@agentprojectcontext/apx 1.80.2 → 1.89.1

package/src/host/daemon/api/exec.js

@@ -1,54 +1,189 @@
-// Per-agent one-shot / multi-turn LLM endpoints. These do NOT run the
-// super-agent tool loop — they go straight to the model with the agent's
-// system prompt. The super-agent endpoints live in api/super-agent.js.
+// Per-agent LLM endpoints — a project agent's own turn, tools and all.
 //
-//   POST /projects/:pid/agents/:slug/exec     one-shot, no history
-//   POST /projects/:pid/agents/:slug/chat     append to (or start) a conversation
-import { callEngine } from "#core/engines/index.js";
+//   POST /projects/:pid/agents/:slug/exec         one-shot, no history
+//   POST /projects/:pid/agents/:slug/chat         append to (or start) a conversation
+//   POST /projects/:pid/agents/:slug/chat/stream  the same turn, as NDJSON events
+//
+// These are NOT the super-agent (that lives in api/super-agent.js). They run the
+// named agent, on its own system prompt, under its own tool allowlist.
 import { readAgents } from "#core/apc/parser.js";
 import { buildAgentSystem } from "#core/agent/build-agent-system.js";
-import { resolveActiveModel } from "#core/agent/model-router.js";
+import { resolveAgentModel } from "#core/agent/agent-model.js";
+import { runAgentTurn } from "#core/agent/run-turn.js";
+import { floorReplyText } from "#core/agent/closing-floor.js";
+import { createWebConfirmAdapter } from "#core/confirmation/adapters/web.js";
+import { CHANNELS } from "#core/constants/channels.js";
 import {
   startConversation,
   appendTurn,
   readConversation,
   setStatus,
 } from "#core/stores/conversations.js";
-import { asyncRoute } from "./shared.js";
-
-// Pick a model for a direct agent chat: explicit override → agent's own model →
-// super-agent default (resolved via the same router the super-agent uses, so
-// it walks the fallback chain when the primary is empty/unhealthy).
-async function pickAgentModel({ modelOverride, agent, config }) {
-  if (modelOverride) return modelOverride;
-  if (agent.fields?.Model) return agent.fields.Model;
-  try {
-    const routing = await resolveActiveModel(config);
-    return routing?.modelId || null;
-  } catch {
+import { buildTurnAttribution, appendAgentReplyToConversation } from "#core/stores/turn-record.js";
+import { answerDeliveries } from "#core/stores/deliveries.js";
+import { attachmentsMeta } from "#core/stores/media-archive.js";
+import { asyncRoute, rejectA2AWrite} from "./shared.js";
+import { readTurnAttachments } from "./media.js";
+import { broadcastTurn } from "../events-ws.js";
+import { startActiveTurn, appendActiveTurn, endActiveTurn, convTurnKey } from "../active-turns.js";
+import { wasAborted, abortedTurnEvent } from "./turn-abort.js";
+
+// How long a streamed turn may go silent before it writes a keepalive byte.
+// Matches api/super-agent.js — comfortably under undici's body timeout.
+const KEEPALIVE_MS = 20_000;
+
+
+/** Resolve the agent and its model, or answer the request and return null. */
+async function resolveTarget(req, res, p, config) {
+  const agents = readAgents(p.path);
+  const agent = agents.find((a) => a.slug === req.params.slug);
+  if (!agent) {
+    res.status(404).json({ error: "agent not found" });
+    return null;
+  }
+  const modelId = await resolveAgentModel({
+    agent,
+    config,
+    override: req.body?.model,
+  });
+  if (!modelId) {
+    res.status(400).json({ error: "agent has no model and none provided" });
     return null;
   }
+  return { agent, modelId };
+}
+
+/**
+ * The conversation this turn belongs to: the one named, or a fresh one.
+ * Returns the path, the id, the turns to replay, and any compacted summary.
+ */
+function openConversation({ p, agent, modelId, system, conversationId, channel }) {
+  if (!conversationId) {
+    const conv = startConversation({
+      storagePath: p.storagePath,
+      agentSlug: agent.slug,
+      engine: modelId,
+      system,
+      channel,
+    });
+    return { path: conv.path, id: conv.id, history: [], compactSummary: null };
+  }
+  const existing = readConversation(p.storagePath, agent.slug, conversationId);
+  if (!existing) return null;
+  // Inject compact summary into system instead of replaying it as a turn.
+  const compactTurn = existing.turns.find((t) => t.role === "compact");
+  const compactSummary = compactTurn
+    ? compactTurn.content.replace(/^\[Compacted \d+ turns.*?\]\n\n?/, "").trim()
+    : null;
+  return {
+    path: existing.path,
+    id: conversationId,
+    history: existing.turns
+      .filter((t) => t.role === "user" || t.role === "assistant")
+      .map((t) => ({ role: t.role, content: t.content })),
+    compactSummary,
+  };
+}
+
+/** The attribution a reopened conversation renders from. */
+function turnAttribution(agent, result) {
+  return buildTurnAttribution({
+    agentSlug: agent.slug,
+    agentName: agent.fields?.Name || agent.slug,
+    model: result.model,
+    usage: result.usage,
+    trace: result.trace,
+  });
+}
+
+/**
+ * The never-silent floor, in the agent's own voice.
+ *
+ * runAgent re-prompts a dud turn and then gives up, and what came back went
+ * straight into the thread: an empty bubble in the panel and an empty assistant
+ * row on disk, which the next turn reads back as the answer this one gave. The
+ * closing is asked of the model first — the AGENT's model, because this thread
+ * is in its voice and not the super-agent's — and the canned line goes out only
+ * if that comes back empty too.
+ *
+ * A turn with a real answer is returned untouched and costs no model call.
+ * Deliberately NOT used on the abort path: an interrupted turn that wrote
+ * nothing is meant to leave no bubble at all.
+ *
+ * @returns {object} the same result, with `text` guaranteed non-empty
+ */
+async function withClosingFloor({ p, agent, modelId, config, result, streamedText = "" }) {
+  const closing = await floorReplyText({
+    globalConfig: p.config || config,
+    model: result.model || modelId,
+    text: result.text,
+    streamedText,
+    trace: result.trace,
+  });
+  if (!closing.floored) return result;
+  // eslint-disable-next-line no-console
+  console.warn(
+    `[apx] ${agent.slug}: empty turn closed by the floor — ` +
+    (closing.authored
+      ? "the model wrote the closing"
+      : "the canned floor spoke, the model could not write it either")
+  );
+  return { ...result, text: closing.text };
+}
+
+/** Collect the last thing the turn actually SAID. The result carries the
+ *  closing, so this is the one piece of context a floored closing needs. */
+function lastSaidCollector() {
+  const seen = { text: "" };
+  return [seen, (ev) => {
+    if (ev?.type === "assistant_text" && String(ev.text || "").trim()) seen.text = String(ev.text).trim();
+  }];
 }
 
