podium

PandaCode

50 of 51 criteria have a sourced finding; 50 were also cross-checked by a second agent. Every comparison that includes PandaCode links here for scope, limitations, atomic answers and sources.

Humans approve the rubric, templates and prototype; they do not review every finding. An agent source cross-check is not a hands-on product test. A missing sourced finding is not evidence that a capability is absent.

Rubric
0.3 · 51 criteria
Source checks
2026-11-16
Agent cross-checked
50 criteria
Comparisons
1

Access and agent choice

ACC-01

Client access

Agent cross-checked partial

Desktop app for macOS 12+ (Apple silicon and Intel), Windows 10 1809+/11 (x64, arm64) and Linux (Ubuntu 22.04+ x86_64, 24.04+ aarch64 AppImage, X11 or XWayland); Windows and Linux builds are previews. iPhone companion through TestFlight beta, Android from source only. No browser client, no editor extension.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds), PandaCode mobile companion (Expo; iPhone TestFlight beta, Android from source), Headless host (pandacode serve) driven from other devices through PandaCode Connect · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: macOS, Windows, Linux · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • desktop os: Six release targets: macOS 12+ arm64 and x64, Windows 10 1809+/11 x64 and arm64, Linux x86_64 (Ubuntu 22.04+) and aarch64 (Ubuntu 24.04+) AppImage needing X11 or XWayland. platform-support.md says Windows and Linux builds are previews with incomplete interactive validation.
  • mobile apps: README: watch, steer and approve from an iPhone or Android phone; tasks never run on the phone. iPhone is a TestFlight beta; Android has no published installer (build from mobile/ with Expo; open issue 381 asks for APKs).
  • browser: No hosted browser client: bounded search of README, docs/ and relay/ found only the desktop webview; npm run dev serves a labelled device-local sample workspace in a regular browser with simulated files, Git and terminals for UI work.
  • mobile browser: No mobile web client; the relay is a Cloudflare Worker API and the phone client is the Expo app (mobile/README.md).
  • cli: pandacode serve (headless host with --sign-in, --sign-out, --device-name, --full-access flags) and npx pandacode (installer); no CLI that creates, steers or reviews tasks (task control is through the app or other devices).
  • editor extensions: No editor extension: package.json, src-tauri/ and README list none; PandaCode imports VS Code themes but ships no VS Code or JetBrains plugin.
  • notifications: Menu bar item lists running, waiting and finished tasks; quick overlay and running-agents widget on desktop; the relay sends generic push alerts to the phone ("A task finished") after the user allows notifications. No email or chat notification channel for tasks (bots use chat channels separately).
9 atomic answers
  • Desktop os [partial] Six release targets: macOS 12+ arm64 and x64, Windows 10 1809+/11 x64 and arm64, Linux x86_64 (Ubuntu 22.04+) and aarch64 (Ubuntu 24.04+) AppImage needing X11 or XWayland. platform-support.md says Windows and Linux builds are previews with incomplete interactive validation. Sources: pandacode:platform, pandacode:readme, pandacode:rel.
  • Mobile apps [partial] README: watch, steer and approve from an iPhone or Android phone; tasks never run on the phone. iPhone is a TestFlight beta; Android has no published installer (build from mobile/ with Expo; open issue 381 asks for APKs). Sources: pandacode:readme, pandacode:mob, pandacode:issues.
  • Browser [unsupported] No hosted browser client: bounded search of README, docs/ and relay/ found only the desktop webview; npm run dev serves a labelled device-local sample workspace in a regular browser with simulated files, Git and terminals for UI work. Sources: pandacode:readme.
  • Mobile browser [unsupported] No mobile web client; the relay is a Cloudflare Worker API and the phone client is the Expo app (mobile/README.md). Sources: pandacode:mob, pandacode:remote.
  • Cli [partial] pandacode serve (headless host with --sign-in, --sign-out, --device-name, --full-access flags) and npx pandacode (installer); no CLI that creates, steers or reviews tasks (task control is through the app or other devices). Sources: pandacode:head, pandacode:readme.
  • Editor extensions [unsupported] No editor extension: package.json, src-tauri/ and README list none; PandaCode imports VS Code themes but ships no VS Code or JetBrains plugin. Sources: pandacode:readme, pandacode:pkg.
  • Task actions [supported] Desktop: start (New task, board, quick overlay), steer and queue messages while an agent runs, approve, review turn diffs and the Changes inspector, resume, stop. Mobile: view, send, steer, approve and stop tasks on connected computers. Sources: pandacode:readme, pandacode:mob, pandacode:board.
  • Notifications [partial] Menu bar item lists running, waiting and finished tasks; quick overlay and running-agents widget on desktop; the relay sends generic push alerts to the phone ("A task finished") after the user allows notifications. No email or chat notification channel for tasks (bots use chat channels separately). Sources: pandacode:readme.
  • Cross device continuity [supported] PandaCode Connect mirrors a task started on another signed-in computer or headless host into the local sidebar with streaming, approvals and steering; the mobile inbox lists every computer task. Sources: pandacode:remote, pandacode:mob.

