@openagents-org/agent-launcher 0.2.168 → 0.2.169

package/src/adapters/deepseek.js

@@ -0,0 +1,1190 @@
+/**
+ * DeepSeek Harness adapter for OpenAgents workspace.
+ *
+ * Bridges DeepSeek's open-source agent harness (https://github.com/deepseek-ai/
+ * deepseek-harness, npm `@deepseek-ai/dsh`) by driving its headless profile:
+ *
+ *   dsh --profile headless --patch <private.yml> "<constant instruction>"
+ *
+ * Like every other agent here it drives its own CLI subprocess; no SDK is
+ * introduced. Mirrors mini.js's one-process-per-message control flow and
+ * aider.js's plain-text result handling.
+ *
+ * Four things are unlike the other adapters, and each has a reason:
+ *
+ *  1. THE PROMPT IS NEVER IN argv. dsh takes its task as a positional argument
+ *     and has no stdin task channel, so putting the workspace prompt there
+ *     would publish the workspace token to /proc/<pid>/cmdline and `ps`. The
+ *     prompt is written to a private task file and argv carries only a constant
+ *     sentence naming it (deepseek-runtime's HEADLESS_TASK_INSTRUCTION). If the
+ *     task file cannot be written the run FAILS — it never falls back to argv,
+ *     because a fallback would reintroduce exactly the exposure the design
+ *     exists to prevent.
+ *
+ *  2. NO SESSION RESUME. headless creates one fresh persisted session per run
+ *     and upstream documents "one submitted task only". Continuity therefore
+ *     comes from a bounded workspace recap injected into the task file, not
+ *     from anything dsh remembers. That also means the sessions directory grows
+ *     once per message, so this adapter garbage-collects it.
+ *
+ *  3. NO STREAMING. headless prints the last assistant message at the end and
+ *     nothing before it. A conventional idle timeout would therefore kill every
+ *     healthy long task; the guard here is a TOTAL run timeout instead. For the
+ *     same reason the child is awaited on 'close' rather than 'exit': the whole
+ *     answer is stdout, and 'exit' can fire with the tail unflushed.
+ *
+ *  4. NON-INTERACTIVE APPROVAL WITHOUT UNSANDBOXING. dsh's composition sets
+ *     approval `ask` everywhere except `danger-full-access`, and a headless run
+ *     has no client to answer. The private patch sets approval `never` while
+ *     leaving the sandbox at workspace-write, so tool use is non-interactive
+ *     but writes stay confined. Reaching the same place via
+ *     DSH_PERMISSION_MODE=danger-full-access would also remove the sandbox,
+ *     which is why that value is never used to solve this.
+ */
+
+'use strict';
+
+const fs = require('fs');
+const os = require('os');
+const path = require('path');
+const { execFileSync, spawn, execSync } = require('child_process');
+
+const BaseAdapter = require('./base');
+const { whichBinary, getEnhancedEnv } = require('../paths');
+const { formatAttachmentsForPrompt } = require('./utils');
+const { buildDeepSeekTaskFile } = require('./workspace-prompt');
+const { sampleRecap } = require('./decision-log');
+const {
+  REASON,
+  classifySpawnError,
+  redactDiagnostic,
+} = require('./health-status');
+const {
+  SUPPORTED_DSH_VERSION,
+  dshEntryCandidates,
+  buildHeadlessArgs,
+  buildDumpConfigArgs,
+  buildPrivatePatch,
+  classifyDshFailure,
+  classifyDshVersion,
+  classifyNodeVersion,
+  cleanStdout,
+  defaultInstallCommand,
+  nodeRequirementText,
+  resolvePermissionMode,
+  safeDshHomeName,
+  selectSessionsForGc,
+  MAX_STDOUT_BYTES,
+  MAX_STDERR_BYTES,
+} = require('./deepseek-runtime');
+
+const IS_WINDOWS = process.platform === 'win32';
+
+/**
+ * TOTAL wall-clock budget for one run — NOT an idle timeout.
+ *
+ * dsh emits nothing on stdout until the task finishes, so the "no output for N
+ * minutes" guard every other adapter uses would terminate perfectly healthy
+ * long tasks. 60 minutes matches pi.js's long-task scale.
+ */
+const RUN_TIMEOUT_MS = 60 * 60 * 1000;
+
+/** Grace period between SIGTERM and SIGKILL when stopping a run. */
+const STOP_GRACE_MS = 5000;
+
+/**
+ * Budget for composing the headless profile.
+ *
+ * `--dump-config` calls no model and does no work beyond reading and merging
+ * the profile's own files, so it is fast or it is broken. The generous ceiling
+ * covers a cold first run that still has to materialise the profile from the
+ * shipped templates.
+ */
+const BOOTSTRAP_TIMEOUT_MS = 3 * 60 * 1000;
+
+/**
+ * The key bootstrap's child is tracked under in `_channelProcesses`.
+ *
+ * The NUL prefix makes it unrepresentable as a real channel name, so a user
+ * `/stop` can reach the bootstrap process through the same registry the
+ * per-channel runs use, with no chance of colliding with one of them.
+ */
+const BOOTSTRAP_KEY = '\u0000bootstrap';
+
+/** How many messages of channel history to pull for the recap. */
+const RECAP_HEAD = 30;
+const RECAP_TAIL = 60;
+
+/** The environment variable the task file tells the agent to read. */
+const TOKEN_ENV = 'OPENAGENTS_WORKSPACE_TOKEN';
+
+/**
+ * The status that ends a run in the workspace UI.
+ *
+ * The client clears its "stopping" state only on a status matching
+ * /stopped|stopping failed/i (chat-view.tsx). Any other final status leaves the
+ * session marked as working forever — and a cancelled run produces nothing
+ * later that could correct it. The wording is therefore load-bearing, and the
+ * tests assert it against the client's own regex.
+ */
+const STOPPED_STATUS = 'Execution stopped by user';
+
+/**
+ * The status for a stop that could NOT be confirmed.
+ *
+ * Contains "stopping failed" because that is the other half of the client's
+ * terminal-status regex. Reporting a clean stop when the child may still be
+ * editing files is worse than reporting nothing: it clears the UI and tells the
+ * user the work is over while it is not.
