@bivy/bivy 0.16.25-staging.9 → 0.17.0-staging.2

package/README.md

@@ -206,6 +206,10 @@ answer questions, approve supported tool calls, or stop the agent.
 
 - Send screenshots, images, logs, and other files from your phone.
 - Download reports and artifacts the agent creates.
+- Publish [session apps](docs/apps.md) with web and CLI/TUI views using
+  `bivy app publish bivy.app.json`. Linked machines automatically discover web
+  preview delivery through Bivy's relay—no per-machine domains, certificates or
+  public ports. Terminal views use the existing connection.
 - Use voice input and read-aloud where supported; provider-backed voice may
   send audio or text to the selected provider.
 - Keep a native terminal workflow or use structured chat, depending on the agent.

package/bin/acp-shim.mjs

@@ -313,6 +313,13 @@ function onSessionUpdate(params) {
     if (!content) return "";
     if (typeof content === "string") return content;
     if (Array.isArray(content)) return content.map(textOf).join("");
+    // ACP's ToolCallContent wraps a ContentBlock: `{ type: "content", content:
+    // { type: "text", text } }`. Unwrap it (recursively) so a tool's real output
+    // — a command's stdout, a fetched document, a read's text — isn't dropped,
+    // which collapsed the tool_result to the bare status ("completed"). opencode
+    // and most spec-compliant ACP agents deliver tool output in this wrapped
+    // shape, so reading only the bare `{type:"text"}` form lost it in production.
+    if (content.type === "content" && content.content != null) return textOf(content.content);
     if (content.type === "text" && typeof content.text === "string") return content.text;
     if (typeof content.text === "string") return content.text;
     return "";

package/bin/agent-manifest.json

@@ -22,7 +22,7 @@
       "hidden": false,
       "supportTier": "supported",
       "certification": "release-tested",
-      "testedVersion": "1.18.30",
+      "testedVersion": "1.18.32",
       "headlessFlags": [
         "run",
         "-s"
@@ -191,7 +191,7 @@
       "hidden": false,
       "supportTier": "supported",
       "certification": "release-tested",
-      "testedVersion": "1.0.0",
+      "testedVersion": "1.0.41",
       "headlessFlags": [
         "--always-approve",
         "-p",

package/bin/bivy.mjs

@@ -2418,6 +2418,57 @@ async function cmdAttach(args = []) {
   console.log(c.green(`Attached ${body.name} (${body.kind}, ${body.size} bytes) to the chat.`));
 }
 
