hilos-agent 0.9.4 → 0.10.0

package/src/handler.mjs

@@ -41,6 +41,7 @@ import { makeStreamParser, createUsageFold } from "./agent-events.mjs";
 import {
   detectVendor,
   codeStreamArgs,
+  codeWebArgs,
   codeDirArgs,
   codeImageArgs,
   codeProjectKey,
@@ -82,6 +83,7 @@ import {
   shouldGateCodexPermissions,
 } from "./codex-mcp-session.mjs";
 import { createUngatedRunNotice } from "./permission-gate.mjs";
+import { webMcpAgentPrompt } from "./webmcp-bridge.mjs";
 
 /**
  * The environment for a coding/chat CLI run. runCli always strips HILOS_* on top
@@ -306,6 +308,11 @@ function openCodePermissionCallbacks({
   threadRoot,
   runId = null,
   provider = "opencode",
+  // 0866 — when the server's get_permission_decision supports a long-poll
+  // hold, each read blocks up to this many ms and answers the moment a human
+  // decides, instead of the gate discovering the decision on its next 1s poll.
+  // 0 (older servers) keeps the plain immediate read.
+  decisionWaitMs = 0,
 }) {
   return {
     requestPermission: async (request) => {
@@ -352,6 +359,8 @@ function openCodePermissionCallbacks({
         vendorSessionId: request.sessionId,
         vendorRequestId: request.vendorRequestId,
         ...(failClosed ? { failClosed: true } : {}),
+        // Never on a failClosed settlement — that call exists to act NOW.
+        ...(decisionWaitMs > 0 && !failClosed ? { waitMs: decisionWaitMs } : {}),
       });
     },
   };
@@ -768,7 +777,37 @@ async function linkPrFromUrl({ tool, channelId, url }) {
   await tool("link_pr", { channelId, repoFullName: m[1], prNumber: Number(m[2]) }).catch(() => {});
 }
 