+ */
+const STOP_FAILED_STATUS =
+  'Stopping failed — the DeepSeek Harness process did not exit and may still be '
+  + 'running. It may continue to change files in the working directory.';
+
+
+/** nodeBin -> `node --version`, or null when it could not be read. */
+const _nodeVersionCache = new Map();
+
+function dshInstallHint() {
+  return defaultInstallCommand(SUPPORTED_DSH_VERSION);
+}
+
+class DeepSeekAdapter extends BaseAdapter {
+  constructor(opts) {
+    super(opts);
+    this.disabledModules = opts.disabledModules || new Set();
+
+    // channel -> running child process (for stop / cleanup)
+    this._channelProcesses = {};
+    // channels the user explicitly stopped (suppress "no response" noise)
+    this._stoppingChannels = new Set();
+
+    // Channels currently inside _handleMessage, whether or not a child exists
+    // yet. A run spends real time before it spawns anything — bootstrap, the
+    // status post, the recap fetch, the browser-setting lookup — and a `/stop`
+    // arriving in that window has no process to kill. Without this set the stop
+    // would be silently dropped and the task would start anyway.
+    this._busyChannels = new Set();
+    // Channels cancelled while busy. Checked at every await boundary and once
+    // more immediately before spawning.
+    this._cancelledChannels = new Set();
+    // channel -> resolve() that releases a waiter parked on the SHARED
+    // bootstrap. Stopping one conversation must let it go without tearing down
+    // the compose that the other channels are still waiting on.
+    this._cancelWaiters = new Map();
+    this._taskCounter = 0;
+    this._bootstrapPromise = null;
+    // Overridable so the timeout paths are testable without 60-minute and
+    // 3-minute waits.
+    this._runTimeoutMs = opts.runTimeoutMs || RUN_TIMEOUT_MS;
+    this._bootstrapTimeoutMs = opts.bootstrapTimeoutMs || BOOTSTRAP_TIMEOUT_MS;
+
+    // A PRIVATE harness home per agent instance. The user's own ~/.dsh is never
+    // read or written: sharing it would leak their profiles, patches and saved
+    // credentials into an agent, and would make two OpenAgents agents share
+    // sessions and settings. The name carries a hash so two agents whose ids
+    // slug identically still get separate homes.
+    this._dshHome = path.join(
+      os.homedir(), '.openagents', 'dsh-homes',
+      safeDshHomeName(this.workspaceId, this.agentName),
+    );
+    this._patchFile = path.join(this._dshHome, 'openagents.patch.yml');
+    this._tasksDir = path.join(this._dshHome, 'tasks');
+
+    const resolved = this._resolveDshCommand();
+    this._nodeBin = resolved ? resolved[0] : null;
+    this._jsEntry = resolved ? resolved[1] : null;
+    if (this._jsEntry) {
+      this._log(`Using DeepSeek Harness: ${this._jsEntry} (node: ${this._nodeBin})`);
+    } else {
+      this._log(`Warning: DeepSeek Harness not found — install with: ${dshInstallHint()}`);
+    }
+  }
+
+  // ------------------------------------------------------------------
+  // Binary resolution
+  // ------------------------------------------------------------------
+
+  /**
+   * Every Node this adapter could launch dsh with, best candidate first.
+   *
+   * Bare 'node' is never among them: a packaged daemon's PATH may not have one.
+   */
+  _nodeCandidates() {
+    const home = os.homedir();
+    const managed = IS_WINDOWS
+      ? [path.join(home, '.openagents', 'nodejs', 'node.exe')]
+      : [path.join(home, '.openagents', 'nodejs', 'node'),
+        path.join(home, '.openagents', 'nodejs', 'bin', 'node')];
+    const out = [];
+    for (const c of managed) {
+      try { if (fs.existsSync(c)) out.push(c); } catch { /* unreadable */ }
+    }
+    if (!out.includes(process.execPath)) out.push(process.execPath);
+    return out;
+  }
+
+  /**
+   * Pick a Node that can actually run dsh.
+   *
+   * Preferring the launcher's managed runtime is right — it is the one version
+   * the product controls — but preferring it *unconditionally* is not: a
+   * managed runtime older than dsh's floor shadows a perfectly good system
+   * Node, and the agent then refuses to start on a machine that could run it.
+   * Observed on Windows with a managed 22.14.0 next to a system 24.15.0.
+   *
+   * So: first candidate that satisfies the version gate wins. When none does,
+   * return the first that exists so preflight can name it in the error rather
+   * than reporting something vague.
+   */
+  _findNodeBin() {
+    const candidates = this._nodeCandidates();
+    for (const c of candidates) {
+      if (classifyNodeVersion(this._readNodeVersionOf(c)).supported) return c;
+    }
+    return candidates[0];
+  }
+
+  /**
+   * Resolve the `dsh` shim to [nodeBin, jsEntry] so the ESM entry point runs
+   * under a Node we control and whose version we have gated.
+   *
+   * dsh is `"type": "module"` with bin `lib/bin.js`. Running the shim instead
+   * would (a) use whatever Node is first on PATH, which may be a version dsh
+   * cannot run on, and (b) on Windows wrap a `.cmd` in cmd.exe. Returns null
+   * when nothing resolves.
+   */
+  _resolveDshCommand() {
+    const bin = whichBinary('dsh');
+    if (!bin) return null;
+    const nodeBin = this._findNodeBin();
+
+    // A resolved path that is already the JS entry point (or a symlink to it).
+    try {
+      let target = bin;
+      if (!IS_WINDOWS && fs.lstatSync(bin).isSymbolicLink()) {
+        target = path.resolve(path.dirname(bin), fs.readlinkSync(bin));
+      }
+      if (/\.(js|mjs)$/i.test(target) && fs.existsSync(target)) return [nodeBin, target];
+    } catch { /* fall through to the package lookup */ }
+
+    // Otherwise locate lib/bin.js inside the installed package. The candidate
+    // layouts are a pure function so every install shape is unit-testable
+    // without a filesystem — see dshEntryCandidates.