Sources: pandacode:readme, pandacode:platform, pandacode:mob, pandacode:rel, pandacode:issues, pandacode:remote, pandacode:head, pandacode:pkg, pandacode:board

Agent cross-checked partial

Runs 20+ existing coding CLIs as harnesses, not an embedded loop: Claude Code, Codex, Kimi Code, OpenCode, Antigravity, Grok Build, Pi, omp, Cursor CLI, Copilot CLI, DeepSeek, Command Code, Cline, Kilo, Devin, Hermes, ZCode, Muse Code and three Claude Code relays (GLM, MiMo, Alibaba Token Plan). Model choice per task; mixed harnesses supported in teams.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • custom adapters: No user-written adapter API found; support comes from built-in adapters (ACP, Pi RPC, ZCode app-server, Muse MSP, Claude/Codex protocols). Any ACP-speaking CLI listed is a built-in entry, and custom OpenAI-compatible endpoints are supported only for specific harnesses.
5 atomic answers
  • Harnesses [supported] README and architect.rs HARNESSES list 21 harnesses: Claude Code, Codex, Kimi Code, OpenCode, GLM, Xiaomi MiMo and Alibaba Token Plan via Claude Code, Antigravity, Grok Build, Pi, DeepSeek, Command Code, Cline, Kilo Code, omp, GitHub Copilot CLI, Muse Code, Devin CLI, Hermes Agent, Cursor CLI, ZCode. Sources: pandacode:readme, pandacode:arch.
  • Model providers [supported] Each harness exposes its own models from the installed catalog; GLM, MiMo, Alibaba Token Plan and DeepSeek are model providers reached through harnesses; custom OpenAI-compatible endpoints (Ollama, local models) are configurable. Sources: pandacode:prov, pandacode:readme.
  • Custom adapters [partial] No user-written adapter API found; support comes from built-in adapters (ACP, Pi RPC, ZCode app-server, Muse MSP, Claude/Codex protocols). Any ACP-speaking CLI listed is a built-in entry, and custom OpenAI-compatible endpoints are supported only for specific harnesses. Sources: pandacode:prov, pandacode:readme.
  • Session integration [supported] Controls sessions through each harness protocol (Codex app-server, Claude CLI stream, ACP JSON-RPC, Pi RPC, ZCode app-server, Muse MSP) with approvals, steering and resume IDs; also an embedded terminal for terminal-only commands. Sources: pandacode:readme, pandacode:prov, pandacode:goals.
  • Mixed harness operation [supported] Architect and each worker can independently use different harnesses; a failed worker moves to the same model on another harness; a task can switch harness with a bounded conversation transfer. Sources: pandacode:arch, pandacode:prov.

Sources: pandacode:readme, pandacode:prov, pandacode:arch, pandacode:goals

Agent cross-checked partial

Reuses each harness own sign-in (Claude Code, Codex, Kimi and others) so existing subscriptions work; also API keys for keyed providers (GLM, MiMo, DeepSeek, Command Code, Cloud) and local OpenAI-compatible endpoints; account pools and named connections switch between several accounts. PandaCode sells no credits or inference.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • vendor credits: README states "PandaCode does not include provider plans or usage"; no PandaCode credits or resold inference found in docs or source.
  • provider restrictions: Docs tell users to pool only accounts they own and that each must be their own subscription; vendor terms for third-party use of subscription logins are not summarised, and Claude-based use depends on Anthropic terms outside the repository.
7 atomic answers
  • Agent subscriptions [supported] Each harness runs with its installed CLI normal login, for example Claude Code and Codex subscriptions; Add account signs another Claude or Codex account in through the CLI own browser login. Sources: pandacode:prov, pandacode:readme.
  • Api accounts [supported] Provider keys for GLM, MiMo, Alibaba Token Plan, DeepSeek and Command Code are stored in secrets.json (0600) or the OS credential store; harness API-key environment variables are honoured. Sources: pandacode:readme, pandacode:prov.
  • Vendor credits [unsupported] README states "PandaCode does not include provider plans or usage"; no PandaCode credits or resold inference found in docs or source. Sources: pandacode:readme.
  • Local models [supported] Custom OpenAI-compatible endpoints (Ollama, FreeLLMAPI, local models) are documented in providers.md; local speech models for voice input. Sources: pandacode:prov.
  • Account switching [supported] Named connections and account pools for Claude Code and Codex: when one hits its usage limit the task moves to the next account and continues the conversation. Sources: pandacode:prov, pandacode:readme.
  • Vendor account requirement [supported] No PandaCode account, license key or telemetry; Connect (optional) needs a GitHub sign-in; a provider login is needed for each harness used. Sources: pandacode:readme, pandacode:remote.
  • Provider restrictions [partial] Docs tell users to pool only accounts they own and that each must be their own subscription; vendor terms for third-party use of subscription logins are not summarised, and Claude-based use depends on Anthropic terms outside the repository. Sources: pandacode:prov.

