@bivy/bivy 0.20.2-staging.1 → 0.20.2-staging.2

package/bin/acp-shim.mjs

@@ -219,6 +219,16 @@ function finishTurnDone() {
   bivy({ type: "session.done" });
 }
 
+// ACP `plan` updates carry the agent's whole todo list with a status per entry.
+// Forward each as a finished `plan` tool call: it persists in the turn like any
+// tool, and clients render the latest one as a checklist (@bivy/core plan.ts).
+let planSeq = 0;
+function emitPlan(entries) {
+  const toolCallId = `plan-${Date.now().toString(36)}-${++planSeq}`;
+  bivy({ type: "tool.observe", toolCallId, name: "plan", input: { entries } });
+  bivy({ type: "tool.result", toolCallId, name: "plan", result: "Plan updated" });
+}
+
 // --- tool-call field normalization -------------------------------------------
 // ACP's `tool_call`/`tool_call_update` carries a free-text `title` (whatever
 // prose the agent chose) AND a small fixed `kind` enum (read/edit/delete/move/
@@ -466,8 +476,7 @@ function onSessionUpdate(params) {
       break;
     }
     case "plan":
-      // Optional planning stream — fold into reasoning so nothing is lost.
-      if (!child && Array.isArray(u.entries)) bivy({ type: "message.reasoning", text: u.entries.map((e) => `• ${e.content ?? ""}`).join("\n") });
+      if (!child && Array.isArray(u.entries)) emitPlan(u.entries);
       break;
     default:
       break;

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

@@ -160,6 +160,16 @@ function emitToolObserved(itemId, name, input) {
   bivy({ type: "tool.observe", toolCallId: itemId, name, input });
 }
 
+// `turn/plan/updated` (the update_plan tool) carries the whole plan with a
+// status per step. Forward it as a finished `plan` tool call, which clients
+// render as a checklist (@bivy/core plan.ts), the same as an ACP plan.
+let planSeq = 0;
+function emitPlan(plan, explanation) {
+  const toolCallId = `plan-${Date.now().toString(36)}-${++planSeq}`;
+  bivy({ type: "tool.observe", toolCallId, name: "plan", input: { plan, ...(explanation ? { explanation } : {}) } });
+  bivy({ type: "tool.result", toolCallId, name: "plan", result: "Plan updated" });
+}
+
 function emitReasoning(itemId, delta) {
   const body = text(delta);
   if (!body) return;
@@ -364,6 +374,9 @@ function onNotification(m) {
     case "item/plan/delta":
       if (fromCurrentThread) emitReasoning(paramsItemId(params), params.delta);
       return;
+    case "turn/plan/updated":
+      if (fromCurrentThread && Array.isArray(params?.plan)) emitPlan(params.plan, params.explanation);
+      return;
     case "item/reasoning/summaryTextDelta":
     case "item/reasoning/textDelta":
       if (fromCurrentThread) emitReasoning(paramsItemId(params), params.delta);

package/bin/tui/transcript.mjs

@@ -25,6 +25,22 @@ export function toolSummary(block) {
   return detail ? `${name} ${detail.replace(/\s+/g, " ").trim()}` : name;
 }
 
+/** "plan · 2 of 4 done · <current step>" for a full plan update (the shapes
+ *  @bivy/core plan.ts reads), or undefined. A partial `merge` update is skipped:
+ *  the full list it amends is already shown. */
+export function planSummary(block) {
+  const input = block?.input ?? block?.arguments ?? {};
+  if (!input || typeof input !== "object" || input.merge === true) return undefined;
+  const list = ["todos", "plan", "entries", "items", "steps", "tasks"].map((k) => input[k]).find(Array.isArray);
+  const steps = (list ?? []).filter((e) => e && typeof e === "object");
+  if (!steps.length) return undefined;
+  const status = (e) => String(e.status ?? e.state ?? "").toLowerCase().replace(/[^a-z]/g, "");
+  const text = (e) => String(e.content ?? e.step ?? e.text ?? e.title ?? e.subject ?? "").trim();
+  const done = steps.filter((e) => ["completed", "complete", "done"].includes(status(e))).length;
+  const current = steps.find((e) => ["inprogress", "active", "running"].includes(status(e)));
+  return `plan · ${done} of ${steps.length} done${current && text(current) ? ` · ${text(current)}` : ""}`;
+}
+
 /** `[{ role: "user"|"agent"|"tool", text }]` from `session.history` messages. */
 export function transcriptEntries(messages = []) {
   const entries = [];
@@ -46,7 +62,15 @@ export function transcriptEntries(messages = []) {
         const id = block.id ?? block.toolCallId;
         if (id && seenTools.has(id)) continue;
         if (id) seenTools.add(id);
-        entries.push({ role: "tool", text: toolSummary(block) });
+        const plan = planSummary(block);
+        if (plan) {
+          // One plan line per turn, showing the latest progress.
+          const previous = entries.findLastIndex((e) => e.role === "user" || e.plan);
+          if (previous >= 0 && entries[previous].plan) entries.splice(previous, 1);
+          entries.push({ role: "tool", text: plan, plan: true });
+        } else if (!/^(plan|todo_?write|update_?plan)$/i.test(String(block.name ?? ""))) {
+          entries.push({ role: "tool", text: toolSummary(block) });
+        }
       }
     }
   }

