORTED", `${tool} cannot fork a session`);
    }
  }

  // A run IS a tmux session, so an absent binary is a precondition failure, not
  // a spawn failure. Without this the caller gets `spawnSync tmux ENOENT` from
  // three directories down — a message that names no fix — and the route maps
  // it to 400, blaming the request for the host's missing package. Checked here
  // rather than in each route because both spawn paths (POST /api/run and
  // POST /api/lanes/:id/start) come through this function.
  if (!tmux.isTmuxAvailable()) {
    throw makeErr(
      "ETMUXMISSING",
      "tmux is not installed, and every run the dashboard starts is a tmux session. " +
        "Install it (sudo apt install tmux / brew install tmux; on Windows run AgentHub under WSL2), " +
        "then restart the dashboard. Everything else keeps working without it."
    );
  }

  const id = runIdForLane(laneId, slot);
  const startedAt = Date.now();
  // Claude Code does not exclude the worktrees it makes, so the main repo
  // would list `.claude/` as untracked; the root's info/exclude hides them.
  if (worktree === true) {
    require("./git-exclude").excludeFromRepoRoot(cwd, [".claude/worktrees/"]);
  }
  // The workspace's own folders, context files and instructions — read from
  // its rows here, the one place both spawn paths pass through, never from a
  // request (server/lib/lane-sources.js).
  const { addDirs, systemPrompt, systemPromptFile } = require("./lane-sources").launchExtras(
    laneId,
    tool
  );
  const routerSpawn = routerForSpawn(tool, routerProfile);
  // A new Claude session gets its id from us, so the run knows the session it
  // holds from the start — a hook that lost AGENTHUB_RUN_ID (an older hook
  // handler) still finds its pane by it (routes/hooks.js → bindRunSession).
  const newSessionId =
    entry.launcher === "claude" && !resumeSessionId ? require("crypto").randomUUID() : null;
  const argv = buildArgv(
    {
      addDirs,
      systemPrompt,
      systemPromptFile,
      tool,
      model,
      permissionMode,
      effort,
      sandbox,
      approval,
      resumeSessionId,
      forkSession: forkSession === true,
      newSessionId,
      worktreeName: worktree === true ? `ah-${laneId}-${Date.now().toString(36)}` : null,
      resumeLast,
      bypass: bypass === true,
      initialPrompt,
      routerArgv: routerSpawn.argv,
    },
    entry
  );
  const pasteLater = entry.prompt?.mode === "after-start" && initialPrompt ? initialPrompt : null;

  const record = () => {
    if (!dashboardRuns) return;
    dashboardRuns.recordRun({
      id,
      // A fork's own id arrives with its first hook (bindRunSession).
      sessionId: newSessionId || (forkSession !== true && resumeSessionId) || null,
      // `dashboard_runs.mode` is NOT NULL. Passing null made every insert throw
      // inside recordRun's best-effort catch, so no tmux-backed run was ever
      // persisted: publicRun then read its cwd/model/prompt/startedAt off a row
      // that did not exist (a blank row dated Jan 1 1970 in the runs modal), and
      // session_id never landed either, so a resumed session stayed listed as a
      // separate `external` row. Every run here is one interactive pane.
      mode: "interactive",
      cwd,
      tool,
      model: model || null,
      permissionMode: permissionMode || "auto",
      effort: effort || null,
      resumeSessionId: resumeSessionId || null,
      worktree: worktree === true || inWorktree === true,
      router: routerProfile ? routerProfile.name : null,
      prompt: initialPrompt || "",
      status: "running",
      startedAt,
      endedAt: null,
      exitCode: null,
      laneId,
    });
  };

  // Every pane the dashboard starts carries its lane id. A `agenthub stage` call
  // from inside it then resolves the lane exactly, instead of matching on cwd —
  // which is ambiguous now that one directory can back both a `claude` lane and
  // a `codex` lane (see bin/agenthub.js's resolveLaneArg).
  // A codex session on an AI Router also needs the key its layer's `env_key`
  // names. A claude one needs nothing here — its `--settings` file carries the
  // router — and a session on the vendor's own API needs nothing either.
  // The graphical session's own DISPLAY/WAYLAND_DISPLAY goes in too: the
  // server normally runs as a systemd user service with neither, so a pane it
  // starts has no clipboard at all — pasting an image into `claude` there
  // answers "no image found in clipboard" while the same CLI pastes fine in a
  // terminal. Never overrides what this spawn already needs.
  // PATH is the dashboard's own, hydrated from the login shell: a pane's env
  // comes from the tmux server, and an npm-installed CLI (`#!/usr/bin/env
  // node`) dies there without the `node` the user's shell would have found.
  const env = {
    AGENTHUB_LANE_ID: String(laneId),
    // Which pane this is: the CLI's hooks inherit it, so the dashboard learns
    // which session each pane holds (routes/hooks.js → bindRunSession) — the
    // sidebar opens a session's own pane by it.
    AGENTHUB_RUN_ID: id,
    ...(process.env.PATH ? { PATH: process.env.PATH } : {}),
    ...desktopEnv(),
    // A codex pane must read the same home the dashboard reads — its rollout
    // journals and the router layer `-p <name>` loads live there — and the
    // tmux server's own env may predate, or never have had, CODEX_HOME.
    ...(tool === "codex" ? { CODEX_HOME: getCodexHome() } : {}),
    ...routerSpawn.env,
    ...(bypass === true && entry.bypassEnv ? entry.bypassEnv : {}),
  };