Sources: pandacode:readme, pandacode:prov, pandacode:remote

Agent cross-checked partial

Fresh install to first task: download the build (or npx pandacode / Homebrew), install Git and at least one coding CLI and sign in with its login, open a project folder, choose harness and model, send. Folders without Git work for chat and tasks; worktrees and PRs need Git and gh. Cloud agents and a headless host have documented setup but are manual.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • managed cloud setup: No PandaCode-operated managed cloud: cloud agents run on machines in the user own provider account that the user connects with API keys (Settings → Cloud).
  • byo vps onboarding: Documented headless setup: install the binary, xvfb, harness logins and a systemd unit, then pandacode serve --sign-in; DigitalOcean, Vultr, Hetzner and AWS machines can be created from the app as cloud-agent targets. Not a one-click VPS installer.
7 atomic answers
  • First task setup [supported] README Install and Use: download or npx/Homebrew, install Git and one coding CLI, sign in with its normal login, open a project folder, New task, pick harness and model. Sources: pandacode:readme.
  • No git [supported] Projectless New chat starts a conversation in its own folder under ~/PandaCode Chats; a project folder without a repository can host tasks (worktree and branch features are Git-only). Sources: pandacode:readme, pandacode:wt.
  • Local git without remote [supported] Worktrees, diffs, staging, commits and local Merge into main work on a local repo; push and PR actions need a remote. Sources: pandacode:readme, pandacode:wt, pandacode:gh.
  • Hosted git [supported] GitHub through the gh CLI: GitHub view, issue import, linked issues and PRs, push and open pull request; projects whose remote is not on GitHub show no GitHub controls. Sources: pandacode:gh.
  • Prerequisite guidance [supported] README lists Git, a coding CLI login, Node 22 and Tauri prerequisites for source builds; Settings → Recommended shows set and unset items and per-harness install notes. Sources: pandacode:readme, pandacode:set.
  • Managed cloud setup [unsupported] No PandaCode-operated managed cloud: cloud agents run on machines in the user own provider account that the user connects with API keys (Settings → Cloud). Sources: pandacode:cloud.
  • Byo vps onboarding [partial] Documented headless setup: install the binary, xvfb, harness logins and a systemd unit, then pandacode serve --sign-in; DigitalOcean, Vultr, Hetzner and AWS machines can be created from the app as cloud-agent targets. Not a one-click VPS installer. Sources: pandacode:head, pandacode:cloud.

Sources: pandacode:readme, pandacode:wt, pandacode:cloud, pandacode:head, pandacode:gh, pandacode:set