-async function applyDecision({ decision, repoPath, branch, task, requester, cfg, tool, channelId, deps, parentId, existingPrUrl, settleId, runId, usageArgs = {} }) {
+function githubProvenance({ cfg, me, message, channelId, runId, summary }) {
+  let siteUrl = "https://hilos.sh";
+  try {
+    const parsed = new URL(cfg.url);
+    parsed.pathname = parsed.pathname.replace(/\/api\/mcp\/?$/, "");
+    parsed.search = "";
+    parsed.hash = "";
+    siteUrl = parsed.toString().replace(/\/$/, "");
+  } catch {
+    /* the body helper will keep links absent if configuration is malformed */
+  }
+  const roomUrl = `${siteUrl}/w/${me?.workspaceId}/c/${channelId}`;
+  const messageUrl = message?.id
+    ? message.parentId
+      ? `${roomUrl}?thread=${message.parentId}&reply=${message.id}`
+      : `${roomUrl}#m-${message.id}`
+    : undefined;
+  return {
+    agentName: me?.agentName,
+    agentId: me?.agentId,
+    personName: message?.author,
+    roomName: message?.channel || "project room",
+    roomUrl,
+    messageId: message?.id,
+    messageUrl,
+    runId,
+    summary,
+  };
+}
+
+async function applyDecision({ decision, repoPath, branch, task, requester, cfg, tool, channelId, deps, parentId, existingPrUrl, settleId, runId, provenance = {}, usageArgs = {} }) {
   const tag = requesterTag(requester);
   const lead = tag ? `${tag} — ` : "";
   // `parentId` here is the run's thread root. Terminal outcomes ask the server
@@ -785,7 +824,7 @@ async function applyDecision({ decision, repoPath, branch, task, requester, cfg,
       });
       return { status: "nothing-staged", branch };
     }
-    const commit = deps.git(repoPath, ["commit", "-m", commitMessage(task)]);
+    const commit = deps.git(repoPath, ["commit", "-m", commitMessage(task, provenance)]);
     if (commit.status !== 0) {
       await tool("post_message", {
         channelId,
@@ -805,7 +844,7 @@ async function applyDecision({ decision, repoPath, branch, task, requester, cfg,
       });
       return { status: "push-failed", branch };
     }
-    const { title, body } = prTitleBody(task, branch);
+    const { title, body } = prTitleBody(task, branch, provenance);
     // Continuing an existing PR (request-changes rework): the push already
     // updated it — reuse its URL instead of opening a duplicate. Otherwise open
     // a fresh PR.
@@ -897,7 +936,7 @@ async function applyDecision({ decision, repoPath, branch, task, requester, cfg,
  * switched to `main` would otherwise make `gh` try to open main → main. Used
  * only when NOT gated (bias-to-action).
  */
-async function shipSelfDriven({ repoPath, branch, task, requester, cfg, tool, channelId, deps, parentId, settleId, runId, usageArgs = {} }) {
+async function shipSelfDriven({ repoPath, branch, task, requester, cfg, tool, channelId, deps, parentId, settleId, runId, provenance = {}, usageArgs = {} }) {
   const tag = requesterTag(requester);
   const lead = tag ? `${tag} — ` : "";
   const currentBranch =
@@ -907,7 +946,7 @@ async function shipSelfDriven({ repoPath, branch, task, requester, cfg, tool, ch
   let prUrl = deps.findPR ? deps.findPR(repoPath, ship.headBranch) : null;
   let prAttempt = null;
   if (!prUrl && push.status === 0) {
-    const { title, body } = prTitleBody(task, ship.headBranch);
+    const { title, body } = prTitleBody(task, ship.headBranch, provenance);
     prAttempt = deps.openPR(repoPath, {
       title,
       body,
@@ -916,7 +955,7 @@ async function shipSelfDriven({ repoPath, branch, task, requester, cfg, tool, ch
     });
     prUrl = prAttempt.ok && prAttempt.url ? prAttempt.url : null;
   }
-  const { title } = prTitleBody(task, ship.headBranch);
+  const { title } = prTitleBody(task, ship.headBranch, provenance);
   const recoveredNote = ship.recoveredFrom
     ? ` The child had switched to \`${ship.recoveredFrom}\`; hilos recovered its HEAD onto \`${ship.headBranch}\`.`
     : "";
@@ -991,6 +1030,7 @@ async function shipSelfDriven({ repoPath, branch, task, requester, cfg, tool, ch
 // Sentinel the router model emits when the latest message is a request to change
 // code. Unusual on purpose so it can't be confused with a real chat reply.
 const CODE_SIGNAL = "__CODE__";
+const CHAT_SIGNAL = "__CHAT__";
 
 // 0860 — the quiet option, offered ONLY on untagged DM wakes (message.implicit
 // === "dm"). In a DM every message reaches the agent without a tag, so some of
@@ -1046,8 +1086,8 @@ export function dmJudgmentBlock(implicitDm) {
 