-export function register(api, { projects, project, config }) {
-  api.post("/projects/:pid/agents/:slug/exec", async (req, res) => {
+function persistAgentReply({ filePath, agent, result }) {
+  // Images the agent attached to THIS reply (attach_media). Archived into
+  // ~/.apx/media on the way, because a skill's picture lives beside its
+  // SKILL.md and the media endpoint serves nothing from outside the media dir —
+  // the row would name a file the viewer is not allowed to fetch.
+  const attribution = { ...turnAttribution(agent, result), ...attachmentsMeta(result.media) };
+  appendAgentReplyToConversation({
+    filePath,
+    reply: result.text,
+    trace: result.trace,
+    attribution,
+  });
+  return attribution;
+}
+
+export function register(api, { projects, project, config, plugins, registries }) {
+  api.post("/projects/:pid/agents/:slug/exec", asyncRoute(async (req, res) => {
     const p = project(req, res);
     if (!p) return;
+    // An a2a thread is a transcript, not an addressable agent.
+    if (rejectA2AWrite(req, res, "written to")) return;
     const {
       prompt,
-      model: modelOverride,
       temperature,
       maxTokens,
+      tools,
+      maxIters,
+      channel,
+      channelMeta,
+      attachments,
     } = req.body || {};
-    if (!prompt) return res.status(400).json({ error: "prompt required" });
-    const agents = readAgents(p.path);
-    const agent = agents.find((a) => a.slug === req.params.slug);
-    if (!agent) return res.status(404).json({ error: "agent not found" });
-    const modelId = await pickAgentModel({ modelOverride, agent, config });
-    if (!modelId)
-      return res
-        .status(400)
-        .json({ error: "agent has no model and none provided" });
+    // Files the composer uploaded to ~/.apx/media and named on this turn,
+    // resolved the way the super-agent turn resolves them: images ride on the
+    // model message, and a marker naming each file is folded into the prompt so
+    // a photo with no caption is still a turn and a non-vision engine is told a
+    // file arrived and where it lives.
+    const turnFiles = readTurnAttachments(attachments);
+    if (!prompt && !turnFiles.markers.length) {
+      return res.status(400).json({ error: "prompt required" });
+    }
+    const turnPrompt = [...turnFiles.markers, prompt].filter(Boolean).join(" ");
+    const target = await resolveTarget(req, res, p, config);
+    if (!target) return;
+    const { agent, modelId } = target;
 