+    const guesses = dshEntryCandidates(bin);
+    for (const g of guesses) if (fs.existsSync(g)) return [nodeBin, g];
+    return null;
+  }
+
+  /** `dsh -V` output, or null. Synchronous: preflight() is a sync contract. */
+  _readDshVersion() {
+    if (!this._jsEntry) return null;
+    try {
+      return execFileSync(this._nodeBin, [this._jsEntry, '-V'], {
+        encoding: 'utf-8',
+        timeout: 15000,
+        env: getEnhancedEnv(this.agentEnv),
+      }).trim();
+    } catch {
+      return null;
+    }
+  }
+
+  /**
+   * `node --version` for a specific binary, memoised.
+   *
+   * Node selection probes every candidate, and this runs on the constructor
+   * path, so the results are cached process-wide — a Node binary does not
+   * change version underneath us.
+   */
+  _readNodeVersionOf(bin) {
+    if (!bin) return null;
+    if (bin === process.execPath) return process.version;
+    if (_nodeVersionCache.has(bin)) return _nodeVersionCache.get(bin);
+    let version = null;
+    try {
+      version = execFileSync(bin, ['--version'], {
+        encoding: 'utf-8',
+        timeout: 10000,
+      }).trim();
+    } catch { /* unreadable or not a Node */ }
+    _nodeVersionCache.set(bin, version);
+    return version;
+  }
+
+  /** `node --version` for the Node we will actually launch dsh with. */
+  _readNodeVersion() {
+    return this._readNodeVersionOf(this._nodeBin);
+  }
+
+  // ------------------------------------------------------------------
+  // Preflight (run by the daemon BEFORE joining the workspace)
+  // ------------------------------------------------------------------
+
+  /**
+   * Gate everything that can be known without starting a task. The daemon skips
+   * the workspace join when this fails, so an unusable runtime never spins a
+   * poll loop it can do nothing with.
+   *
+   * Deliberately ordered cheapest-first: a missing binary makes the version
+   * probes meaningless, and both version probes must pass before bootstrap is
+   * ever allowed to start a dsh process.
+   */
+  preflight() {
+    if (!this._jsEntry) {
+      const resolved = this._resolveDshCommand();
+      if (resolved) { this._nodeBin = resolved[0]; this._jsEntry = resolved[1]; }
+    }
+    if (!this._jsEntry) {
+      return {
+        ok: false,
+        reason: REASON.RUNTIME_MISSING,
+        message: `DeepSeek Harness (dsh) not found — install with: ${dshInstallHint()}`,
+      };
+    }
+
+    const node = classifyNodeVersion(this._readNodeVersion());
+    if (node.supported === false) {
+      // Name the binary: the usual cause is a stale managed runtime under
+      // ~/.openagents/nodejs shadowing a newer system Node, and without the
+      // path there is no way to tell which one to upgrade.
+      return {
+        ok: false,
+        reason: REASON.VERSION_INCOMPATIBLE,
+        message:
+          `DeepSeek Harness cannot run on Node ${node.version} `
+          + `(${this._nodeBin}) — it requires ${nodeRequirementText()}. `
+          + 'No compatible Node was found; upgrade the runtime at that path, '
+          + 'or install a supported Node.',
+      };
+    }
+
+    const entry = this._registryEntry();
+    const supported = (entry && entry.install && entry.install.supported_version)
+      || SUPPORTED_DSH_VERSION;
+    const installCmd = (entry && entry.install && entry.install[this._platformKey()])
+      || defaultInstallCommand(supported);
+    const gate = classifyDshVersion(this._readDshVersion(), supported, installCmd);
+    if (gate.compatible === false || gate.compatible === null) {
+      return { ok: false, reason: REASON.VERSION_INCOMPATIBLE, message: gate.message };
+    }
+
+    if (!this._apiKey()) {
+      return {
+        ok: false,
+        reason: REASON.LOGIN_REQUIRED,
+        message:
+          'DeepSeek Harness is not configured — set DEEPSEEK_API_KEY. The agent '
+          + 'runs with a private, empty harness home, so there is no saved dsh '
+          + 'login to fall back on.',
+      };
+    }
+
+    this._log(`DeepSeek Harness ${gate.detected} resolved (node ${node.version})`);
+    return { ok: true };
+  }
+
+  _platformKey() {
+    if (IS_WINDOWS) return 'windows';
+    return process.platform === 'darwin' ? 'macos' : 'linux';
+  }
+
+  /** The bundled registry entry for this agent type, or null. */
+  _registryEntry() {
+    try {
+      // eslint-disable-next-line global-require
+      const entries = require('../../registry.json');
+      return entries.find((e) => e && e.name === 'deepseek') || null;
+    } catch {
+      return null;
+    }
+  }
+
+  _apiKey() {
+    return String(this.agentEnv.DEEPSEEK_API_KEY || this.agentEnv.LLM_API_KEY || '').trim();
+  }
+
+  _baseUrl() {
+    return String(this.agentEnv.DEEPSEEK_BASE_URL || this.agentEnv.LLM_BASE_URL || '').trim();
+  }
+
+  _model() {
+    return String(this.agentEnv.DEEPSEEK_MODEL || this.agentEnv.LLM_MODEL || '').trim();
+  }
+
+  // ------------------------------------------------------------------
+  // Bootstrap + private patch
+  // ------------------------------------------------------------------
+
+  /**
+   * Compose the headless profile once, before any channel runs a task.
+   *
+   * A fresh DSH_HOME auto-initialises its headless profile from bundled
+   * templates on first use. BaseAdapter dispatches channels concurrently, so
+   * without this several dsh processes would race to initialise the same
+   * profile directory. `--dump-config` composes and prints the tree without
+   * booting it: no model call, no API key needed, no task run.
+   *
+   * The promise is CLEARED on failure. Caching a rejection would make one bad
+   * first run permanent for the lifetime of the daemon process.
+   */
+  _ensureBootstrap() {
+    if (!this._bootstrapPromise) {
+      this._bootstrapPromise = this._bootstrap().catch((e) => {
+        this._bootstrapPromise = null;
+        throw e;
+      });
+    }
+    return this._bootstrapPromise;
+  }
+
+  /**
+   * Wait for the SHARED bootstrap, but let THIS channel give up on its own.