 /**
  * Decide — with the LLM, not a word list — whether the latest message wants a
- * code change or a conversational reply, and produce the payload in the SAME
- * call. The model reads the whole conversation, so it judges by intent and
+ * code change or a conversational reply. The model reads the whole
+ * conversation, so it judges by intent and
  * context, in any language: "just code it", "dale, hazlo", "yeah go for it" after
  * a request → code; "how does this work?", "thoughts?", "thanks" → chat.
  *
@@ -1059,12 +1099,12 @@ export function dmJudgmentBlock(implicitDm) {
  * Returns one of:
  *   { aborted: true }
  *   { code: true, task }           → run the coding flow with `task` as the spec
- *   { code: false, reply, error }  → post `reply` as a chat message
+ *   { code: false, error }         → run the separate conversational responder
  *
  * `error` is set when the model produced nothing (so the caller can be honest
  * about a timeout vs a missing binary instead of inventing a reply).
  */
-async function routeIntent({ name, repoFullName, transcript, workspaceMemory, cfg, signal, hasActiveRun = false, runCliFn, implicitDm = false }) {
+export async function routeIntent({ name, repoFullName, transcript, workspaceMemory, cfg, signal, hasActiveRun = false, runCliFn, implicitDm = false }) {
   const doRun = runCliFn || runCli; // folder mode injects deps.runCli; repo flow uses the import
   const cmd = chatCmdFor(cfg);
   const parts = cmd.split(" ").filter(Boolean);
@@ -1093,12 +1133,12 @@ async function routeIntent({ name, repoFullName, transcript, workspaceMemory, cf
     `using the conversation. Example:\n${CODE_SIGNAL}\nAdd a hover popover to message reactions ` +
     `that lists who reacted with each emoji.\n\n` +
     `Otherwise — a question, a greeting, general discussion, or they explicitly don't want code ` +
-    `yet — just reply to them concisely and directly as a single chat message (no preamble, no ` +
-    `headings). When unsure, stay conversational; implementation and PR flow are for clear ` +
-    `action language.` +
+    `yet — output ONLY the token ${CHAT_SIGNAL}. When unsure, choose ${CHAT_SIGNAL}; ` +
+    `implementation and PR flow are for clear action language.` +
     `${followupBlock}\n\n` +
-    `You are only routing here — output ONLY your text response. Do NOT use any tools, do NOT ` +
-    `edit files, do NOT run commands; a separate step does the actual coding.` +
+    `You are only routing here — output ONLY the routing token (plus the imperative spec for ` +
+    `${CODE_SIGNAL}). Do NOT use any tools, edit files, or run commands; separate tool-capable ` +
+    `steps write the conversational answer or do the coding.` +
     `${dmJudgmentBlock(implicitDm)}\n\n` +
     `${memoryPreamble(workspaceMemory)}Conversation so far:\n${transcript}`;
   const run = await doRun({
@@ -1134,7 +1174,10 @@ async function routeIntent({ name, repoFullName, transcript, workspaceMemory, cf
     const quiet = parseNoReply(out);
     if (quiet.noReply) return { code: false, noReply: true, emoji: quiet.emoji, followupSignal };
   }
-  return { code: false, reply: out, followupSignal };
+  // A classifier never authors the final answer. Even if a model ignores the
+  // token contract and emits prose, discard it and let the normal responder —
+  // with the vendor's real tools — answer the room.
+  return { code: false, reply: null, followupSignal };
 }
 
 /**
@@ -1221,6 +1264,11 @@ function memoryPreamble(workspaceMemory) {
   return m ? `Workspace context (the project's soul):\n${m}\n\n` : "";
 }
 
+/** Shared context for a model that may act, not the read-only intent router. */
+function agentPreamble(workspaceMemory, cfg) {
+  return memoryPreamble(workspaceMemory) + webMcpAgentPrompt(cfg);
+}
+
 /**
  * Recent conversation the agent should reply within: the thread it was pinged in
  * (threads are where conversations live), else the channel tail. Returns the raw
@@ -1326,7 +1374,7 @@ async function respondConversationally({ message, channelId, tool, me, cfg, repo
     `concisely and directly as a single chat message — no preamble, no headings. ` +
     `${repoLine}` +
     `${dmJudgmentBlock(implicitDm)}\n\n` +
-    `${memoryPreamble(workspaceMemory)}` +
+    `${agentPreamble(workspaceMemory, cfg)}` +
     `Conversation so far:\n${transcript}`;
 