Agent cross-checked partial

No public API, SDK or webhooks for third parties. Agents launched by PandaCode get a loopback MCP bridge and Codex dynamic tools with pandacode_* app tools; a headless host and CLI flags exist; bots expose a local web chat and channel webhooks to chat services.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • cli: pandacode serve flags and the npx installer only; no task create, steer or review CLI.
  • api: No public HTTP API: the relay API serves PandaCode Connect devices with encrypted frames and device-scoped auth; the app protocol is internal (docs/remote-access.md App protocol).
  • sdk: No SDK package: packages/ contains only the npx installer.
  • mcp server: A run-local authenticated MCP bridge exposes pandacode_* tools to harnesses PandaCode itself launches (task, worktree, browser, schedule and machine tools); it is not a server for other products to connect to.
  • webhooks: Bots receive inbound webhooks from Slack, Discord, Telegram and WhatsApp; no webhook starts a generic task or reports task events outward.
  • operations: Agents can discover, create, start, follow up, steer, rename and archive tasks, set board stages, create worktrees and message other machines; humans use the app. Not exposed to third parties.
  • stability: No API stability or versioning statement exists for the internal tools; releases ship several times per week.
8 atomic answers
  • Cli [partial] pandacode serve flags and the npx installer only; no task create, steer or review CLI. Sources: pandacode:head.
  • Api [unsupported] No public HTTP API: the relay API serves PandaCode Connect devices with encrypted frames and device-scoped auth; the app protocol is internal (docs/remote-access.md App protocol). Sources: pandacode:remote.
  • Sdk [unsupported] No SDK package: packages/ contains only the npx installer. Sources: pandacode:readme, pandacode:pkg.
  • Mcp server [partial] A run-local authenticated MCP bridge exposes pandacode_* tools to harnesses PandaCode itself launches (task, worktree, browser, schedule and machine tools); it is not a server for other products to connect to. Sources: pandacode:tools.
  • Webhooks [partial] Bots receive inbound webhooks from Slack, Discord, Telegram and WhatsApp; no webhook starts a generic task or reports task events outward. Sources: pandacode:bots.
  • Operations [partial] Agents can discover, create, start, follow up, steer, rename and archive tasks, set board stages, create worktrees and message other machines; humans use the app. Not exposed to third parties. Sources: pandacode:tools, pandacode:mach.
  • Authentication [supported] MCP bridge is per-run authenticated; Connect uses GitHub device flow, per-device keys and end-to-end encryption. Sources: pandacode:tools, pandacode:remote.
  • Stability [unknown] No API stability or versioning statement exists for the internal tools; releases ship several times per week. Sources: pandacode:rel.

Sources: pandacode:tools, pandacode:head, pandacode:bots, pandacode:remote, pandacode:readme, pandacode:pkg, pandacode:mach, pandacode:rel

Agent cross-checked partial

MCP and skills are inherited from each harness; PandaCode adds a skill manager, rules library, a read-only capabilities inventory, harness slash commands and a built-in browser and simulator tools. No PandaCode plugin API.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • mcp client: Harness MCP servers run under each harness config; PandaCode reads and lists MCP server names in the capabilities matrix and injects its own loopback MCP bridge, but does not provide its own general MCP client or config UI.
  • plugins: Claude Code plugins and hooks are listed read-only; PandaCode has add-ons (Jev, Panda Router) but no third-party plugin API.
  • custom tools: Custom tools come from harness MCP and skills; recipes (JSON steps) and pandacode-run.sh scripts are PandaCode-level customisation.
  • authentication: Provider keys and GitHub use the OS credential store or a 0600 secrets file; MCP server authentication is handled by each harness.
  • permission scope: Host tools follow the approval modes; skills can be switched off per project or globally; MCP permissions are the harness own.
7 atomic answers
  • Mcp client [partial] Harness MCP servers run under each harness config; PandaCode reads and lists MCP server names in the capabilities matrix and injects its own loopback MCP bridge, but does not provide its own general MCP client or config UI. Sources: pandacode:cap, pandacode:tools.
  • Plugins [partial] Claude Code plugins and hooks are listed read-only; PandaCode has add-ons (Jev, Panda Router) but no third-party plugin API. Sources: pandacode:cap, pandacode:readme.
  • Custom tools [partial] Custom tools come from harness MCP and skills; recipes (JSON steps) and pandacode-run.sh scripts are PandaCode-level customisation. Sources: pandacode:recipes, pandacode:run.
  • Instructions import [supported] Skills added from skills.sh, GitHub, GitLab or a folder; rules imported once and enabled per harness and scope. Sources: pandacode:skill, pandacode:readme.
  • Authentication [partial] Provider keys and GitHub use the OS credential store or a 0600 secrets file; MCP server authentication is handled by each harness. Sources: pandacode:readme, pandacode:cap.
  • Permission scope [partial] Host tools follow the approval modes; skills can be switched off per project or globally; MCP permissions are the harness own. Sources: pandacode:tools, pandacode:skill.
  • Integration owner [supported] Inherited from harnesses for MCP and hooks; product-managed for skills, rules, GitHub (gh), cloud providers, bots and the browser. Sources: pandacode:cap, pandacode:skill, pandacode:gh.