+   *
+   * A scoped stop deliberately leaves the bootstrap child running, because
+   * every other channel is waiting on the same compose. Without a per-waiter
+   * escape the stopped channel stayed parked on that promise anyway — the user
+   * was told the run had stopped while the channel stayed busy, and the next
+   * message queued behind it for up to the full bootstrap timeout.
+   *
+   * Returns true when bootstrap completed, false when this channel was
+   * cancelled while waiting. Rethrows a genuine bootstrap failure.
+   */
+  async _awaitBootstrap(channel) {
+    const shared = this._ensureBootstrap();
+    // This waiter may lose the race; without a handler its rejection would
+    // surface as an unhandled promise rejection.
+    shared.catch(() => {});
+
+    const cancelled = new Promise((resolve) => {
+      this._cancelWaiters.set(channel, resolve);
+    });
+    try {
+      return await Promise.race([
+        shared.then(() => true),
+        cancelled.then(() => false),
+      ]);
+    } finally {
+      this._cancelWaiters.delete(channel);
+    }
+  }
+
+  async _bootstrap() {
+    fs.mkdirSync(this._dshHome, { recursive: true, mode: 0o700 });
+    fs.mkdirSync(this._tasksDir, { recursive: true, mode: 0o700 });
+    try { fs.chmodSync(this._dshHome, 0o700); } catch { /* best effort (Windows) */ }
+    this._writePrivatePatch();
+
+    const args = buildDumpConfigArgs({ jsEntry: this._jsEntry, patchFile: this._patchFile });
+    await new Promise((resolve, reject) => {
+      const proc = spawn(this._nodeBin, args, {
+        stdio: ['ignore', 'ignore', 'pipe'],
+        env: this._buildSubprocessEnv('workspace-write'),
+        cwd: this.workingDir,
+        detached: !IS_WINDOWS,
+        windowsHide: true,
+      });
+      // Bootstrap runs BEFORE any per-run timeout exists and every waiting
+      // channel is blocked on it, so a wedged `--dump-config` would hang the
+      // whole agent with nothing able to interrupt it. Tracking it under a
+      // reserved key makes it reachable from /stop; the timeout is the backstop
+      // for a user who never presses it.
+      this._channelProcesses[BOOTSTRAP_KEY] = proc;
+
+      let err = '';
+      let settled = false;
+      let bootTimedOut = false;
+      const bootTimeoutError = () => new Error(
+        `dsh profile bootstrap timed out after ${this._bootstrapTimeoutMs / 1000}s. `
+        + 'The harness composes its profile without calling a model, so this '
+        + 'usually means the install is incomplete — reinstall with: '
+        + dshInstallHint(),
+      );
+      const finish = (fn, value) => {
+        if (settled) return;
+        settled = true;
+        clearTimeout(timer);
+        delete this._channelProcesses[BOOTSTRAP_KEY];
+        fn(value);
+      };
+      const timer = setTimeout(async () => {
+        // Mark BEFORE killing: the kill produces a `close` with a null exit
+        // code, and without this flag that generic "exit null" would win the
+        // race and hide the actual reason from the user.
+        bootTimedOut = true;
+        this._log(`DeepSeek Harness profile bootstrap exceeded ${this._bootstrapTimeoutMs / 1000}s — terminating`);
+        await this._stopProcess(proc).catch(() => false);
+        finish(reject, bootTimeoutError());
+      }, this._bootstrapTimeoutMs);
+
+      proc.stderr.on('data', (d) => {
+        if (err.length < MAX_STDERR_BYTES) err += d.toString();
+      });
+      proc.stderr.on('error', () => {});
+      proc.on('error', (e) => finish(reject, bootTimedOut ? bootTimeoutError() : e));
+      proc.on('close', (code) => {
+        if (bootTimedOut) return finish(reject, bootTimeoutError());
+        if (code === 0) return finish(resolve, undefined);
+        finish(reject, new Error(
+          `dsh profile bootstrap failed (exit ${code}): ${redactDiagnostic(err)}`,
+        ));
+      });
+    });
+    this._log(`DeepSeek Harness profile ready in ${this._dshHome}`);
+  }
+
+  /**
+   * (Re)write the private `--patch` overlay.
+   *
+   * Rewritten on every run so a changed model setting takes effect without the
+   * user having to recreate the agent, and written atomically so a run that
+   * starts mid-write can never read a half-file.
+   */
+  _writePrivatePatch() {
+    const { text } = buildPrivatePatch({ model: this._model() });
+    const tmp = `${this._patchFile}.${process.pid}.tmp`;
+    fs.writeFileSync(tmp, text, { encoding: 'utf-8', mode: 0o600 });
+    fs.renameSync(tmp, this._patchFile);
+  }
+
+  // ------------------------------------------------------------------
+  // Subprocess environment
+  // ------------------------------------------------------------------
+
+  /**
+   * The child environment is the ONLY channel for secrets: nothing below ever
+   * reaches argv, the task file or a log.
+   *
+   * Note that the daemon merges process.env into the agent environment, so this
+   * is "the effective child environment is authoritative", not "the Launcher is
+   * the only possible source". What matters for isolation is DSH_HOME: the
+   * agent never reads the user's shared harness credentials.
+   */
+  _buildSubprocessEnv(permissionMode) {
+    const env = getEnhancedEnv(this.agentEnv);
+    if (env.NO_COLOR === undefined) env.NO_COLOR = '1';
+
+    env.DSH_HOME = this._dshHome;
+    env.DSH_PERMISSION_MODE = permissionMode;
+    // Telemetry uploads mirror session-log records with no redaction rule, so it
+    // is off unconditionally. dsh treats ANY non-empty value as opt-out.
+    env.DSH_TELEMETRY_DISABLED = '1';
+
+    const key = this._apiKey();
+    if (key) env.DEEPSEEK_API_KEY = key;
+    const base = this._baseUrl();
+    if (base) env.DEEPSEEK_BASE_URL = base;
+
+    // The task file references this by NAME; the value never appears in it.