   // Chat uses the FAST one-shot command (the coding vendor's own print mode when
@@ -2048,8 +2096,10 @@ async function handleFolderTask({ message, channelId, tool, me, caps, cfg, deps,
     // 0779: [] for every vendor without a verified image flag — their argv is
     // byte-identical to before, and the prompt note still names the files.
     const imageArgs = codeImageArgs(vendor, localImages);
+    const webArgs = codeWebArgs(vendor, parts.slice(1));
     const codeArgs = [
       ...parts.slice(1),
+      ...webArgs,
       ...modelArgs,
       ...dirArgs,
       ...imageArgs,
@@ -2131,7 +2181,7 @@ async function handleFolderTask({ message, channelId, tool, me, caps, cfg, deps,
           cmd: parts[0],
           vendor,
           cwd: folderPath,
-          prompt: memoryPreamble(workspaceMemory) + promptText,
+          prompt: agentPreamble(workspaceMemory, cfg) + promptText,
           timeoutMs: cfg.runTimeoutMs,
           signal,
           env: scrubHilosEnv(codingChildEnv(cfg) || process.env),
@@ -2141,6 +2191,7 @@ async function handleFolderTask({ message, channelId, tool, me, caps, cfg, deps,
             tool,
             channelId,
             threadRoot,
+            decisionWaitMs: caps.decisionWaitMs,
             provider: vendor,
           }),
         });
@@ -2149,7 +2200,7 @@ async function handleFolderTask({ message, channelId, tool, me, caps, cfg, deps,
           cmd: parts[0],
           args: codeArgs,
           cwd: folderPath,
-          prompt: memoryPreamble(workspaceMemory) + promptText,
+          prompt: agentPreamble(workspaceMemory, cfg) + promptText,
           timeoutMs: cfg.runTimeoutMs,
           signal,
           env: scrubHilosEnv(codingChildEnv(cfg) || process.env),
@@ -2158,6 +2209,7 @@ async function handleFolderTask({ message, channelId, tool, me, caps, cfg, deps,
             tool,
             channelId,
             threadRoot,
+            decisionWaitMs: caps.decisionWaitMs,
           }),
         });
       } else if (gateCodexPermissions) {
@@ -2168,7 +2220,7 @@ async function handleFolderTask({ message, channelId, tool, me, caps, cfg, deps,
           cmd: parts[0],
           codeArgs,
           cwd: folderPath,
-          prompt: memoryPreamble(workspaceMemory) + promptText,
+          prompt: agentPreamble(workspaceMemory, cfg) + promptText,
           model: resolvedModelId,
           signal,
           onData: handleCliData,
@@ -2177,6 +2229,7 @@ async function handleFolderTask({ message, channelId, tool, me, caps, cfg, deps,
             tool,
             channelId,
             threadRoot,
+            decisionWaitMs: caps.decisionWaitMs,
             provider: vendor,
           }),
         });
@@ -2187,7 +2240,7 @@ async function handleFolderTask({ message, channelId, tool, me, caps, cfg, deps,
           onGateDropped: () => noticeUngatedRun(parts[0]),
           cmd: parts[0],
           codeArgs,
-          prompt: memoryPreamble(workspaceMemory) + promptText,
+          prompt: agentPreamble(workspaceMemory, cfg) + promptText,
           cwd: folderPath,
           signal,
           onData: handleCliData,
@@ -2196,6 +2249,7 @@ async function handleFolderTask({ message, channelId, tool, me, caps, cfg, deps,
             tool,
             channelId,
             threadRoot,
+            decisionWaitMs: caps.decisionWaitMs,
             provider: vendor,
           }),
         });
@@ -2204,7 +2258,7 @@ async function handleFolderTask({ message, channelId, tool, me, caps, cfg, deps,
           cmd: parts[0],
           args: [
             ...codeArgs,
-            memoryPreamble(workspaceMemory) + promptText,
+            agentPreamble(workspaceMemory, cfg) + promptText,
           ],
           cwd: folderPath,
           timeoutMs: cfg.runTimeoutMs,
@@ -2715,12 +2769,8 @@ export async function handleTask({ message, channelId, tool, me, caps = {} }, cf
           onStopRequested,
         });
       }
-      // Not a coding task → post the router's reply if it produced one, else fall
-      // through to a normal conversational reply.
-      if (routed.reply) {
-        await tool("post_message", { channelId, parentId, body: routed.reply });
-        return { status: "chat" };
-      }
+      // Not a coding task → the classifier is finished. The separate responder
+      // below owns the answer and keeps its normal tools.
     }
     await respondConversationally({ message, channelId, tool, me, cfg, repoLink, parentId, workspaceMemory, signal, context, caps });
     return { status: "chat" };
@@ -2780,14 +2830,19 @@ export async function handleTask({ message, channelId, tool, me, caps = {} }, cf
       await reactQuietly({ tool, caps, message, emoji: routed.emoji });
       return { status: "quiet" };
     }
-    let body = routed.reply;
-    if (!body) {
-      body =
-        routed.error?.code === "ENOENT"
-          ? `(my chat command \`${chatCmdFor(cfg)}\` isn't installed or on PATH.)`
-          : `Still thinking on this — it's taking longer than usual. I'll follow up shortly.`;
-    }
-    await tool("post_message", { channelId, parentId, body });
+    await respondConversationally({
+      message,
+      channelId,
+      tool,
+      me,
+      cfg,
+      repoLink,
+      parentId,
+      workspaceMemory,
+      signal,
+      context,
+      caps,
+    });
     return { status: "chat" };
   }
   // routed.code → fall through to the coding flow below.
@@ -3300,6 +3355,11 @@ export async function handleTask({ message, channelId, tool, me, caps = {} }, cf
     let drainProgress = async () => {};
     // 0782 — the silent-work stop poll, torn down in the same finally as the run.
     let stopPoller = null;
+    // Structured transports expose the final assistant summary as an event;
+    // argv transports fold it into the progress emitter. Keep one local value
+    // so GitHub title/provenance selection never reaches for an out-of-scope
+    // parser variable (0884).
+    let resultText = "";
     if (streamOn) {
       if (!progressId) {
         try {
@@ -3377,9 +3437,10 @@ export async function handleTask({ message, channelId, tool, me, caps = {} }, cf
     // `exec`, so it has to sit ahead of it. [] for every vendor without a
     // verified flag, leaving their argv byte-identical to before.
     const imageArgs = codeImageArgs(vendor, localImages);
+    const webArgs = codeWebArgs(vendor, parts.slice(1));