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

@@ -4,7 +4,7 @@ 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 = "0.3.285";
+export const CLAUDE_TESTED_VERSION = "0.3.286";
 const CLAUDE_CAPABILITIES = withExactCapabilitySurface({
     toolInterception: true,
     modelSelection: true,

package/dist/agents/codex/integration.js

@@ -10,7 +10,7 @@ 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 = "0.159.1";
+export const CODEX_TESTED_VERSION = "0.159.2";
 function codexCommand() {
     return process.env.BIVY_CODEX_BIN?.trim() || "codex";
 }

package/dist/agents/pi/integration.js

@@ -6,7 +6,7 @@ import path from "node:path";
 import { defineAgentIntegration } from "../definition.js";
 import { PI_CAPABILITIES } from "./capabilities.js";
 import { bridgeInstalled, enableBridgeResolution } from "../../agent-bridges.mjs";
-export const PI_TESTED_VERSION = "0.99.1";
+export const PI_TESTED_VERSION = "0.99.2";
 export function piAgentDir() {
     return process.env.PI_CODING_AGENT_DIR?.trim() || path.join(os.homedir(), ".pi", "agent");
 }

package/dist/apps/styles.css

@@ -689,6 +689,22 @@ input, textarea, select { font-family: inherit; font-size: 16px; }
 .tool-detail-value.output pre,
 .tool-detail-value.output code { white-space: pre-wrap; word-break: break-word; overflow-wrap: anywhere; }
 
+/* Agent plan: the work-line (progress) over a quiet checklist. */
+.plan-chevron { transition: transform var(--dur-fast) var(--ease-standard); }
+.plan-card .tool-group-line[aria-expanded="true"] .plan-chevron { transform: rotate(90deg); }
+.plan-steps { list-style: none; margin: 0 0 var(--space-1); padding: 0; display: grid; gap: var(--space-1); font-size: var(--text-sm); }
+.plan-step { display: flex; align-items: baseline; gap: var(--space-2); color: var(--ink); }
+.plan-mark {
+  flex: none; width: 14px; height: 14px; box-sizing: border-box; align-self: center; display: inline-flex; align-items: center; justify-content: center;
+  border: 1.5px solid var(--line-strong); border-radius: var(--radius-full); font-size: 9px; font-weight: 700; line-height: 1; color: var(--muted);
+}
+.plan-step.is-completed .plan-mark { border-color: var(--ok); background: var(--ok); color: var(--bg); }
+.plan-step.is-completed .plan-mark::before { content: "\2713"; }
+.plan-step.is-completed .plan-text { color: var(--muted); }
+.plan-step.is-in_progress .plan-mark { border-color: var(--accent); box-shadow: inset 0 0 0 2.5px var(--bg), inset 0 0 0 7px var(--accent); }
+.plan-step.is-in_progress .plan-text { font-weight: var(--weight-medium); }
+.plan-step.is-cancelled .plan-text { color: var(--muted); text-decoration: line-through; }
+
 .diff-viewer { max-width: 100%; overflow: auto; border: 1px solid var(--line); border-radius: var(--radius-md); background: var(--surface-2); font-family: ui-monospace, SFMono-Regular, Menlo, monospace; font-size: 12px; line-height: 1.45; }
 .diff-hunk + .diff-hunk { border-top: 1px solid var(--line); }
 .diff-hunk-title { position: sticky; top: 0; z-index: var(--z-base); display: flex; align-items: center; justify-content: space-between; gap: 8px; padding: 7px 10px; color: var(--muted); background: var(--surface-3); border-bottom: 1px solid var(--line); font-family: inherit; font-size: 12px; font-weight: 700; }

package/dist/certification/generated.js

@@ -3,9 +3,9 @@
 export const CERTIFICATION_MATRIX = {
     schemaVersion: 1,
     agents: [
-        { id: "claude-code-sdk", status: "active", executionMode: "protocol", pinnedVersion: "0.3.285", capabilities: ["toolInterception", "modelSelection", "resume"] },
-        { id: "codex-approvals", status: "active", executionMode: "protocol", pinnedVersion: "0.159.1", capabilities: ["toolInterception", "modelSelection", "resume"] },
-        { id: "pi", status: "active", executionMode: "protocol", pinnedVersion: "0.99.1", capabilities: ["toolInterception", "modelSelection", "resume"] },
+        { id: "claude-code-sdk", status: "active", executionMode: "protocol", pinnedVersion: "0.3.286", capabilities: ["toolInterception", "modelSelection", "resume"] },
+        { id: "codex-approvals", status: "active", executionMode: "protocol", pinnedVersion: "0.159.2", capabilities: ["toolInterception", "modelSelection", "resume"] },
+        { id: "pi", status: "active", executionMode: "protocol", pinnedVersion: "0.99.2", capabilities: ["toolInterception", "modelSelection", "resume"] },
         { id: "opencode", status: "active", executionMode: "protocol", pinnedVersion: "1.18.33", capabilities: ["toolInterception", "modelSelection", "resume"] },
         { id: "grok", status: "active", executionMode: "protocol", pinnedVersion: "1.0.44", capabilities: ["toolInterception", "modelSelection", "resume"] },
     ]

package/dist/server.js

@@ -34,6 +34,8 @@ import { AppRegistry } from "./apps/registry.js";
 import { AppGateway } from "./apps/gateway.js";
 import { RemotePreview } from "./apps/remote-preview.js";
 import { AppService } from "./apps/service.js";
+import { scanListeners, scanMachineListeners } from "./apps/listeners.js";
+import { AGENT_PROFILES } from "./agents/profiles.js";
 import { pngSize, reviewHint } from "./apps/review.js";
 import { createDisplayHost } from "./apps/display.js";
 import { createAppCommands } from "./controllers/app-commands.js";
@@ -2252,6 +2254,10 @@ function createPinSink() {
         },
     };
 }
+// An agent CLI's own server is not an app the user runs: OpenCode's ACP process
+// listens on :4096 in the workspace, and was offered as a preview in every session.
+const agentCommands = new Set(Object.values(AGENT_PROFILES).map((profile) => profile.command));
+const notAgentServers = (offers) => offers.filter((offer) => !agentCommands.has(offer.command.split(" ")[0]));
 const appService = new AppService(appRegistry, appGateway ?? remotePreview, {
     start: async (spec) => {
         let failure = "Could not start the app terminal.";
@@ -2266,7 +2272,11 @@ const appService = new AppService(appRegistry, appGateway ?? remotePreview, {
     },
     has: (id) => terminals.has(id),
     close: (id) => { terminals.close(id); },
-}, { screenshots: { enabled: appScreenshotsEnabled }, displays: appDisplays, reviews: createReviewSink(), pins: createPinSink() });
+}, {
+    scan: async (workspace) => notAgentServers(await scanListeners(workspace)),
+    scanMachine: async (root) => notAgentServers(await scanMachineListeners(root)),
+    screenshots: { enabled: appScreenshotsEnabled }, displays: appDisplays, reviews: createReviewSink(), pins: createPinSink(),
+});
 // A reviewer's note shows up in an open Apps sheet without reopening it, and
 // the owner gets one push per burst of notes on a view ("2 notes on
 // Storefront"), which opens the Apps sheet at that app. Counts, names and IDs