     try {
       const system = buildAgentSystem(p, agent, { invocation: "engine" });
@@ -58,22 +193,21 @@ export function register(api, { projects, project, config }) {
         engine: modelId,
         system,
       });
-      appendTurn({ filePath: conv.path, role: "user", content: prompt });
+      appendTurn({ filePath: conv.path, role: "user", content: turnPrompt, meta: turnFiles.media || undefined });
 
-      const result = await callEngine({
-        modelId,
-        system,
-        messages: [{ role: "user", content: prompt }],
-        config: p.config || config,
-        temperature,
-        maxTokens,
+      const [said, observe] = lastSaidCollector();
+      const raw = await runAgentTurn({
+        p, agent, modelId, system, prompt: turnPrompt,
+        attachments: turnFiles.attachments,
+        channel: channel || CHANNELS.API,
+        channelMeta,
+        temperature, maxTokens, tools, maxIters,
+        projects, plugins, registries, config,
+        onEvent: observe,
       });
+      const result = await withClosingFloor({ p, agent, modelId, config, result: raw, streamedText: said.text });
 
-      appendTurn({
-        filePath: conv.path,
-        role: "assistant",
-        content: result.text,
-      });
+      const attribution = persistAgentReply({ filePath: conv.path, agent, result });
       setStatus(conv.path, "closed");
 
       p.logMessage({
@@ -93,7 +227,7 @@ export function register(api, { projects, project, config }) {
         actor_kind: "agent",
         author: agent.slug,
         body: result.text,
-        meta: { conversation: conv.id, usage: result.usage },
+        meta: { conversation: conv.id, ...attribution },
       });
 
       projects.rebuild(p.id);
@@ -101,111 +235,293 @@ export function register(api, { projects, project, config }) {
         conversation: { id: conv.id, filename: conv.filename, path: conv.path },
         text: result.text,
         usage: result.usage,
-        engine: modelId,
+        engine: result.model,
+        trace: result.trace,
+        allowed_tools: result.allowedTools,
       });
     } catch (e) {
       res.status(500).json({ error: e.message });
     }
-  });
+  }));
 
   api.post("/projects/:pid/agents/:slug/chat", asyncRoute(async (req, res) => {
     const p = project(req, res);
     if (!p) return;
+    // An a2a thread is a transcript, not an addressable agent.
+    if (rejectA2AWrite(req, res, "written to")) return;
     const {
       prompt,
       conversation_id,
-      model: modelOverride,
       temperature,
       maxTokens,
+      tools,
+      maxIters,
       channel,
+      channelMeta,
+      attachments,
     } = req.body || {};
-    if (!prompt) return res.status(400).json({ error: "prompt required" });
-    const agents = readAgents(p.path);
-    const agent = agents.find((a) => a.slug === req.params.slug);
-    if (!agent) return res.status(404).json({ error: "agent not found" });
-    const modelId = await pickAgentModel({ modelOverride, agent, config });
-    if (!modelId)
-      return res
-        .status(400)
-        .json({ error: "agent has no model and none provided" });
+    // Files the composer uploaded to ~/.apx/media and named on this turn,
+    // resolved the way the super-agent turn resolves them: images ride on the
+    // model message, and a marker naming each file is folded into the prompt so
+    // a photo with no caption is still a turn and a non-vision engine is told a
+    // file arrived and where it lives.
+    const turnFiles = readTurnAttachments(attachments);
+    if (!prompt && !turnFiles.markers.length) {
+      return res.status(400).json({ error: "prompt required" });
+    }
+    const turnPrompt = [...turnFiles.markers, prompt].filter(Boolean).join(" ");
+    const target = await resolveTarget(req, res, p, config);
+    if (!target) return;
+    const { agent, modelId } = target;
 