     const codeArgs = streamOn
-      ? [...parts.slice(1), ...modelArgs, ...dirArgs, ...imageArgs, ...resumeArgs, ...streamArgs]
-      : [...parts.slice(1), ...modelArgs, ...dirArgs, ...imageArgs, ...resumeArgs];
+      ? [...parts.slice(1), ...webArgs, ...modelArgs, ...dirArgs, ...imageArgs, ...resumeArgs, ...streamArgs]
+      : [...parts.slice(1), ...webArgs, ...modelArgs, ...dirArgs, ...imageArgs, ...resumeArgs];
     const handleCliData = (c) => {
       // Keep tracking lastLine as a fallback (legacy heartbeat / honesty).
       const lines = String(c).split("\n").map((s) => s.trim()).filter(Boolean);
@@ -3420,6 +3481,9 @@ export async function handleTask({ message, channelId, tool, me, caps = {} }, cf
       } catch {
         /* a progress fold must never break the run */
       }
+      if (ev?.t === "result" && typeof ev.summary === "string") {
+        resultText = ev.summary;
+      }
     };
     let run;
     try {
@@ -3468,7 +3532,7 @@ export async function handleTask({ message, channelId, tool, me, caps = {} }, cf
           cmd: parts[0],
           vendor,
           cwd: repoPath,
-          prompt: memoryPreamble(workspaceMemory) + promptText,
+          prompt: agentPreamble(workspaceMemory, cfg) + promptText,
           // 0778: approvals AND continuity. `resume:false` (the never-worse
           // retry) drops it exactly like buildResumeArgs does.
           resumeSessionId: resume ? resumeSessionId : null,
@@ -3481,6 +3545,7 @@ export async function handleTask({ message, channelId, tool, me, caps = {} }, cf
             tool,
             channelId,
             threadRoot,
+            decisionWaitMs: caps.decisionWaitMs,
             runId,
             provider: vendor,
           }),
@@ -3490,7 +3555,7 @@ export async function handleTask({ message, channelId, tool, me, caps = {} }, cf
           cmd: parts[0],
           args: codeArgs,
           cwd: repoPath,
-          prompt: memoryPreamble(workspaceMemory) + promptText,
+          prompt: agentPreamble(workspaceMemory, cfg) + promptText,
           timeoutMs: cfg.runTimeoutMs,
           signal,
           // A model running inside the server must never inherit the daemon's
@@ -3502,6 +3567,7 @@ export async function handleTask({ message, channelId, tool, me, caps = {} }, cf
             tool,
             channelId,
             threadRoot,
+            decisionWaitMs: caps.decisionWaitMs,
             runId,
           }),
         });
@@ -3513,7 +3579,7 @@ export async function handleTask({ message, channelId, tool, me, caps = {} }, cf
           cmd: parts[0],
           codeArgs,
           cwd: repoPath,
-          prompt: memoryPreamble(workspaceMemory) + promptText,
+          prompt: agentPreamble(workspaceMemory, cfg) + promptText,
           model: resolvedModelId,
           // Codex's own resume over this transport. `resume:false` (the
           // never-worse retry) drops it exactly like buildResumeArgs does.
@@ -3525,6 +3591,7 @@ export async function handleTask({ message, channelId, tool, me, caps = {} }, cf
             tool,
             channelId,
             threadRoot,
+            decisionWaitMs: caps.decisionWaitMs,
             runId,
             provider: vendor,
           }),
@@ -3538,7 +3605,7 @@ export async function handleTask({ message, channelId, tool, me, caps = {} }, cf
           // codeArgs already carries this run's resume flags, so a gated
           // iterate resumes AND raises cards — the two never traded off.
           codeArgs,
-          prompt: memoryPreamble(workspaceMemory) + promptText,
+          prompt: agentPreamble(workspaceMemory, cfg) + promptText,
           cwd: repoPath,
           signal,
           onData: handleCliData,
@@ -3548,6 +3615,7 @@ export async function handleTask({ message, channelId, tool, me, caps = {} }, cf
             tool,
             channelId,
             threadRoot,
+            decisionWaitMs: caps.decisionWaitMs,
             runId,
             provider: vendor,
           }),
@@ -3559,7 +3627,7 @@ export async function handleTask({ message, channelId, tool, me, caps = {} }, cf
         // why nothing above the unit tests could ever drive it (0787).
         run = await deps.runCli({
           cmd: parts[0],
-          args: [...codeArgs, memoryPreamble(workspaceMemory) + promptText],
+          args: [...codeArgs, agentPreamble(workspaceMemory, cfg) + promptText],
           cwd: repoPath,
           timeoutMs: cfg.runTimeoutMs,
           label: "coding",
@@ -3630,6 +3698,13 @@ export async function handleTask({ message, channelId, tool, me, caps = {} }, cf
         }
       }
     }
+    if (!resultText && emitter) {
+      try {
+        resultText = emitter.snapshot()?.lastLine || "";
+      } catch {
+        /* title fallback still has the routed task */
+      }
+    }
     if (run.aborted || signal?.aborted) {
       console.log("  code → cancelled");
       return { aborted: true };
@@ -3661,15 +3736,33 @@ export async function handleTask({ message, channelId, tool, me, caps = {} }, cf
         Number((git(repoPath, ["rev-list", "--count", `${baseRef}..HEAD`]).stdout || "0").trim()) || 0;
       if (ahead > 0) {
         console.log(`  code → agent self-committed (${ahead} commit(s) ahead); reconciling`);
-        return { empty: true, failed: false, ahead };
+        return {
+          empty: true,
+          failed: false,
+          ahead,
+          summary: (resultText || (streamArgs.length ? "" : run.stdout) || "").trim(),
+        };
       }
       console.log("  code → no changes produced");
-      return { empty: true, failed: false, ahead: 0 };
+      return {
+        empty: true,
+        failed: false,
+        ahead: 0,
+        summary: (resultText || (streamArgs.length ? "" : run.stdout) || "").trim(),
+      };
     }
     console.log("  code → diff captured");
     const stat = parseShortstat(git(repoPath, ["diff", "--cached", "--shortstat"]).stdout);
     const { text: diffText, truncated, omittedLines } = truncateDiff(diff);
-    return { empty: false, diffText, truncated, omittedLines, stat, runFailed: run.status !== 0 };
+    return {
+      empty: false,
+      diffText,
+      truncated,
+      omittedLines,
+      stat,
+      runFailed: run.status !== 0,
+      summary: (resultText || (streamArgs.length ? "" : run.stdout) || "").trim(),
+    };
   };
 