package/dist/session/handoff-seed.js

@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright (c) 2026 Petter André Sjulstad
+//
+// A hand-off seed is the first prompt that carries a conversation into an agent
+// that couldn't import it natively: a cross-agent fork, or a session imported
+// from outside Bivy. It is written for the agent, not the reader — the
+// transcript above it already shows the same conversation — so clients render it
+// as a compact "context handed over" line instead of a wall of user text.
+//
+// The node writes these (src/session/transcript-normal.ts buildSeedPrompt and
+// src/session/native-import.ts). One row per shape; a new seed is a new row.
+const SEEDS = [
+    { kind: "fork", pattern: /^I am continuing an existing Bivy session \(forked from (.+?) to .+?\)\./ },
+    { kind: "import", pattern: /^I am continuing an? (.+?) session that was started outside Bivy and imported here\./ },
+];
+/** The hand-off a user message carries, or undefined for an ordinary prompt. */
+export function handoffSeedOf(text) {
+    if (!text)
+        return undefined;
+    for (const seed of SEEDS) {
+        const match = seed.pattern.exec(text);
+        if (match)
+            return { kind: seed.kind, from: match[1] };
+    }
+    return undefined;
+}

package/dist/session/plan.js

@@ -0,0 +1,99 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright (c) 2026 Petter André Sjulstad
+//
+// An agent's plan (todo list) as one shape. Agents publish it three ways: a todo
+// tool call (`todos: [{content, status}]`), Codex's `turn/plan/updated`
+// (`plan: [{step, status}]`), and ACP's `plan` session update
+// (`entries: [{content, status}]`). The shims forward the last two as a `plan`
+// tool call, so every plan reaches the transcript as a tool input, and this
+// module reads any of them. Spellings are table rows, not per-agent code.
+const LIST_KEYS = ["todos", "plan", "entries", "items", "steps", "tasks"];
+const TEXT_KEYS = ["content", "step", "text", "title", "subject", "description", "task"];
+const STATUS = {
+    pending: "pending", todo: "pending", notstarted: "pending", open: "pending",
+    inprogress: "in_progress", active: "in_progress", running: "in_progress", doing: "in_progress", current: "in_progress", started: "in_progress",
+    completed: "completed", complete: "completed", done: "completed", finished: "completed", success: "completed",
+    cancelled: "cancelled", canceled: "cancelled", skipped: "cancelled", abandoned: "cancelled",
+};
+function planStatus(value) {
+    return STATUS[String(value ?? "").toLowerCase().replace(/[^a-z]/g, "")];
+}
+function firstString(o, keys) {
+    for (const key of keys) {
+        const value = o[key];
+        if (typeof value === "string" && value.trim())
+            return value.trim();
+    }
+    return undefined;
+}
+/**
+ * The plan a tool input carries, or undefined when it holds no list of steps.
+ * `merge` marks a partial update (entries keyed by id, text may be missing) that
+ * applies on top of the previous plan rather than replacing it.
+ */
+export function planUpdateOf(input) {
+    if (!input || typeof input !== "object" || Array.isArray(input))
+        return undefined;
+    const o = input;
+    const list = LIST_KEYS.map((key) => o[key]).find(Array.isArray);
+    if (!list?.length)
+        return undefined;
+    const entries = [];
+    for (const raw of list) {
+        if (typeof raw === "string") {
+            entries.push({ text: raw, status: "pending" });
+            continue;
+        }
+        if (!raw || typeof raw !== "object")
+            continue;
+        const item = raw;
+        const text = firstString(item, TEXT_KEYS);
+        const status = planStatus(item.status ?? item.state);
+        const id = item.id ?? item.taskId;
+        entries.push({ ...(text ? { text } : {}), ...(status ? { status } : {}), ...(id != null && id !== "" ? { id: String(id) } : {}) });
+    }
+    const merge = o.merge === true;
+    if (!merge && !entries.some((e) => e.text))
+        return undefined;
+    return { entries, merge };
+}
+/** Apply one update to the previous plan: replace it, or merge entries by id. */
+export function applyPlanUpdate(previous, update) {
+    if (!update.merge) {
+        return update.entries.filter((e) => e.text).map((e) => ({ text: e.text, status: e.status ?? "pending", ...(e.id ? { id: e.id } : {}) }));
+    }
+    const next = previous.map((e) => ({ ...e }));
+    for (const change of update.entries) {
+        const existing = change.id ? next.find((e) => e.id === change.id) : undefined;
+        if (existing) {
+            if (change.text)
+                existing.text = change.text;
+            if (change.status)
+                existing.status = change.status;
+        }
+        else if (change.text) {
+            next.push({ text: change.text, status: change.status ?? "pending", ...(change.id ? { id: change.id } : {}) });
+        }
+    }
+    return next;
+}
+/** The plan after a sequence of plan tool inputs, oldest first. */
+export function latestPlan(inputs) {
+    let plan = [];
+    for (const input of inputs) {
+        const update = planUpdateOf(input);
+        if (update)
+            plan = applyPlanUpdate(plan, update);
+    }
+    return plan;
+}
+const CHECKLIST = {
+    completed: { box: "[x]", note: "" },
+    in_progress: { box: "[ ]", note: " (in progress)" },
+    pending: { box: "[ ]", note: "" },
+    cancelled: { box: "[ ]", note: " (cancelled)" },
+};
+/** A plan as a Markdown checklist, for prompts and transcript files. */
+export function planChecklist(plan) {
+    return plan.map((e) => `- ${CHECKLIST[e.status].box} ${e.text}${CHECKLIST[e.status].note}`).join("\n");
+}