+// Universal app publishing: every agent that can write JSON and run a command
+// uses the same contract. Relative static directories are manifest-relative.
+async function cmdApp(args = []) {
+  const [action, ...rest] = args;
+  if (!action || rest.some((arg) => arg === "--help" || arg === "-h") || ["help", "--help", "-h"].includes(action)) {
+    console.log(`Usage: bivy app publish <manifest.json> [--session <id>]
+       bivy app list [--session <id>]
+       bivy app remove <app-id> [--session <id>]
+
+Manifest: {"version":1,"name":"My app","views":[
+  {"kind":"web","name":"Website","source":{"kind":"service","port":3000}},
+  {"kind":"terminal","name":"Console","command":"bin/rails","args":["console"]}
+]}
+
+Static web source: {"kind":"static","directory":"./dist"}
+Publishing does not start commands. Open Apps in the session menu to view or run.
+Terminal commands run with the node user's permissions, not in a new sandbox.
+Web previews require operator setup; see docs/apps.md.`);
+    return;
+  }
+  if (!["publish", "list", "remove"].includes(action)) throw new Error("Unknown app command. Run bivy app --help.");
+  const sessionIndex = rest.indexOf("--session");
+  if (sessionIndex >= 0 && (!rest[sessionIndex + 1] || rest[sessionIndex + 1].startsWith("--"))) throw new Error("--session requires an id.");
+  const sessionId = resolveAttachSessionId({ sessionFlag: sessionIndex >= 0 ? rest[sessionIndex + 1] : undefined, env: process.env });
+  if (!sessionId) throw new Error("Set --session <id> or run inside a Bivy agent session.");
+  const positional = rest.filter((_, i) => i !== sessionIndex && (sessionIndex < 0 || i !== sessionIndex + 1));
+  if (positional.some((a) => a.startsWith("-")) || positional.length !== (action === "list" ? 0 : 1)) throw new Error("Invalid arguments. Run bivy app --help.");
+  const body = { sessionId };
+  if (action === "publish") {
+    const file = path.resolve(positional[0]);
+    if (fs.statSync(file).size > 64 * 1024) throw new Error("App manifest exceeds 64 KiB.");
+    const manifest = JSON.parse(fs.readFileSync(file, "utf8"));
+    if (!manifest || typeof manifest !== "object") throw new Error("App manifest must be a JSON object.");
+    if (Array.isArray(manifest.views)) for (const view of manifest.views) {
+      if (view?.kind === "web" && view.source?.kind === "static" && typeof view.source.directory === "string") {
+        view.source.directory = path.resolve(path.dirname(file), view.source.directory);
+      }
+    }
+    body.manifest = manifest;
+  } else if (action === "remove") body.appId = positional[0];
+  const config = loadConfig();
+  if (!(await ensureNodeRunning(config))) throw new Error("Could not reach the Bivy node.");
+  const token = await localDeviceToken(config);
+  const response = await fetch(`${url(config)}/api/apps/${action}`, {
+    method: "POST", headers: { authorization: `Bearer ${token}`, "content-type": "application/json" }, body: JSON.stringify(body),
+  });
+  const result = await response.json();
+  if (!response.ok || result.error) throw new Error(result.error || `App command failed (${response.status}).`);
+  console.log(JSON.stringify(result, null, 2));
+}
+
 // Map a saved session's runtime id to the `bivy run` agent whose native CLI can
 // resume it in a terminal. Only agents with a real native resume qualify; other
 // runtimes (generic-cli, SDK-only) have no terminal resume and open in the web app.
@@ -5045,6 +5096,7 @@ ${c.bold("bivy")} — Bivy node CLI
   ${c.cyan("bivy token")}       Print a device token for this node (for 'bivy nodes add' elsewhere)
   ${c.cyan("bivy sessions")}     List recent sessions (live + saved) and resume one (alias: ls)
   ${c.cyan("bivy resume")} [n|id] Resume a session directly (default: most recent)
+  ${c.cyan("bivy app")}             Publish and list session apps (web and terminal views)
   ${c.cyan("bivy send <id>")} "..."  Send a prompt to an existing session and stream the reply
   ${c.cyan("bivy kill <id>")}    Stop a session/terminal (--delete also removes a saved session)
   ${c.cyan("bivy prune")}         Delete old sessions/workspaces/worktrees (--keep N, --older-than 7d, --dry-run)
@@ -5167,6 +5219,9 @@ An agent's own --help passes through, e.g. 'bivy run claude --help'.`);
     case "send":
       await cmdSend(args);
       break;
+    case "app":
+      await cmdApp(args);
+      break;
     case "attach":
       await cmdAttach(args);
       break;

package/bin/codex-app-server-shim.mjs

@@ -415,6 +415,16 @@ function onNotification(m) {
       bivy({ type: "session.status", status: "working" });
       return;
     case "turn/completed": {
+      // The completion event is authoritative about a terminal failure: a turn
+      // that hit a usage limit (or any provider error) arrives here as
+      // `status: "failed"` carrying `turn.error`. Capture it directly instead of
+      // depending on a separate `error` notification landing first — otherwise a
+      // completion that omits that notification seals as session.done and the
+      // user sees a silent empty turn (fail loud, never swallow the error).
+      const turnError =
+        params?.turn?.error?.message ||
+        (params?.turn?.status === "failed" ? params?.error?.message || params?.message || "Codex turn failed" : undefined);
+      if (turnError && !sawError) sawError = turnError;
       // Codex 0.147 can publish turn/completed just before the final
       // collabAgentToolCall item/completed notification. Sealing immediately
       // leaves the sub-agent card running forever and drops its persisted

package/bin/patch-pi-dependencies.mjs

@@ -37,8 +37,8 @@ const piPackage = findDependency(path.join("@earendil-works", "pi-coding-agent")
 if (!piPackage) process.exit(0);
 
 const patches = [
-  { name: "brace-expansion", version: "5.0.9" },
-  { name: "undici", version: "8.10.0" },
+  { name: "brace-expansion", version: "5.0.12" },
+  { name: "undici", version: "8.10.2" },
 ];
 