   // Persist the run's coding-agent session (0282) so a LATER iterate can resume it.
@@ -3886,6 +3979,7 @@ export async function handleTask({ message, channelId, tool, me, caps = {} }, cf
       parentId: threadRoot,
       settleId: selfSettleId,
       runId,
+      provenance: githubProvenance({ cfg, me, message, channelId, runId, summary: staged.summary }),
       usageArgs: reportUsageArgs(runUsage, { vendor, modelId: resolvedModelId, runId }),
     });
     await recordSession(selfResult.prUrl);
@@ -3962,6 +4056,7 @@ export async function handleTask({ message, channelId, tool, me, caps = {} }, cf
       existingPrUrl: continuingPrUrl,
       settleId,
       runId,
+      provenance: githubProvenance({ cfg, me, message, channelId, runId, summary: staged.summary }),
       usageArgs: reportUsageArgs(runUsage, { vendor, modelId: resolvedModelId, runId }),
     });
     // The settle already flipped the card to the report (body + metadata); editing
@@ -4024,6 +4119,7 @@ export async function handleTask({ message, channelId, tool, me, caps = {} }, cf
     parentId: threadRoot,
     existingPrUrl: continuingPrUrl,
     runId,
+    provenance: githubProvenance({ cfg, me, message, channelId, runId, summary: staged.summary }),
     // Normally {} — a gated run already reported its spend on the proposal card
     // it is now shipping. Anything the CLI spent after that still comes home.
     usageArgs: reportUsageArgs(runUsage, { vendor, modelId: resolvedModelId, runId }),