     try {
-      let convPath;
-      let convId;
-      let history = [];
-      let compactSummary = null;
-
-      if (conversation_id) {
-        const existing = readConversation(
-          p.storagePath,
-          agent.slug,
-          conversation_id
-        );
-        if (!existing)
-          return res
-            .status(404)
-            .json({ error: `conversation ${conversation_id} not found` });
-        convPath = existing.path;
-        convId = conversation_id;
-        // Inject compact summary into system instead of replaying it as a turn.
-        const compactTurn = existing.turns.find((t) => t.role === "compact");
-        if (compactTurn) {
-          compactSummary = compactTurn.content
-            .replace(/^\[Compacted \d+ turns.*?\]\n\n?/, "")
-            .trim();
-        }
-        history = existing.turns
-          .filter((t) => t.role === "user" || t.role === "assistant")
-          .map((t) => ({ role: t.role, content: t.content }));
+      const turn = prepareChatTurn({ p, agent, modelId, conversation_id, channel });
+      if (!turn) {
+        return res
+          .status(404)
+          .json({ error: `conversation ${conversation_id} not found` });
       }
 
-      const extraParts = compactSummary
-        ? [`## Previous Conversation Context (Compacted)\n${compactSummary}`]
-        : [];
-      const system = buildAgentSystem(p, agent, {
-        invocation: "engine",
-        extraParts,
-      });
+      appendTurn({ filePath: turn.conv.path, role: "user", content: turnPrompt, meta: turnFiles.media || undefined });
+      // Manu replied in this agent's chat → close its open deliveries (and cancel
+      // any grace-window notify still pending). See core/stores/deliveries.js.
+      try { answerDeliveries(p.storagePath, agent.slug); } catch { /* best-effort */ }
 
-      if (!conversation_id) {
-        const conv = startConversation({
-          storagePath: p.storagePath,
-          agentSlug: agent.slug,
-          engine: modelId,
-          system,
-          channel,
-        });
-        convPath = conv.path;
-        convId = conv.id;
-      }
-
-      appendTurn({ filePath: convPath, role: "user", content: prompt });
-      history.push({ role: "user", content: prompt });
-
-      const result = await callEngine({
-        modelId,
-        system,
-        messages: history,
-        config: p.config || config,
-        temperature,
-        maxTokens,
-      });
-      appendTurn({
-        filePath: convPath,
-        role: "assistant",
-        content: result.text,
+      const [said, observe] = lastSaidCollector();
+      const raw = await runAgentTurn({
+        p, agent, modelId,
+        system: turn.system,
+        prompt: turnPrompt,
+        attachments: turnFiles.attachments,
+        previousMessages: turn.conv.history,
+        channel: channel || CHANNELS.API,
+        channelMeta,
+        temperature, maxTokens, tools, maxIters,
+        projects, plugins, registries, config,
+        onEvent: observe,
       });
+      const result = await withClosingFloor({ p, agent, modelId, config, result: raw, streamedText: said.text });
+
+      persistAgentReply({ filePath: turn.conv.path, agent, result });
       projects.rebuild(p.id);
 