+    env[TOKEN_ENV] = this.token;
+    return env;
+  }
+
+  /** The shell expression the task file uses to reference the token. */
+  _tokenExpr() {
+    return IS_WINDOWS ? `$env:${TOKEN_ENV}` : `$${TOKEN_ENV}`;
+  }
+
+  // ------------------------------------------------------------------
+  // Control actions (stop)
+  // ------------------------------------------------------------------
+
+  async _onControlAction(action, payload) {
+    if (action === 'stop') {
+      // The workspace sends the conversation it wants stopped. Honouring it is
+      // what keeps one user's stop from tearing down every other concurrent
+      // run — claude.js already scopes its stop this way; mini/amp do not, and
+      // following them here was the mistake. A stop with no channel still
+      // means "everything".
+      const target = (payload && typeof payload === 'object' && payload.channel)
+        ? payload.channel
+        : null;
+      await this._stopChannels(target);
+      return;
+    }
+    await super._onControlAction(action, payload);
+  }
+
+  /**
+   * Stop one channel, or every channel when `target` is null.
+   *
+   * Each affected channel is told exactly once what happened — and told the
+   * TRUTH: a kill that could not be confirmed reports a failed stop rather than
+   * a clean one, because the child may still be writing to the workspace.
+   */
+  async _stopChannels(target) {
+    // channel -> was termination confirmed
+    const outcome = new Map();
+
+    const busy = target
+      ? (this._busyChannels.has(target) ? [target] : [])
+      : [...this._busyChannels];
+    for (const channel of busy) {
+      this._cancelledChannels.add(channel);
+      this._stoppingChannels.add(channel);
+      // Release a waiter parked on the shared bootstrap. Marking alone is not
+      // enough: the flag is only read between awaits, and this one can last
+      // the whole bootstrap timeout.
+      const release = this._cancelWaiters.get(channel);
+      if (release) release();
+      // Nothing has spawned yet, so cancellation is complete by definition.
+      outcome.set(channel, true);
+    }
+
+    for (const [channel, proc] of Object.entries(this._channelProcesses)) {
+      if (channel === BOOTSTRAP_KEY) {
+        // Bootstrap is SHARED: every channel waits on the same profile
+        // compose. Tearing it down because one conversation was stopped would
+        // fail the others too, so a scoped stop leaves it alone.
+        if (target) continue;
+      } else if (target && channel !== target) {
+        continue;
+      }
+
+      // Untrack only once the kill has been attempted, so the child is never
+      // missing from the registry while it may still be running.
+      const ended = await this._stopProcess(proc);
+      delete this._channelProcesses[channel];
+
+      if (channel === BOOTSTRAP_KEY) continue;
+      this._stoppingChannels.add(channel);
+      outcome.set(channel, ended);
+    }
+
+    // The client clears its stopping state only on a terminal status, and a
+    // cancelled run produces nothing later that could supply one.
+    for (const [channel, ended] of outcome) {
+      try {
+        await this.sendStatus(channel, ended ? STOPPED_STATUS : STOP_FAILED_STATUS);
+      } catch { /* the run is already over; a failed post must not throw here */ }
+    }
+  }
+
+  /** True when a child has ended, by exit code OR by signal. */
+  _hasExited(proc) {
+    return !!proc && (proc.exitCode !== null || proc.signalCode !== null);
+  }
+
+  /**
+   * Terminate a child, reporting whether its exit was observed.
+   *
+   * Best-effort, exactly like every other adapter here (mini, amp, aider,
+   * goose): SIGTERM the process group, escalate to SIGKILL after a grace
+   * period, give up after that. This adapter briefly carried extra machinery —
+   * process-group polling and a registry of children whose kill could not be
+   * confirmed — to retry failed kills and hold off session GC. It was removed:
+   * it caused several regressions of its own, and the two risks it covered are
+   * remote (a session in flight is the NEWEST by mtime, so the GC's
+   * keep-newest-50 and keep-7-days bounds already spare it, and re-sending
+   * SIGKILL to something SIGKILL did not end rarely helps).
+   *
+   * The boolean is still worth having: it decides whether the user is told the
+   * run stopped or that stopping FAILED. It reflects the leader — on POSIX a
+   * descendant of the group could in principle outlive it — so it is a claim
+   * about the process we started, not a proof that every grandchild is gone.
+   */
+  async _stopProcess(proc) {
+    if (!proc) return true;
+    if (this._hasExited(proc)) return true;
+    try {
+      if (IS_WINDOWS) {
+        // /T takes the whole tree, so there is no separate group to signal.
+        try {
+          execSync(`taskkill /F /T /PID ${proc.pid}`, { timeout: 5000 });
+          return true;
+        } catch {
+          return this._hasExited(proc);
+        }
+      }
+      // POSIX: signal the whole process group (the child is detached) so the
+      // shell commands dsh spawned for its bash tool are reaped too.
+      try { process.kill(-proc.pid, 'SIGTERM'); } catch { proc.kill('SIGTERM'); }
+      return await new Promise((resolve) => {
+        const timer = setTimeout(() => {
+          try { process.kill(-proc.pid, 'SIGKILL'); } catch { proc.kill('SIGKILL'); }
+          setTimeout(() => resolve(this._hasExited(proc)), 250);
+        }, STOP_GRACE_MS);
+        proc.on('exit', () => { clearTimeout(timer); resolve(true); });
+      });
+    } catch {
+      return false;
+    }
+  }
+
+  /** True when an error is a child_process spawn failure (vs. a runtime error). */
+  _isSpawnError(e) {
+    if (!e) return false;
+    const code = e.code || e.errno;
+    return (
+      e.syscall === 'spawn'
+      || String(e.syscall || '').startsWith('spawn ')
+      || code === 'ENOENT'
+      || code === 'EACCES'
+      || code === 'EPERM'
+    );
+  }
+
+  // ------------------------------------------------------------------
+  // Channel recap
+  // ------------------------------------------------------------------
+
+  /**
+   * Bounded channel history for the task file.
+   *
+   * Two queries, matching claude.js: an ascending window (the opening messages
+   * — what the channel is *for*) and a descending one (the most recent
+   * messages — what just happened).
+   *
+   * BOTH arrive in chronological order. `getRecentMessages` reverses a `desc`
+   * window itself so every caller gets the same ordering, so reversing the tail
+   * again here would hand `sampleRecap` a newest-first list, and its
+   * `tail.slice(-tailKeep)` would then keep the OLDEST entries and discard the
+   * newest context — the exact opposite of the intent.