package/src/mcp.mjs

@@ -48,13 +48,20 @@ export function makeClient({ url, token }) {
     }
   }
 
-  async function listToolNames() {
+  // Full tool objects, schemas included — capability sniffing reads these
+  // (0866: an ARG-level addition like waitMs shows up in a tool's inputSchema,
+  // not in the tool list's names).
+  async function listTools() {
     try {
-      return ((await rpc("tools/list"))?.tools ?? []).map((t) => t.name);
+      return (await rpc("tools/list"))?.tools ?? [];
     } catch {
       return [];
     }
   }
 
-  return { rpc, tool, listToolNames };
+  async function listToolNames() {
+    return (await listTools()).map((t) => t.name);
+  }
+
+  return { rpc, tool, listTools, listToolNames };
 }

package/src/progress-emitter.mjs

@@ -57,7 +57,8 @@ export function detectVendor(codingCmd) {
  * Code installed just to answer chat (0521). Every command is the vendor's
  * verified non-interactive print mode; the daemon appends the prompt as the last
  * arg. codex carries --skip-git-repo-check because chat (and the read-only
- * review sandbox) can run outside a git checkout. cursor carries
+ * review sandbox) can run outside a git checkout, plus the explicit web-search
+ * config because `codex exec` otherwise leaves that tool off. cursor carries
  * --output-format text (explicit, so a CLI default change can never post raw
  * JSONL into the channel) and --trust (its Jan-2026 workspace-trust gate fails
  * headless runs at spawn in untrusted directories — 0572; pre-2026 CLIs reject
@@ -70,7 +71,9 @@ export function detectVendor(codingCmd) {
  */
 export function fastChatCmd(vendor) {
   if (vendor === "claude_code") return "claude -p --model haiku";
-  if (vendor === "codex") return "codex exec --skip-git-repo-check";
+  if (vendor === "codex") {
+    return "codex exec --skip-git-repo-check -c tools.web_search=true";
+  }
   if (vendor === "cursor") return "cursor-agent -p --output-format text --trust";
   if (vendor === "opencode") return "opencode run";
   if (vendor === "antigravity") return "agy -p";
@@ -107,6 +110,31 @@ export function codeStreamArgs(vendor) {
   return [];
 }
 
+/**
+ * Make Codex's built-in public web search available to code runs. This is a
+ * capability flag, not an instruction to browse; Codex decides whether the
+ * task needs it. An operator's explicit true/false override wins unchanged.
+ * Other vendors already expose their own web tools and receive no guessed
+ * flags.
+ * @param {'claude_code'|'codex'|'cursor'|'opencode'|'antigravity'|'hermes'|'unknown'} vendor
+ * @param {string[]} baseArgs
+ * @returns {string[]}
+ */
+export function codeWebArgs(vendor, baseArgs = []) {
+  if (vendor !== "codex") return [];
+  const hasOverride = baseArgs.some((arg, index) => {
+    if (/^tools\.web_search=/.test(arg)) return true;
+    if (
+      (baseArgs[index - 1] === "-c" || baseArgs[index - 1] === "--config") &&
+      /^tools\.web_search=/.test(arg)
+    ) {
+      return true;
+    }
+    return /^--config=tools\.web_search=/.test(arg);
+  });
+  return hasOverride ? [] : ["-c", "tools.web_search=true"];
+}
+
 /**
  * Extra args that hand the code run an IMAGE, appended to the code run's argv
  * (0779). Verified against the installed binaries, per the 0521 rule — never