package/dist/session/transcript-normal.js

@@ -1,3 +1,5 @@
+import { handoffSeedOf } from "./handoff-seed.js";
+import { applyPlanUpdate, planChecklist, planUpdateOf } from "./plan.js";
 /** Cap a tool payload down to a short, log-safe one-liner. */
 function compactValue(value, max = 200) {
     let s;
@@ -105,8 +107,10 @@ function toTs(value) {
  */
 export function normalizeMessages(messages, header) {
     const turns = [];
+    let plan = [];
     for (const message of messages ?? []) {
         const m = message;
+        plan = planAfter(plan, m.content);
         const { text, tools, toolResultOnly } = readContent(m.content);
         const role = normalizeRole(m.role, toolResultOnly);
         const namedTool = tools.find((t) => t.name)?.name;
@@ -122,8 +126,25 @@ export function normalizeMessages(messages, header) {
             ...(toTs(m.timestamp ?? m.ts) !== undefined ? { ts: toTs(m.timestamp ?? m.ts) } : {}),
         });
     }
-    return { header, turns };
+    return { header, turns, ...(plan.length ? { plan } : {}) };
 }
+/** Fold a message's top-level plan updates (see @bivy/core plan.ts) into `plan`. */
+function planAfter(plan, content) {
+    if (!Array.isArray(content))
+        return plan;
+    for (const block of content) {
+        if (block?.type !== "tool_use" || block.parentToolUseId || block.parentToolCallId)
+            continue;
+        if (block.detail?.kind && block.detail.kind !== "plan")
+            continue;
+        const update = planUpdateOf(block.input);
+        if (update)
+            plan = applyPlanUpdate(plan, update);
+    }
+    return plan;
+}
+/** Cap for the latest request when the seed repeats it in full. */
+const REQUEST_CHARS = 4000;
 function truncate(text, max) {
     const compact = String(text || "").replace(/\s+/g, " ").trim();
     return compact.length > max ? `${compact.slice(0, Math.max(0, max - 1))}…` : compact;
@@ -168,6 +189,14 @@ export function buildSeedPrompt(transcript, opts = {}) {
     picked.reverse();
     const omitted = formatted.length - picked.length;
     const recent = picked.length ? picked.join("\n") : "- (no prior turns were available)";
+    // The user's latest request is what the new agent must act on. Repeat it in
+    // full when the recent block had to shorten it (or leave it out). An earlier
+    // hand-off seed is context, not a request.
+    const lastRequest = [...transcript.turns].reverse().find((t) => t.role === "user" && t.text.trim() && !handoffSeedOf(t.text))?.text.trim();
+    const requestCut = lastRequest && lastRequest.replace(/\s+/g, " ").length > perTurnChars;
+    const pinnedRequest = lastRequest && (requestCut || !picked.some((line) => line.startsWith("- user:")))
+        ? (lastRequest.length > REQUEST_CHARS ? `${lastRequest.slice(0, REQUEST_CHARS - 1)}…` : lastRequest)
+        : undefined;
     const file = opts.transcriptFile;
     const where = file ? "the conversation file above" : opts.transcriptUrl ? "the full transcript linked above" : "the full transcript";
     const lines = [
@@ -185,6 +214,8 @@ export function buildSeedPrompt(transcript, opts = {}) {
             ? `Recent conversation (most recent last; ${omitted} earlier turn${omitted === 1 ? "" : "s"} omitted — see ${where}):`
             : "Recent conversation (most recent last):",
         recent,
+        ...(pinnedRequest ? ["", "The user's latest request, in full:", pinnedRequest] : []),
+        ...(transcript.plan?.length ? ["", "The plan so far (from the previous agent):", planChecklist(transcript.plan)] : []),
         "",
         file
             ? "Read the conversation file above only if this summary is missing something you need; otherwise continue from here."
@@ -236,6 +267,13 @@ export function buildForkHistory(transcript) {
         else
             history.push({ role, text });
     }
+    // The plan's latest state, in the agent's voice on its last reply, so the new
+    // agent picks up the open steps rather than re-deriving them from tool
+    // payloads. Never after a trailing user turn: that request is still open.
+    const lastReply = history.map((turn) => turn.role).lastIndexOf("assistant");
+    if (transcript.plan?.length && lastReply >= 0) {
+        history[lastReply].text += `\n\n[plan] Where the plan stands:\n${planChecklist(transcript.plan)}`;
+    }
     return history;
 }
 /**