  // `after-start` tools take no prompt on their command line: paste it once
  // the CLI has had `delayMs` to draw its input box. ponytail: fixed delay,
  // add readiness detection (Orca's) if a tool drops prompts.
  const schedulePaste = () => {
    if (!pasteLater) return;
    const timer = setTimeout(() => {
      try {
        if (tmux.hasSession(id)) tmux.pasteText(id, pasteLater);
      } catch {
        /* pane gone mid-paste — nothing to deliver to */
      }
    }, entry.prompt.delayMs);
    if (timer.unref) timer.unref();
  };

  if (!tmux.hasSession(id)) {
    tmux.newSession({ name: id, cwd, argv, env });
    record();
    schedulePaste();
  } else if (SHELL_COMMANDS.has(tmux.paneCommand(id) || "")) {
    // The session exists but its pane is sitting at a bare shell prompt — a
    // `agenthub lanes shell`, or a `claude` that already exited. Adopting it
    // silently here would swallow the whole request: a Resume would spawn no
    // `--resume` and an initial prompt would never be typed, while the API
    // still answered 200. Run the argv in the pane the user already sees
    // instead of erroring or opening a second session.
    tmux.sendCommand(id, argv, env, cwd);
    record();
    schedulePaste();
  }
  // Pane is running something (a live `claude`, an editor, a build): adopt
  // silently, same convention as this repo's server port-adoption logic — no
  // error, no duplicate session. Attaching shows the user what is running.
  //
  // Adopting is only honest when the request asked for nothing more than a
  // pane. A Resume or an initial prompt cannot be delivered into a pane that is
  // already running an agent (typing over it is what the shell-pane branch
  // above exists to avoid), so answering 200 there reports work that never
  // happened: the user sees "resumed" while no `--resume` was ever spawned.
  else if (resumeSessionId || initialPrompt) {
    throw makeErr(
      "ERUNLIVE",
      `workspace ${laneId}'s pane is already running \`${tmux.paneCommand(id) || "a process"}\` — stop it before resuming a session or sending a prompt`
    );
  }

  return getRun(id);
}

// Only a dashboard lane pane is a run. Without this, `DELETE /api/run/main`
// would kill the user's own tmux session named `main`.
function killRun(id) {
  if (laneIdFromRunId(id) === null || !tmux.hasSession(id)) return false;
  tmux.killSession(id);
  if (dashboardRuns) {
    dashboardRuns.patchRun({ id, status: "killed", endedAt: Date.now() });
  }
  return true;
}

function publicRun(id) {
  const laneId = laneIdFromRunId(id);
  const live = tmux.hasSession(id);
  const row = dashboardRuns ? dashboardRuns.getRun(id) : null;
  return {
    id,
    laneId,
    slot: slotFromRunId(id),
    status: live ? "running" : "gone",
    tool: row?.tool || "claude",
    cwd: row?.cwd || null,
    model: row?.model || null,
    permissionMode: row?.permission_mode || null,
    effort: row?.effort || null,
    resumeSessionId: row?.resume_session_id || null,
    sessionId: row?.session_id || null,
    // The pane works in a git worktree of its own (`claude --worktree`).
    worktree: !!row?.worktree,
    startedAt: row?.started_at || null,
    promptPreview: row?.prompt_preview || null,
  };
}

function getRun(id) {
  if (laneIdFromRunId(id) === null) return null;
  return publicRun(id);
}

/** Computed fresh from tmux state every call — see file header. */
function listRuns() {
  return tmux.listSessions("agenthub-lane-").map(publicRun);
}

module.exports = {
  spawnRun,
  killRun,
  getRun,
  listRuns,
  laneIdFromRunId,
  runIdForLane,
  MAX_PANE_SLOT,
  SHELL_COMMANDS,
};