 for (const patch of patches) {

package/dist/agents/claude-code/integration.js

@@ -4,13 +4,21 @@ import { defineAgentIntegration } from "../definition.js";
 import { withExactCapabilitySurface } from "../../runtime/types.js";
 import { createCredentialStore } from "../../runtime/credentials.js";
 import { ClaudeCodeRuntime, claudeRuntimeFromEnv, claudeSdkInstalled, } from "./runtime.js";
-export const CLAUDE_TESTED_VERSION = "0.3.270";
+export const CLAUDE_TESTED_VERSION = "0.3.281";
 const CLAUDE_CAPABILITIES = withExactCapabilitySurface({
     toolInterception: true,
     modelSelection: true,
     packages: false,
     resume: true,
     fork: true,
+    // Must mirror ClaudeCodeRuntime.capabilities (runtime.ts): the catalog is a
+    // consumed source of truth (the fork-fidelity matrix and any fork-target
+    // fidelity preview read these), so under-reporting here makes Claude look
+    // like a seeded-only fork target when its runtime actually delivers a
+    // byte-exact same-runtime fork (forkTransport) and a full cross-runtime
+    // transcript replay INTO it (forkHistoryImport, via importHistoryForFork).
+    forkTransport: true,
+    forkHistoryImport: true,
     usageReporting: true,
     nativeSessionDiscovery: true,
     nativeSessionAdoption: true,

package/dist/agents/codex/integration.js

@@ -9,7 +9,7 @@ import { codexCredentialPreflight } from "../../runtime/codex-preflight.js";
 import { deleteCodexSession, discoverNativeCodexSessions, loadCodexTranscript, writeCodexRollout, exportCodexRollout, importCodexRollout, } from "../../runtime/codex-sessions.js";
 import { ProtocolRuntime } from "../../runtime/protocol.js";
 import { codexSlashCommands } from "../../runtime/slash-commands.js";
-export const CODEX_TESTED_VERSION = "0.154.0";
+export const CODEX_TESTED_VERSION = "0.156.1";
 const CODEX_AVAILABLE_CACHE = new Map();
 function codexCommand() {
     return process.env.BIVY_CODEX_BIN?.trim() || "codex";
@@ -146,6 +146,14 @@ export function codexIntegration(origin) {
                     resume: true,
                     packages: false,
                     fork: true,
+                    // Mirror the ProtocolRuntime this integration builds (createCodexRuntime
+                    // above wires exportForFork+importForFork and writeHistory, which flip
+                    // capabilities.forkTransport/forkHistoryImport on). The catalog is a
+                    // consumed source of truth (fork-fidelity matrix / fork-target preview),
+                    // so omitting these made Codex look seeded-only when its runtime does a
+                    // byte-exact codex→codex fork and a full transcript replay INTO it.
+                    forkTransport: true,
+                    forkHistoryImport: true,
                     usageReporting: true,
                     interactiveTui: installed,
                     nativeSessionDiscovery: true,

package/dist/agents/pi/integration.js

@@ -5,7 +5,7 @@ import os from "node:os";
 import path from "node:path";
 import { defineAgentIntegration } from "../definition.js";
 import { PI_CAPABILITIES } from "./capabilities.js";
-export const PI_TESTED_VERSION = "0.85.1";
+export const PI_TESTED_VERSION = "0.87.1";
 export function piAgentDir() {
     return process.env.PI_CODING_AGENT_DIR?.trim() || path.join(os.homedir(), ".pi", "agent");
 }

package/dist/agents/pi/runtime.js

@@ -268,8 +268,29 @@ class PiSession {
      * reload. Async because that resolution reads the store.
      */
     async getModels() {
-        const models = await this.session.modelRuntime.getAvailable();
-        return models.map((model) => toModelInfo(model, true));
+        const runtime = this.session.modelRuntime;
+        try {
+            const models = await runtime.getAvailable();
+            return models.map((model) => toModelInfo(model, true));
+        }
+        catch (error) {
+            // The runtime's availability refresh is a Promise.all across every
+            // provider's auth check, so one broken provider (an unreachable custom
+            // endpoint, a single failing credential read) rejects the whole catalog
+            // and the picker shows nothing. Degrade to per-provider listing and
+            // skip only the broken ones; rethrow only when truly nothing survives.
+            const models = [];
+            for (const provider of runtime.getProviders()) {
+                try {
+                    for (const model of await runtime.getAvailable(provider.id))
+                        models.push(toModelInfo(model, true));
+                }
+                catch { /* this provider is broken — the rest still list */ }
+            }
+            if (models.length)
+                return models;
+            throw error instanceof Error ? error : new Error(String(error));
+        }
     }
     /**
      * Every model Pi knows about — built-in and custom, regardless of whether
@@ -281,7 +302,10 @@ class PiSession {
     async getAllModels() {
         const runtime = this.session.modelRuntime;
         // Refresh availability first so `hasConfiguredAuth` reflects current creds.
-        await runtime.getAvailable();
+        // Best-effort: a failed refresh (one broken provider rejects the shared
+        // Promise.all) must not blank the full catalog — the flags then reflect
+        // the last successful refresh, which the picker can still act on.
+        await runtime.getAvailable().catch(() => { });
         return runtime.getModels().map((model) => toModelInfo(model, runtime.hasConfiguredAuth(model.provider)));
     }
     /** Reload the projected models.json in-place so a live session immediately