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bivy/bivy",
-  "version": "0.20.2-staging.1",
+  "version": "0.20.2-staging.2",
   "type": "module",
   "license": "AGPL-3.0-only",
   "description": "Run coding agents on machines you own. Open-source, self-hostable agent workspace.",
@@ -68,8 +68,8 @@
   "readme": "# Bivy\n\n[![npm](https://img.shields.io/npm/v/@bivy/bivy?color=2b6cb0&label=%40bivy%2Fbivy)](https://www.npmjs.com/package/@bivy/bivy)\n[![license: AGPL-3.0-only](https://img.shields.io/badge/license-AGPL--3.0--only-2b6cb0)](LICENSE)\n[![node](https://img.shields.io/badge/node-%E2%89%A520-2b6cb0)](https://nodejs.org)\n\n**Local coding agents that show their work. Anywhere.**\n\nBivy is the open-source workspace for coding agents. It runs Claude Code, Codex,\nPi, OpenCode and other agents on your own machine. Reach them from a browser or\nyour phone, with the live session, its terminal, its approvals, and the app\nthey're building right in the chat. Try it, mark what's wrong, and send it back.\n\n<p align=\"center\">\n  <img src=\"docs/images/preview-feedback-loop.gif\" width=\"360\"\n       alt=\"On a phone: open the app an agent built, circle the packed items, send a note, and compare the result before and after the agent's fix.\">\n</p>\n\nBivy is not another coding agent and not a cloud development machine. The agent\nstays local and does the coding with your model provider. Your repos, tools,\ndatabases, and services stay where they are. Bivy gives you remote access to all\nof it, plus live previews, automations, and review.\n\n- **Live previews.** The running app beside the chat. Mark it, send it back,\n  share it with someone who has no Bivy account.\n- **Automations.** GitHub issues, failed CI, Linear, Slack, schedules, and\n  webhooks start the work. It comes back as a preview link or a pull request.\n- **Any agent.** Hit a usage limit? Fork the session to another agent, or let\n  Bivy retry when the limit resets.\n- **Your machine, anywhere.** Start at your desk, pick the session up on your\n  phone. The machine dials out, so there are no ports to open, and session\n  traffic is end-to-end encrypted.\n\n**[Start free on Bivy Cloud](https://app.bivy.sh)** ·\n**[Quickstart](docs/quickstart.md)** ·\n**[Documentation](docs/README.md)** ·\n**[Self-host](docs/deploy-images.md)** ·\n**[Website](https://bivy.sh)**\n\n**Recommended:** sign in at [app.bivy.sh](https://app.bivy.sh), then copy your\npersonalized **Connect a Machine** command into a terminal on your Mac or Linux\ncomputer. It installs and enrolls the machine without another Bivy login. The\nbrowser connects automatically; choose a repository and send your first task\nright there.\n\nPrefer starting from the terminal?\n\n```bash\ncurl -fsSL https://bivy.sh/install.sh | bash  # install + guided setup\ncd your-repo\nbivy run claude                            # or codex, pi, opencode\nbivy open                                  # continue in the web app (needs remote setup)\n```\n\nBivy Cloud hosts the app, control plane, and relay—not the machines running your\nagents. Connect a Mac, Linux computer, or existing server and bring your own\nagent subscription, model API key, or local model. You can also self-host the\nentire remote-access stack.\n\n> **Bivy is 0.x software.** Claude Code, Codex, Pi, OpenCode, and Grok are the\n> release-tested paths. Credential sync, resume, handoffs, approvals, and\n> sandboxing depend on the runtime. See the\n> [runtime support matrix](docs/runtime-support-matrix.md).\n\n## Don't just read the diff. Try the app.\n\nBivy lets you **use what the agent built and show it what to fix**, without a\nseparate deployment step.\n\n1. **Ask for a change.** The agent works in your repo with your existing tools\n   and local services.\n2. **Try the app.** Bivy finds the dev server the agent starts and opens it\n   beside the chat: web apps, terminals, and desktop apps. Use the running app,\n   not a screenshot of what the agent says it finished.\n3. **Mark what needs work.** Press and hold anything in the preview to mark it,\n   or drag to circle it, then say what you want. **Mark another** keeps\n   separate notes separate. The marks go to the agent with the element and page\n   context. No screenshots to paste.\n4. **Follow each note.** Every mark stays in the chat as a pin that resolves\n   itself: **Changed** when a later run changes what you marked, **Element\n   gone** when it leaves the page, **Done** when you say so.\n5. **Review the next version.** A new version waits on the preview pill instead\n   of reloading under you. Take it when you're ready, compare before and after,\n   inspect the diff and checks, and decide when the work is done.\n6. **Get a second opinion.** Share a preview link for 1 hour, 1 day, or 7 days.\n   Teammates and clients try the running app and leave notes without a Bivy\n   account. Make it view-only, or **Stop sharing** to end every link at once.\n\nOn Bivy Cloud, preview delivery is built in: no per-app domains, certificates,\npublic ports, or tunnels to configure. Self-hosters configure preview delivery\nonce for their Bivy deployment, not for every app. The app still needs its normal\nbuild or dev-server setup, and **the machine serving it must stay awake and\nonline**. These are development previews, not production hosting.\n\n**Share deliberately:** anyone with a preview link can use that app, including\nits live backend, until the link expires or you stop sharing.\nReviewer notes aren't sent to the agent automatically; you send them or explicitly\nallow agent access. Preview traffic uses HTTPS through the preview relay, not\nsession end-to-end encryption; the relay operator can see it.\n\n[App previews, visual feedback, and sharing →](docs/apps.md)\n\n## One workflow, from trigger to review\n\n```text\nPrompt · GitHub issue · CI failure · Linear · Slack · Schedule · Webhook\n                                 │\n                                 ▼\n                   Choose machine + agent + model\n                     + supported credentials\n                                 │\n                                 ▼\n                         Live agent session\n                     Join · steer · approve · stop\n                                 │\n                                 ▼\n                  Try app · inspect changes · checks\n                                 │\n                                 ▼\n                  Mark up · send feedback · iterate\n                                 │\n                                 ▼\n                      Share preview · review PR\n```\n\nA **Machine** is a computer or server you connect. A **Session** is live agent\nwork on that machine. A **Run** is delegated background work that creates a\nsession and tracks its outcome. An **Automation** is a reusable definition that\ncreates runs when an event matches.