       res.json({
-        conversation_id: convId,
+        conversation_id: turn.conv.id,
         text: result.text,
         usage: result.usage,
-        engine: modelId,
-        compacted: !!compactSummary,
+        engine: result.model,
+        trace: result.trace,
+        allowed_tools: result.allowedTools,
+        compacted: !!turn.conv.compactSummary,
       });
     } catch (e) {
       res.status(500).json({ error: e.message });
     }
   }));
+
+  // The same turn, streamed. Identical event vocabulary to the super-agent's
+  // /chat/stream, so a client that already renders tool progress for Roby needs
+  // no second reader for Magui.
+  api.post("/projects/:pid/agents/:slug/chat/stream", asyncRoute(async (req, res) => {
+    const p = project(req, res);
+    if (!p) return;
+    // An a2a thread is a transcript, not an addressable agent.
+    if (rejectA2AWrite(req, res, "written to")) return;
+    const {
+      prompt,
+      conversation_id,
+      temperature,
+      maxTokens,
+      tools,
+      maxIters,
+      channel,
+      channelMeta,
+      attachments,
+    } = req.body || {};
+    // Files the composer uploaded to ~/.apx/media and named on this turn,
+    // resolved the way the super-agent turn resolves them: images ride on the
+    // model message, and a marker naming each file is folded into the prompt so
+    // a photo with no caption is still a turn and a non-vision engine is told a
+    // file arrived and where it lives.
+    const turnFiles = readTurnAttachments(attachments);
+    if (!prompt && !turnFiles.markers.length) {
+      return res.status(400).json({ error: "prompt required" });
+    }
+    const turnPrompt = [...turnFiles.markers, prompt].filter(Boolean).join(" ");
+    const target = await resolveTarget(req, res, p, config);
+    if (!target) return;
+    const { agent, modelId } = target;
+
+    const turn = prepareChatTurn({ p, agent, modelId, conversation_id, channel });
+    if (!turn) {
+      return res
+        .status(404)
+        .json({ error: `conversation ${conversation_id} not found` });
+    }
+
+    res.setHeader("content-type", "application/x-ndjson; charset=utf-8");
+    res.setHeader("cache-control", "no-cache, no-transform");
+    res.setHeader("x-accel-buffering", "no");
+    res.flushHeaders?.();
+
+    let lastWriteAt = Date.now();
+    const send = (event) => {
+      lastWriteAt = Date.now();
+      res.write(JSON.stringify(event) + "\n");
+    };
+    // A bare newline is a valid no-op in NDJSON: one tool can own the turn for
+    // minutes, and to an HTTP client that is an idle body undici drops.
+    const keepalive = setInterval(() => {
+      if (Date.now() - lastWriteAt < KEEPALIVE_MS) return;
+      lastWriteAt = Date.now();
+      try { res.write("\n"); } catch { /* the socket is gone; close clears us */ }
+    }, KEEPALIVE_MS);
+    keepalive.unref?.();
+    res.on("close", () => clearInterval(keepalive));
+
+    appendTurn({ filePath: turn.conv.path, role: "user", content: turnPrompt, meta: turnFiles.media || undefined });
+    // Manu replied in this agent's chat → close its open deliveries (and cancel
+    // any grace-window notify still pending). See core/stores/deliveries.js.
+    try { answerDeliveries(p.storagePath, agent.slug); } catch { /* best-effort */ }
+
+    // Register the turn and push it over the shared feed too, so a surface that
+    // did NOT open this stream (another tab, or this one after a refresh) can
+    // catch up on the partial and follow the tokens live. The turn keeps running
+    // here regardless of whether this NDJSON socket stays open.
+    const turnKey = convTurnKey(p.id, turn.conv.id);
+    // The run's kill switch. Closing this socket deliberately does NOT stop the
+    // turn — that is what lets another tab catch up — so stopping it has to be
+    // asked for out loud, via POST .../turns/abort, which reaches `abort` here.
+    const turnAbort = new AbortController();
+    const active = startActiveTurn(turnKey, {
+      project_id: p.id, agent_slug: agent.slug, conversation_id: turn.conv.id, model: modelId,
+      abort: () => turnAbort.abort(),
+    });
+    // The steps as they happen. runAgentTurn throws on abort and its trace goes
+    // with it, so an interrupted turn would otherwise be persisted as prose with
+    // the work erased — and the turn that continues it would not know which
+    // tools had already run for real.
+    const partialTrace = [];
+    const turnFrame = (phase, extra) => broadcastTurn({
+      phase, project_id: p.id, agent_slug: agent.slug, conversation_id: turn.conv.id,
+      turn_id: active.id, ...extra,
+    });
+    turnFrame("start");
+    // The conversation's identity, up front. It used to ride only on `final`,
+    // which meant the first turn of a new chat could not be addressed until it
+    // was over — no way to stop it, and nothing for another tab to follow.
+    send({
+      type: "start",