+   *
+   * The current message is excluded BY ID. `sampleRecap`'s own guard compares
+   * message TEXT, which silently drops an older message that happens to repeat
+   * the current one verbatim — common for short instructions like "continue".
+   * Filtering by id first makes that impossible; the text guard stays as a
+   * fallback for events that carry no id.
+   */
+  async _buildRecap(channel, msg) {
+    try {
+      const [head, tail] = await Promise.all([
+        this.client.getRecentMessages(this.workspaceId, channel, this.token, RECAP_HEAD, { sort: 'asc' }),
+        this.client.getRecentMessages(this.workspaceId, channel, this.token, RECAP_TAIL),
+      ]);
+      const currentId = msg && (msg.messageId || msg.eventId);
+      const drop = (list) => (list || []).filter(
+        (m) => !(currentId && (m.messageId === currentId || m.eventId === currentId)),
+      );
+      // Both windows are already chronological — see the note above.
+      const tailAsc = drop(tail);
+      // When the current event carries an id it has already been removed above,
+      // and passing its TEXT as well would make sampleRecap drop any older
+      // message that happens to repeat it verbatim. The text guard is only a
+      // fallback for events with no id.
+      // sampleRecap returns an ARRAY of formatted lines.
+      const lines = sampleRecap(
+        drop(head), tailAsc, currentId ? '' : ((msg && msg.content) || ''),
+      );
+      return (lines || []).join('\n');
+    } catch (e) {
+      this._log(`Could not build channel recap: ${redactDiagnostic(e && e.message)}`);
+      return '';
+    }
+  }
+
+  // ------------------------------------------------------------------
+  // Message handler
+  // ------------------------------------------------------------------
+
+  /**
+   * True when the user stopped this channel since the handler started.
+   *
+   * Called after every await in the pre-spawn window. Each of those awaits —
+   * bootstrap, the status post, the recap fetch, the browser lookup — can take
+   * long enough for a `/stop` to land, and before this existed the stop was
+   * dropped and the task ran to completion anyway.
+   */
+  _cancelled(channel) {
+    return this._cancelledChannels.has(channel);
+  }
+
+  async _handleMessage(msg) {
+    const content = (msg.content || '').trim();
+    const attachments = msg.attachments || [];
+    if (!content && attachments.length === 0) return;
+
+    const msgChannel = msg.sessionId || this.channelName;
+    const sender = msg.senderName || msg.senderType || 'user';
+    this._log(`Processing message from ${sender} in ${msgChannel}: ${content.length} chars`);
+
+    // A stop belongs to the run that was in flight when it was pressed, so the
+    // flag is cleared here — before the first await — and never again.
+    this._cancelledChannels.delete(msgChannel);
+    this._busyChannels.add(msgChannel);
+    try {
+      await this._handleMessageInner(msg, msgChannel, content, attachments);
+    } finally {
+      this._busyChannels.delete(msgChannel);
+    }
+  }
+
+  async _handleMessageInner(msg, msgChannel, content, attachments) {
+
+    if (!this._jsEntry) {
+      const resolved = this._resolveDshCommand();
+      if (resolved) { this._nodeBin = resolved[0]; this._jsEntry = resolved[1]; }
+    }
+    if (!this._jsEntry) {
+      const message = `DeepSeek Harness (dsh) not found — install with: ${dshInstallHint()}`;
+      this._reportStatus(REASON.RUNTIME_MISSING, message);
+      await this.sendError(msgChannel, message);
+      return;
+    }
+
+    await this._autoTitleChannel(msgChannel, content);
+    this._stoppingChannels.delete(msgChannel);
+    if (this._cancelled(msgChannel)) return;
+
+    try {
+      // Returns false when this channel was stopped while waiting; the shared
+      // bootstrap keeps running for whoever else is queued behind it.
+      if (!await this._awaitBootstrap(msgChannel)) return;
+    } catch (e) {
+      if (this._cancelled(msgChannel)) return;
+      const message = `DeepSeek Harness could not initialise its profile: ${redactDiagnostic(e && e.message)}`;
+      this._reportStatus(REASON.SPAWN_FAILED, message);
+      await this.sendError(msgChannel, message);
+      return;
+    }
+    // Bootstrap is the longest pre-spawn wait — a cold first run composes the
+    // whole profile — so it is the likeliest place for a stop to land.
+    if (this._cancelled(msgChannel)) return;
+
+    // No streaming: this is the only progress the workspace will see until the
+    // run finishes.
+    await this.sendStatus(msgChannel, 'DeepSeek Harness is working...');
+    if (this._cancelled(msgChannel)) return;
+
+    let result;
+    try {
+      result = await this._runDsh({ content, attachments, msgChannel, msg });
+    } catch (e) {
+      if (this._isSpawnError(e)) {
+        const { reason, message } = classifySpawnError(e, {
+          label: 'DeepSeek Harness',
+          bin: this._jsEntry,
+        });
+        this._log(message);
+        this._reportStatus(reason, message);
+        await this.sendError(msgChannel, message);
+      } else {
+        const message = `Error processing message: ${redactDiagnostic(e && e.message)}`;
+        this._log(message);
+        await this.sendError(msgChannel, message);
+      }
+      return;
+    }
+
+    if (this._stoppingChannels.has(msgChannel)) {
+      this._stoppingChannels.delete(msgChannel);
+      return;
+    }
+
+    const { text, error } = result;
+    if (error) {
+      await this.sendError(msgChannel, error);
+      return;
+    }
+    this._reportStatus(null);
+    if (text) {
+      await this.sendResponse(msgChannel, text);
+    } else {
+      await this.sendResponse(
+        msgChannel,
+        'DeepSeek Harness finished with no textual output (any file changes were '
+        + 'applied to the workspace directory).',
+      );
+    }
+  }
+
+  // ------------------------------------------------------------------
+  // Subprocess execution
+  // ------------------------------------------------------------------
+
+  async _runDsh({ content, attachments, msgChannel, msg }) {
+    const permissionMode = resolvePermissionMode({
+      workspaceMode: this._mode,
+      configured: this.agentEnv.DSH_PERMISSION_MODE,
+    });
+    if (this.agentEnv.DSH_PERMISSION_MODE
+        && permissionMode !== this.agentEnv.DSH_PERMISSION_MODE
+        && this._mode !== 'plan') {
+      // A typo in a permission setting must not silently downgrade to a
+      // default: the user asked for a specific boundary.