\n\nManual and automated work use the same kind of live session. You can join a run\nwhen it needs help rather than wait for a black-box job to finish.\n\n### Let events start the work\n\nAutomations turn recurring or incoming work into sessions you can join,\nsupervise, and review. An agent picks the task up on your machine and posts\nback a preview link or a pull request, so you try the result where the task\nlives and reply in the same session. Choose the repository, machine, agent,\nmodel, approval mode, sandbox setting, and maximum attempts.\n\n| Trigger | Example workflow |\n|---|---|\n| **GitHub issues and mentions** | Label an issue `bivy` or `bivy/<machine>`, or mention your Bivy GitHub App, to work toward a pull request. |\n| **Failed CI** | Match a failed workflow, ask the agent to reproduce it, make a fix, and run the affected checks. |\n| **Linear** | Label an issue to start work without copying its description into an agent. |\n| **Slack** | Send a request from the conversation where the work came up. |\n| **Schedules** | Run a weekly dependency review, recurring maintenance, or a one-time task. |\n| **Signed webhooks** | Connect alerts, internal tools, or your own event sources. |\n\nConfigure automations in the app or version them with your repository in\n`.bivy/automations.yaml`:\n\n```bash\nbivy automation init\n# Edit the generated definition for your repository and workflow.\nbivy automation validate\nbivy automation test --event .bivy/events/failed-ci.yaml  # supply a local event fixture\nbivy automation apply\n```\n\nOr delegate a one-off job without creating an automation:\n\n```bash\nbivy runs start \"Review outdated dependencies and propose a small, tested update.\"\nbivy runs wait <id>\n```\n\nRuns keep routing and lifecycle evidence, check results, and output references\nin a reviewable Receipt. For unattended issue work, Bivy runs declared repository\nchecks after the agent's turn; failed required checks fail the run even if the\nagent reports success. A completed process alone is not proof that the task\nsucceeded.\n\n[Automation recipes →](docs/capability-recipes.md#let-events-start-runs) ·\n[Automations as code →](docs/automations-as-code.md) ·\n[Run outcomes and reliability limits →](docs/automation-runs.md)\n\n### Use every agent you pay for. Switch mid-task.\n\nRun Claude Code for one task, Codex for another, and Pi or OpenCode where they\nfit, side by side in one session list. Bivy supplies the shared session,\nremote-access, automation, and review surfaces; your chosen agent still does\nthe coding and uses your model provider.\n\n- **Hit a usage limit? Keep going.** When a turn fails on a limit, fork the\n  session, with its code and conversation, to another agent and model, or let\n  Bivy retry when the limit resets (the provider has to report a reset time).\n- **Get a second opinion.** `bivy delegate` hands a task to another agent, on\n  the same machine or a different one, and brings back its answer, branch, and\n  PR. `--to codex,claude@linux` sends one task to several agents to compare.\n- Choose an agent and, where supported, a model for each session or run.\n- Import existing Claude Code and Codex sessions.\n- Fork or move work to another agent or machine when a different setup fits\n  better. Continuation fidelity varies: some paths preserve native history,\n  while others replay portable turns or hand over a summary with the earlier\n  transcript as a local file.\n- Use agent-native logins, Bivy-managed credentials, or local inference.\n  Bivy's custom OpenAI-compatible endpoint registry currently feeds Pi;\n  other agents may need their own provider configuration.\n- Register your own ACP or headless process agent with `bivy agent add`.\n\nBivy does not replace your agent, provide model inference, or make every agent's\nfeatures identical. Consult the [support matrix](docs/runtime-support-matrix.md)\nand [handoff recipes](docs/capability-recipes.md#fork-or-move-a-session).\n\n### Less signing in. Less copying secrets.\n\nBivy syncs **Bivy-managed API keys and supported OAuth credentials** across\nenrolled machines for compatible runtimes. Connect supported credentials once\nand reuse them where you run work, rather than manually distributing keys to\neach machine.\n\nFor ordinary account sync, credentials are encrypted on the node before upload.\nThe control plane stores ciphertext and wrapped-key metadata; enrolled nodes\nshare access by wrapping the vault key to one another. Bivy Cloud does not\nreceive plaintext credentials through this sync path.\n\nYou can also keep credentials local, use labeled keys and project presets, or\nreference environment variables and 1Password instead of embedding secrets in\nconfiguration:\n\n```bash\nbivy provider login\nbivy credentials add anthropic work\nbivy secrets ref github.repo-token op://Bivy/GitHub/repo-token\n```\n\n**Not every CLI login syncs.** Native agent logins may still be per-machine;\nGitHub App private-key sync is separately opt-in. If you lose every node and\ndevice able to unwrap a vault, you must sign in to providers again. Explicit\nhosted-provisioning custody grants are separate from ordinary encrypted sync.\n\n[Credential sync and runtime coverage →](docs/credential-sync.md) ·\n[Credentials guide →](docs/credentials-guide.md) ·\n[Key storage →](docs/key-management.md)\n\n### Work in the environment you already have\n\nA clean cloud sandbox isn't always enough. Your agent may need the database\nrunning on localhost, an uncommitted change, an internal API behind your VPN,\nor a model running on your GPU. Bivy runs the agent where those things already\nexist, subject to that machine's permissions and the runtime's protection.\n\nConnect several machines to the same account: a laptop for interactive work,\na Linux server for background jobs, or a GPU box for local inference. Choose\nthe machine for each session or pin it in an automation. Repository runs can\nuse isolated Git worktrees without rebuilding the whole development environment.\n\n**The execution machine must stay awake and online.