+      conversation_id: turn.conv.id,
+      turn_id: active.id,
+      agent_slug: agent.slug,
+      model: modelId,
+    });
+
+    const [said, observeSaid] = lastSaidCollector();
+    try {
+      const raw = await runAgentTurn({
+        p, agent, modelId,
+        system: turn.system,
+        prompt: turnPrompt,
+        attachments: turnFiles.attachments,
+        previousMessages: turn.conv.history,
+        // `api`, the same default as the sibling /exec and /chat routes and as
+        // the super-agent's own streamed turn (resolveSuperAgentContext). The
+        // channel is whatever the CLIENT says it is: the web panel always sends
+        // `channel: "web"`, so it keeps the web chat's behaviour, and nothing
+        // here has to guess. This route used to default to CHANNELS.WEB, which
+        // was invisible only because the panel was its one caller — the next
+        // client to omit `channel` would have silently inherited the web chat's
+        // run-to-completion tool budget (channelToolIters) instead of the
+        // bounded conversational one, on a surface nobody was watching.
+        channel: channel || CHANNELS.API,
+        channelMeta,
+        temperature, maxTokens, tools, maxIters,
+        projects, plugins, registries, config,
+        signal: turnAbort.signal,
+        onEvent: (ev) => {
+          if (ev?.type === "tool_result" && ev.trace) partialTrace.push(ev.trace);
+          observeSaid(ev);
+          send(ev);
+        },
+        onToken: (chunk) => { send({ type: "assistant_delta", delta: chunk }); appendActiveTurn(active.id, chunk); turnFrame("delta", { delta: chunk }); },
+        // A streamed turn CAN answer a confirmation round-trip (the client
+        // posts to /super-agent/confirm/:id). A caller that cannot — `apx exec`
+        // renders a spinner and nothing else — sends confirm:false and falls
+        // back to the configured policy, same as the blocking endpoint.
+        requestConfirmation:
+          req.body?.confirm === false ? null : createWebConfirmAdapter({ onEvent: send }),
+      });
+
+      const result = await withClosingFloor({ p, agent, modelId, config, result: raw, streamedText: said.text });
+      persistAgentReply({ filePath: turn.conv.path, agent, result });
+      projects.rebuild(p.id);
+
+      const finalResult = {
+        conversation_id: turn.conv.id,
+        text: result.text,
+        usage: result.usage,
+        name: agent.slug,
+        model: result.model,
+        trace: result.trace,
+        allowed_tools: result.allowedTools,
+        compacted: !!turn.conv.compactSummary,
+      };
+      turnFrame("final", { result: finalResult });
+      send({ type: "final", result: finalResult });
+      clearInterval(keepalive);
+      res.end();
+    } catch (e) {
+      // Interrupted, not broken. Whatever streamed is real work the user saw, so
+      // it stays in the thread — the message that interrupted this turn opens
+      // the next one, and reads what got done here as its history. Same contract
+      // Telegram has: "whatever streamed so far is already sent + logged; the
+      // newer message's run continues the thread."
+      if (wasAborted(e, turnAbort)) {
+        const partial = (active.text || "").trim();
+        if (partial || partialTrace.length) {
+          persistAgentReply({
+            filePath: turn.conv.path,
+            agent,
+            result: { text: partial, trace: partialTrace, model: modelId, media: [] },
+          });
+          projects.rebuild(p.id);
+        }
+        const ev = abortedTurnEvent({ text: partial, trace: partialTrace });
+        turnFrame("aborted", ev);
+        clearInterval(keepalive);
+        send(ev);
+        res.end();
+        return;
+      }
+      turnFrame("error", { error: e.message });
+      clearInterval(keepalive);
+      send({ type: "error", error: e.message });
+      res.end();
+    } finally {
+      endActiveTurn(active.id);
+    }
+  }));
+
+  /** Everything both chat endpoints need before the model is called. */
+  function prepareChatTurn({ p, agent, modelId, conversation_id, channel }) {
+    // The system prompt has to exist before the conversation file that records
+    // it, and the compacted summary has to be inside it — so the file is opened
+    // first when it already exists, and created after when it does not.
+    const existing = conversation_id
+      ? openConversation({ p, agent, modelId, system: "", conversationId: conversation_id, channel })
+      : null;
+    if (conversation_id && !existing) return null;
+
+    const extraParts = existing?.compactSummary
+      ? [`## Previous Conversation Context (Compacted)\n${existing.compactSummary}`]
+      : [];
+    const system = buildAgentSystem(p, agent, { invocation: "engine", extraParts });
+
+    const conv =
+      existing ||
+      openConversation({ p, agent, modelId, system, conversationId: null, channel });
+    return { system, conv };
+  }
 }