+      return {
+        text: '',
+        error:
+          `Invalid DSH_PERMISSION_MODE "${this.agentEnv.DSH_PERMISSION_MODE}". `
+          + 'Valid values are read-only, workspace-write, danger-full-access.',
+      };
+    }
+
+    // Keep the patch current (model changes) before every run. A failure here
+    // is reported as what it is: an ENOENT from this write would otherwise
+    // reach the caller's spawn-error classifier and be announced as
+    // "executable not found", sending the user after a binary that is present.
+    try {
+      this._writePrivatePatch();
+    } catch (e) {
+      return {
+        text: '',
+        error:
+          'DeepSeek Harness could not write its private configuration '
+          + `(${redactDiagnostic(e && e.message)}). The harness home is `
+          + `${this._dshHome}.`,
+      };
+    }
+
+    const recap = await this._buildRecap(msgChannel, msg);
+    if (this._cancelled(msgChannel)) return { text: '', error: null };
+
+    const attachmentText = formatAttachmentsForPrompt(
+      attachments, 'skills', IS_WINDOWS,
+      { tokenExpr: this._tokenExpr(), endpoint: this.endpoint },
+    ) || '';
+
+    // `_browserEnabledCache` starts null and is populated lazily by this call.
+    // Reading the field directly (as claude.js does, where another code path
+    // has already warmed it) would leave the browser directive permanently off
+    // here, because nothing else in this adapter ever loads it.
+    let browserEnabled = false;
+    try { browserEnabled = await this.getBrowserEnabled(); } catch { /* default off */ }
+    if (this._cancelled(msgChannel)) return { text: '', error: null };
+
+    const taskText = buildDeepSeekTaskFile({
+      agentName: this.agentName,
+      workspaceId: this.workspaceId,
+      channelName: msgChannel,
+      endpoint: this.endpoint,
+      tokenExpr: this._tokenExpr(),
+      mode: this._mode,
+      disabledModules: this.disabledModules,
+      browserEnabled,
+      recap,
+      request: content,
+      attachments: attachmentText,
+    });
+
+    this._taskCounter += 1;
+    const taskFile = path.join(
+      this._tasksDir, `dsh-task-${process.pid}-${this._taskCounter}.md`,
+    );
+
+    try {
+      fs.mkdirSync(this._tasksDir, { recursive: true, mode: 0o700 });
+      // 0o600 is meaningful on POSIX only; on Windows the protection comes from
+      // the per-user profile directory's ACL, not from this bit.
+      fs.writeFileSync(taskFile, taskText, { encoding: 'utf-8', mode: 0o600 });
+    } catch (e) {
+      // FAIL CLOSED. The alternative — passing the prompt on the command line —
+      // would put the workspace token in the process list, which is the exact
+      // exposure the task file exists to avoid.
+      return {
+        text: '',
+        error:
+          'DeepSeek Harness could not stage its task file '
+          + `(${redactDiagnostic(e && e.message)}). The run was not started: the `
+          + 'prompt is never passed on the command line.',
+      };
+    }
+
+    try {
+      // Last gate before anything is started. Everything above this line is
+      // reversible; a process is not.
+      if (this._cancelled(msgChannel)) return { text: '', error: null };
+      return await this._spawnDsh({ taskFile, msgChannel, permissionMode });
+    } finally {
+      try { fs.rmSync(taskFile, { force: true }); } catch { /* best effort */ }
+      this._gcSessions();
+    }
+  }
+
+  _spawnDsh({ taskFile, msgChannel, permissionMode }) {
+    return new Promise((resolve, reject) => {
+      const args = buildHeadlessArgs({
+        jsEntry: this._jsEntry,
+        taskFile,
+        patchFile: this._patchFile,
+      });
+
+      const proc = spawn(this._nodeBin, args, {
+        // stdin is /dev/null: dsh has no stdin task channel, and a prompt that
+        // ever waited on it would hang the run instead of failing.
+        stdio: ['ignore', 'pipe', 'pipe'],
+        env: this._buildSubprocessEnv(permissionMode),
+        cwd: this.workingDir,
+        detached: !IS_WINDOWS,
+        windowsHide: true,
+      });
+      this._channelProcesses[msgChannel] = proc;
+
+      // Log the invocation WITHOUT the task: argv's task element is a constant
+      // sentence, but the file it names holds the user's content.
+      this._log(
+        `Running DeepSeek Harness: --profile headless --patch <private> `
+        + `${path.basename(taskFile)} (${fs.statSync(taskFile).size} bytes, mode=${permissionMode})`,
+      );
+
+      // Buffers, not strings: the caps below are BYTE budgets, and a string's
+      // .length counts UTF-16 code units, which for multi-byte output
+      // understates the memory actually held.
+      const stdoutChunks = [];
+      let stdoutBytes = 0;
+      let stderrTail = Buffer.alloc(0);
+      let overflow = false;
+      let settled = false;
+      let timedOut = false;
+      let timer = null;
+      let exitBackstop = null;
+
+      const timeoutMessage = (suffix) =>
+        `DeepSeek Harness run timed out after ${this._runTimeoutMs / 60000} minutes `
+        + 'and was terminated. The harness does not report progress while a task '
+        + 'is running, so no partial result is available.' + (suffix || '');
+
+      // Four routes can end this run — timeout, stop, spawn error and close.
+      // Exactly one of them may settle the promise.
+      //
+      // Untracking happens HERE and only here, and every route reaches it after
+      // the child is known to be gone (or after the backstop below gives up on
+      // it). Untracking while the child may still be alive would hand the run's
+      // `finally` a green light to garbage-collect sessions under a live
+      // process, and would drop the handle `/stop` needs to try again.