** To close your laptop and\nleave work running, run the agent on a different, always-on machine.\n\n[Environment and multi-machine recipes →](docs/capability-recipes.md)\n\n### Start at your desk. Continue anywhere.\n\nOpen the same session in the browser, phone PWA, or terminal. Watch work live,\nanswer questions, approve supported tool calls, or stop the agent.\n\n- Send screenshots, images, logs, and other files from your phone.\n- Download reports and artifacts the agent creates.\n- Open [session apps](docs/apps.md) from chat or the **Apps** menu: live web\n  previews, desktop app views, and CLI/TUI tools. Agents can publish views with\n  `bivy app publish bivy.app.json` and create share links with `bivy app share`.\n- Use voice input and read-aloud where supported; provider-backed voice may\n  send audio or text to the selected provider.\n- Keep a native terminal workflow or use structured chat, depending on the agent.\n\n```bash\nbivy run claude --no-follow  # start without attaching\nbivy open                    # open the web app\nbivy resume                  # return to the session in your terminal\nbivy link                    # pair a device directly via QR\n```\n\nNo phone app installation is required. Open [app.bivy.sh](https://app.bivy.sh)\nin your browser; adding it to your home screen is optional.\n\n[Remote access →](docs/remote-access.md) ·\n[Voice, files, and terminal recipes →](docs/capability-recipes.md)\n\n### Agents can reach you, too\n\nEvery session gets the `bivy` CLI, so any agent with a shell can talk back\nthrough the app, not only agents with their own built-in tools:\n\n```bash\nbivy notify \"Tests are green, PR is up\"     # chat card + phone push when you're away\nbivy ask \"Ship to staging?\" --option Yes --option No  # waits for your answer\nbivy attach report.png --caption \"Before/after\"       # show a file in the chat\nbivy suggest \"Add a dark theme to settings\"           # a next task you start in one tap\nbivy context --json                                   # session, workspace, machine, apps\n```\n\n`bivy help --json` lists every command, and with `BIVY_OUTPUT=json` failures\ncome back as structured errors with distinct exit codes, so an agent can tell\nwhat went wrong and what to run next. `bivy guide` prints short playbooks for\nagents, and the same commands are served as MCP tools for agents that prefer\nthem. Add your own standing instructions for\nevery session in **Settings → Agent instructions**.\n\n[Commands for agents inside a session →](docs/cli-reference.md#inside-an-agent-session) ·\n[Agent instructions →](docs/agent-instructions.md)\n\n## Get started\n\n### Install\n\nBivy supports **macOS and Linux with Node.js 20+**. The installer installs the\n`@bivy/bivy` package, runs guided setup, and starts a launchd or systemd service:\n\n```bash\ncurl -fsSL https://bivy.sh/install.sh | bash\n```\n\nSetup helps you choose an agent and configure remote access. Existing agents\nkeep their command, login, and configuration. The installer may use `sudo` to\ninstall a missing Node.js, but never for `npm install`. To inspect it first,\ndownload it with `curl -fsSL https://bivy.sh/install.sh -o install.sh`.\n\nAlready have Node.js and want to avoid sudo?\n\n```bash\nnpm install -g @bivy/bivy\nbivy setup\n```\n\nThen try one small task:\n\n```bash\ncd your-repo\nbivy run claude\n# Ask: \"Explain this repo and make one small, safe improvement. Run the relevant checks.\"\nbivy open\n```\n\nOpen that same session on your phone while it runs. For a web app, ask the agent\nto start its dev server, then open **Apps → Running in this workspace → Preview**.\nTry a page, point to something you'd change, and send the feedback back to the\nagent. No separate preview deployment needed.\n\nOnce that works, share a preview for review, connect another machine, or add\nyour first automation.\n\n**Local-only works too.** `bivy run`, `bivy resume`, and `bivy sessions` need no\naccount or server. Choose **local only for now** during setup; use `bivy login`\nlater. Browser and phone access need a hosted or self-hosted control plane;\nthe node itself does not serve a web UI.\n\n[Full quickstart →](docs/quickstart.md) ·\n[Installer options, service management, and uninstall →](docs/install.md)\n\n### Choose hosted or self-hosted\n\n| Option | What you get |\n|---|---|\n| **Free Cloud — $0** | Every launch feature, including automations; 10 new remote sessions per rolling seven days. No credit card required. |\n| **Cloud — $15/month** | The same features with unlimited remote sessions. |\n| **Self-hosted Core** | Operate the app, control plane, and relay yourself, with no Bivy usage limits. |\n\nManual and automated sessions share the Cloud allowance. Resuming existing\nsessions and viewing history do not consume it. Agent subscriptions and model\nprovider charges are separate. See [current pricing](https://bivy.sh#pricing).\n\n**Self-host anywhere:** deploy the public control-plane (including the web app)\nand relay images with Postgres and [a small set of environment variables](docs/deploy-images.md).\nYour server or container platform handles HTTPS. Set up owner access in the\nbrowser—no SSH, external authentication provider, or Bivy Cloud account required.\nFor a bare VPS, the [Compose installer](docs/self-host-quickstart.md) automates the\nsame stack. These onboarding features require a release containing them.\n\nStart on Cloud and self-host later if you prefer. Deploy the stack, reconnect\nmachines with `bivy relay:setup`, and pair devices to your server. This is not a\none-click migration of your Cloud account; your local repos and agent\nconfiguration stay in place.\n\n```bash\nbivy relay:setup \\\n  --control-plane https://bivy.example.com \\\n  --relay wss://relay.example.com\n```\n\nSelf-hosting is community-supported: you own TLS, backups, upgrades, and\nhardening. Public multi-architecture images are available as\n`ghcr.io/bivysh/bivy-control-plane` and `ghcr.io/bivysh/bivy-relay`; pin a release\nversion or full commit SHA.