package/dist/agents/profiles.js

@@ -43,7 +43,7 @@ export const AGENT_PROFILES = {
         // Approve/Deny + session/load resume + a real model picker), the same bar Pi,
         // Claude Code, and Codex clear. See `acp` below for the version fallback.
         supportTier: "supported",
-        testedVersion: "1.18.30",
+        testedVersion: "1.18.32",
         blurb: "The most widely used open-source coding harness (OpenCode CLI).",
         // `opencode run -s <id> "<prompt>"` continues a prior session by its own id
         // (`-s, --session  session id to continue`, per `opencode run --help`).
@@ -356,8 +356,11 @@ export const AGENT_PROFILES = {
         behaviors: { preflight: "grok", prepare: "grok-auth", nativeSessions: "grok" },
         command: "grok",
         packageName: "grok (curl -fsSL https://x.ai/cli/install.sh | bash)",
+        // Supported tier: Grok runs on the governed ACP path by default (per-tool
+        // Approve/Deny + native resume + live model state), the same bar Pi, Claude
+        // Code, Codex, and OpenCode clear. See `acp` below for the version fallback.
         supportTier: "supported",
-        testedVersion: "1.0.0",
+        testedVersion: "1.0.41",
         authOwner: "mixed",
         blurb: "xAI's official Grok coding agent (Grok CLI) — SuperGrok/X subscription or API key.",
         // Official CLI: `grok -p "<prompt>"` (alias `--single`) runs one headless
@@ -398,14 +401,32 @@ export const AGENT_PROFILES = {
             flag: "-m",
             models: [
                 // Official Grok CLI (1.x) advertises grok-4.6 as the default subscription
-                // model (verified against `grok models`); the older grok-4.5 / grok-4-latest
-                // / grok-code-fast-1 ids now return "unknown model id". Keep the list
-                // honest; operators can override with BIVY_GROK_MODELS if their install
-                // exposes more.
+                // model (verified against the authenticated `~/.grok/models_cache.json`,
+                // origin cli-chat-proxy.grok.com/v1/models); the older grok-4.5 /
+                // grok-4-latest / grok-code-fast-1 ids now return "unknown model id".
+                // Keep the list honest; operators can override with BIVY_GROK_MODELS if
+                // their install exposes more.
                 { id: "grok-4.6", name: "Grok 4.6", provider: "xai" },
             ],
         },
+        // Grok's reasoning models expose a `--reasoning-effort <EFFORT>` flag
+        // (verified against `grok --help`, 1.0.0); the authenticated model catalog
+        // lists low/medium/high/xhigh with `high` as the default. `{level}` fills
+        // the value; insertAt: 0 keeps the trailing `-p` prompt flag last so the
+        // prompt still lands as its value. Threads through the resume path too, so a
+        // continued turn keeps the chosen effort. Override with BIVY_GROK_THINKING.
+        thinking: { levels: ["low", "medium", "high", "xhigh"], default: "high", template: ["--reasoning-effort", "{level}"], insertAt: 0 },
         promptMode: "argv",
+        // `grok agent stdio` ("Run the agent over stdio") is a full ACP server
+        // (protocolVersion 1: session/new, session/prompt, session/request_permission,
+        // loadSession resume, live model state) — the governed ProtocolRuntime path
+        // that supersedes the `--always-approve` pipe above with real per-tool
+        // Approve/Deny. Validated against grok 1.0.41, so it is ON by default
+        // (`preferred`) — gated on the binary actually describing the `agent`
+        // subcommand ("Run Grok without the interactive UI" in `grok --help`), so an
+        // older CLI falls back to the pipe path rather than opening a dead session.
+        // Force the pipe path back with BIVY_GROK_ACP=0.
+        acp: { args: ["agent", "stdio"], helpToken: "without the interactive ui", preferred: true },
         install: {
             kind: "curl",
             display: "curl -fsSL https://x.ai/cli/install.sh | bash",