+      const settle = (fn, value) => {
+        if (settled) return;
+        settled = true;
+        if (timer) clearTimeout(timer);
+        if (exitBackstop) clearTimeout(exitBackstop);
+        delete this._channelProcesses[msgChannel];
+        fn(value);
+      };
+
+      proc.stdout.on('data', (d) => {
+        const buf = Buffer.isBuffer(d) ? d : Buffer.from(d);
+        if (stdoutBytes >= MAX_STDOUT_BYTES) { overflow = true; return; }
+        // Trim the chunk to the remaining budget rather than testing before
+        // appending: a single oversized chunk would otherwise sail past the cap.
+        const room = MAX_STDOUT_BYTES - stdoutBytes;
+        if (buf.length > room) {
+          stdoutChunks.push(buf.subarray(0, room));
+          stdoutBytes = MAX_STDOUT_BYTES;
+          overflow = true;
+        } else {
+          stdoutChunks.push(buf);
+          stdoutBytes += buf.length;
+        }
+      });
+      proc.stderr.on('data', (d) => {
+        const buf = Buffer.isBuffer(d) ? d : Buffer.from(d);
+        // A rolling TAIL, so memory is bounded no matter how much a wedged CLI
+        // writes, and the end — where dsh puts the terminal error — survives.
+        stderrTail = Buffer.concat([stderrTail, buf]);
+        if (stderrTail.length > MAX_STDERR_BYTES) {
+          stderrTail = stderrTail.subarray(stderrTail.length - MAX_STDERR_BYTES);
+        }
+      });
+      // A killed child can emit a benign EPIPE/ECONNRESET on its stdio a tick
+      // later; without these listeners that becomes an uncaughtException.
+      proc.stdout.on('error', () => {});
+      proc.stderr.on('error', () => {});
+
+      proc.on('error', (e) => settle(reject, e));
+
+      timer = setTimeout(async () => {
+        timedOut = true;
+        this._log(`DeepSeek Harness exceeded ${this._runTimeoutMs / 60000}min — terminating`);
+        // Terminate first and let 'close' settle, so the child stays tracked
+        // (and therefore stoppable, and visible to the session GC guard) until
+        // it is actually gone. `timedOut` is what stops the close handler from
+        // reporting this as a generic "killed by SIGTERM" failure.
+        await this._stopProcess(proc).catch(() => {});
+        // If even SIGKILL did not produce a 'close', settle anyway rather than
+        // wedging the channel forever — and say plainly that a process may have
+        // survived, instead of implying a clean termination.
+        exitBackstop = setTimeout(() => {
+          this._log('DeepSeek Harness did not exit after SIGKILL — releasing the channel');
+          settle(resolve, {
+            text: '',
+            error: timeoutMessage(
+              ' The process did not exit after being killed and may still be running.',
+            ),
+          });
+        }, STOP_GRACE_MS);
+      }, this._runTimeoutMs);
+
+      // 'close' rather than 'exit': the entire answer is stdout, and 'exit' can
+      // fire before the pipe has been fully drained.
+      proc.on('close', (code, signal) => {
+        const stdout = Buffer.concat(stdoutChunks).toString('utf-8');
+        const stderr = stderrTail.toString('utf-8');
+        if (timedOut) {
+          settle(resolve, { text: '', error: timeoutMessage() });
+          return;
+        }
+        if (this._stoppingChannels.has(msgChannel)) {
+          settle(resolve, { text: '', error: null });
+          return;
+        }
+        if (code === 0) {
+          const { text, truncated } = cleanStdout(stdout);
+          if (truncated || overflow) {
+            this._log('DeepSeek Harness output exceeded the reply cap and was truncated');
+          }
+          settle(resolve, { text, error: null });
+          return;
+        }
+        // A failed run's stdout is NOT an answer: dsh only guarantees the final
+        // assistant text on success.
+        const { category, message } = classifyDshFailure({ code, signal, stderr });
+        this._log(`DeepSeek Harness failed (${category}, exit ${code})`);
+        settle(resolve, {
+          text: '',
+          error: `DeepSeek Harness failed (${category}): ${redactDiagnostic(message, 2000)}`,
+        });
+      });
+    });
+  }
+
+  // ------------------------------------------------------------------
+  // Session garbage collection
+  // ------------------------------------------------------------------
+
+  /**
+   * Trim the private harness home's persisted sessions.
+   *
+   * dsh writes one session per run — `sessions/<workspace-slug>/session-<uuid>/`
+   * holding a compressed jsonl (~20 KB) — and has no resume, so the directory
+   * grows once per workspace message and nothing ever reclaims it.
+   *
+   * Only runs when this adapter has no live child: a session belonging to a run
+   * still in flight is indistinguishable from a stale one by mtime alone.
+   * Nothing outside `sessions/` is touched — profiles, settings.yaml and
+   * .credentials.yaml are not ours to delete.
+   */
+  _gcSessions() {
+    if (Object.keys(this._channelProcesses).length > 0) return;
+    const root = path.join(this._dshHome, 'sessions');
+    try {
+      if (!fs.existsSync(root)) return;
+      for (const scope of fs.readdirSync(root)) {
+        const scopeDir = path.join(root, scope);
+        // Reject symlinks outright rather than following them.
+        if (!fs.lstatSync(scopeDir).isDirectory()) continue;
+
+        const entries = [];
+        for (const name of fs.readdirSync(scopeDir)) {
+          const full = path.join(scopeDir, name);
+          const st = fs.lstatSync(full);
+          if (!st.isDirectory()) continue;
+          entries.push({ name, mtimeMs: st.mtimeMs });
+        }
+
+        for (const name of selectSessionsForGc(entries)) {
+          const victim = path.join(scopeDir, name);
+          // Belt and braces: never delete anything whose real path escaped the
+          // private sessions root.
+          const real = fs.realpathSync(victim);
+          if (!real.startsWith(fs.realpathSync(root) + path.sep)) continue;
+          fs.rmSync(victim, { recursive: true, force: true });
+        }
+      }
+    } catch (e) {
+      // GC is housekeeping: a failure must never affect the reply.
+      this._log(`Session cleanup skipped: ${redactDiagnostic(e && e.message)}`);
+    }
+  }
+}
+
+module.exports = DeepSeekAdapter;