\n\n[Deploy the images anywhere →](docs/deploy-images.md) ·\n[Optional VPS installer →](docs/self-host-quickstart.md) ·\n[Operations reference →](docs/self-host.md)\n\n## Architecture\n\nYour environment, with clear security boundaries:\n\n```text\nYour machine                         Hosted or self-hosted\n┌──────────────────────┐             ┌──────────────────────┐\n│ Node daemon          │──outbound──▶│ Relay                │\n│ Agents, repos, tools │             │ Encrypted frames     │\n│ Local credentials    │             └──────────┬───────────┘\n└──────────────────────┘                        │\n                                    Browser / phone\n                                    + control plane\n                                    (app, accounts, metadata)\n```\n\n- **Execution stays on your machine.** Bivy Cloud does not run your agents.\n  Your model provider still sees whatever the agent sends it.\n- **Interactive session traffic is end-to-end encrypted** between the node and\n  paired devices. The relay forwards opaque session frames; your node dials out,\n  so no inbound public port is required.\n- **App previews have a separate security boundary.** Preview traffic uses HTTPS\n  and an outbound tunnel, not session E2E encryption. The preview ingress/relay\n  operator can see it. Public preview links grant anyone holding them access to\n  that view and its live backend until expiry or revocation. See\n  [preview security and sharing](docs/apps.md#runtime-and-security-boundaries).\n- **Ordinary credential sync uploads ciphertext, not plaintext keys.**\n  Supported credentials and recovery limits are documented separately.\n- **Encryption is not universal across integrations.** Slack commands and\n  generic webhook instructions reach the control plane in plaintext. Do not\n  put secrets in them. Routing and bounded run metadata are also visible there.\n- **Device authorization matters.** QR pairing authorizes a device directly\n  through the node. Hosted account pairing trusts the control plane to authorize\n  devices and serve the web app that holds client keys.\n- **Bivy is not an OS-level sandbox.** The default approval mode is\n  `autonomous`; protection depends on the runtime. Some agents enforce sandbox\n  tiers, while process agents may run with your full user permissions.\n  Heuristic tool checks help prevent accidents but are not isolation.\n\nReview the runtime's Protection label and configure approval/sandbox settings\nfor the task, especially before enabling unattended work.\n\n[Security model and known limitations →](docs/security-model.md) ·\n[Runtime protection matrix →](docs/runtime-support-matrix.md) ·\n[Configuration →](docs/configuration.md)\n\n## Agents and everyday commands\n\n**Claude Code, Codex, Pi, OpenCode, and Grok are release-tested.** Additional\nadapters include Gemini CLI, Qwen Code, Goose, Aider, Cline, Crush, Cursor, GitHub\nCopilot, Amp, Auggie, Droid, Continue, Kilo Code, and Rovo Dev. Installation,\nresume, model selection, and tool protection vary—see the\n[support matrix](docs/runtime-support-matrix.md) and [agent guides](docs/agents/README.md).\n\nRun an arbitrary command with `bivy run -- ./your-agent --flags`, register a\nreusable entry with `bivy agent add`, or package a declarative integration with\nexperimental [plugins](docs/plugins.md).\n\n```bash\nbivy run claude       # launch a durable session; also codex, pi, opencode\nbivy sessions         # list live and saved sessions\nbivy resume           # resume the most recent session\nbivy open             # open the web app (requires remote setup)\nbivy nodes            # list connected account machines\nbivy runs list        # inspect delegated work\nbivy automation init  # scaffold repo-owned automations\nbivy provider login   # connect supported model credentials\nbivy agent add        # register an ACP or process agent\nbivy doctor           # check installation and connectivity\nbivy logs -f          # follow node logs\nbivy update           # update and restart the service\n```\n\n`bivy update` uses your original installation method and waits for an active\nturn to finish before restarting. Use `--force` to skip that wait.\n\n[CLI reference →](docs/cli-reference.md) ·\n[Node and project configuration →](docs/config-as-code.md) ·\n[GitHub setup →](docs/github-setup.md) ·\n[Linear setup →](docs/linear-work-queue.md)\n\n## Development and contributions\n\n```bash\npnpm install\npnpm run dev          # node daemon on http://localhost:4317\npnpm run dev:web      # web client dev server\n```\n\n| Directory | Contents |\n|---|---|\n| `src/`, `bin/` | Node daemon, CLI, runtime adapters, sessions, approvals, secrets |\n| `packages/core/` | Shared protocol, pairing, and wire format |\n| `packages/web/`, `packages/ui/` | React PWA and shared design system |\n| `services/relay/` | Self-hostable encrypted relay |\n| `services/control-plane/` | Self-hostable control plane |\n| `deploy/` | Deployment examples |\n\n```bash\npnpm run typecheck\npnpm run typecheck:web\npnpm run lint\npnpm run test:unit\npnpm run test:core\npnpm run check:licenses\npnpm run check:secrets\n```\n\nSee [CONTRIBUTING.md](CONTRIBUTING.md) for the development workflow. Releases\nare published from CI with provenance attestations; see\n[release verification](docs/releasing.md).\n\n**Found a security issue?** Use\n[GitHub private vulnerability reporting](https://github.com/bivysh/bivy/security/advisories/new),\nnot a public issue. See [SECURITY.md](SECURITY.md).\n\n### In development—not available at launch\n\nAutomatically provisioned, short-lived **ephemeral machines** are in development\nfor hosted and bring-your-own-cloud deployments. Neither path is ready or\nsupported for this launch. Use an existing computer or server you operate.\nExperimental provisioning has different credential-custody and encryption\nboundaries; see the [provisioning trust model](docs/hosted-provisioning-trust-model.md).\n\n## License\n\nEverything in this repository—node, CLI, web/PWA, relay, and control plane—is\nfree and open-source **AGPL-3.0-only Core**, with no Bivy usage limits. You may\nuse, modify, and self-host it under that license. If users interact with your\nmodified version over a network, section 13 requires you to offer its\ncorresponding source. See [LICENSE](LICENSE).\n\nBivy Cloud is the hosted operation of that stack plus billing and plans, in a\nseparate private repository. Contributions use the\n[DCO](CONTRIBUTING.md#certificate-of-origin); there is no CLA.\n",
   "readmeFilename": "README.md",
   "agentBridges": {
-    "@anthropic-ai/claude-agent-sdk": "0.3.285",
-    "@earendil-works/pi-ai": "0.99.1",
-    "@earendil-works/pi-coding-agent": "0.99.1"
+    "@anthropic-ai/claude-agent-sdk": "0.3.286",
+    "@earendil-works/pi-ai": "0.99.2",
+    "@earendil-works/pi-coding-agent": "0.99.2"
   }
 }