Sources: pandacode:cap, pandacode:skill, pandacode:tools, pandacode:readme, pandacode:recipes, pandacode:run, pandacode:gh

Agent cross-checked supported

Desktop: Tauri 2 with Rust backend and a plain TypeScript webview UI (system WKWebView, WebView2, WebKitGTK; no Electron). Mobile: Expo React Native. No web client.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds), PandaCode mobile companion (Expo; iPhone TestFlight beta, Android from source) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
4 atomic answers
  • Desktop implementation [supported] Tauri 2 app (ai.pandacode.desktop) with a Rust backend and TypeScript UI in the platform webview; platform-support.md states no Electron, resident Node server or bundled Chromium. Sources: pandacode:tauri, pandacode:pkg, pandacode:platform.
  • Mobile implementation [supported] Expo (React Native) app per mobile/README.md with native Stack headers and SecureStore. Sources: pandacode:mob.
  • Hybrids [supported] Webview UI hybrid with native Rust, AppKit materials on macOS, ConPTY on Windows; mobile is React Native. Sources: pandacode:platform, pandacode:mob.
  • Implementation evidence [supported] Pinned source: package.json, src-tauri/tauri.conf.json, docs/platform-support.md, mobile/README.md. Sources: pandacode:pkg, pandacode:tauri, pandacode:platform, pandacode:mob.

Sources: pandacode:tauri, pandacode:pkg, pandacode:platform, pandacode:mob

Planning and context

Agent cross-checked partial

Tasks are created locally from the composer, board, quick overlay or bots; GitHub issues import as Pending tasks once (no sync) and one issue and one PR can be linked per task. The native board is authoritative; edits to issues require going back to GitHub.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • chat git triggers: Chat channels trigger tasks through bots; GitHub PR and issue rows start review and solve threads; no push or comment event trigger.
  • stable ids: Tasks have local IDs; a linked issue shows #123 chips; no stable cross-system ID mapping beyond one linked issue and one PR per task.
  • sync: Import is a one-shot action with no background synchronization; link states refresh in a batched call on window focus.
  • supported edits: Update task refreshes the title and description of an unstarted task from a changed issue; nothing is written back to the issue except PR text with Closes #N.
9 atomic answers
  • Local creation [supported] Board tasks with title and description queue without starting an agent; New task from the composer, quick launcher and agents. Sources: pandacode:board, pandacode:readme.
  • Prompt intake [supported] Prompt box, quick overlay, bots in Telegram, Slack, Discord, WhatsApp and a local web chat all start ordinary tasks. Sources: pandacode:readme, pandacode:bots.
  • External import [supported] Board Import from GitHub lists up to fifty open issues and turns selected ones into Pending tasks with the issue URL. Sources: pandacode:gh.
  • Chat git triggers [partial] Chat channels trigger tasks through bots; GitHub PR and issue rows start review and solve threads; no push or comment event trigger. Sources: pandacode:bots, pandacode:gh.
  • Authoritative tracker [supported] The PandaCode board (local SQLite) is authoritative for task state; GitHub is only linked. Sources: pandacode:board, pandacode:gh.
  • Stable ids [partial] Tasks have local IDs; a linked issue shows #123 chips; no stable cross-system ID mapping beyond one linked issue and one PR per task. Sources: pandacode:gh, pandacode:board.
  • Sync [unsupported] Import is a one-shot action with no background synchronization; link states refresh in a batched call on window focus. Sources: pandacode:gh.
  • Supported edits [partial] Update task refreshes the title and description of an unstarted task from a changed issue; nothing is written back to the issue except PR text with Closes #N. Sources: pandacode:gh.
  • External roundtrips [supported] Issue edits, labels, comments and review happen on GitHub; the GitHub view links out. Sources: pandacode:gh.

Sources: pandacode:board, pandacode:gh, pandacode:readme, pandacode:bots

Agent cross-checked partial

A task keeps a saved description and the transcript; Plan mode asks the harness for a plan and Codex /goal holds an objective with budget. No separate plan object with acceptance criteria or revision history.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • plans: Plan mode (Claude plan mode, Codex read-only plan) produces a plan inside the transcript; Codex /goal stores a persistent objective; architect teams keep an objective and assignments.
  • acceptance criteria: No acceptance-criteria field on tasks found in docs/task-board.md or the task schema; criteria live in the prompt text.
  • editors: The task title and description are editable before start; goals editable (pause, resume, edit, clear).
  • revisions: Edit forks the thread before a prompt (new task) but no plan revision history is stored.
  • implementation links: Attributed per-turn changes, branch, linked PR and issue chips tie a task to its output.
7 atomic answers
  • Plans [partial] Plan mode (Claude plan mode, Codex read-only plan) produces a plan inside the transcript; Codex /goal stores a persistent objective; architect teams keep an objective and assignments. Sources: pandacode:appr, pandacode:goals, pandacode:arch.
  • Acceptance criteria [unsupported] No acceptance-criteria field on tasks found in docs/task-board.md or the task schema; criteria live in the prompt text. Sources: pandacode:board.
  • Editors [partial] The task title and description are editable before start; goals editable (pause, resume, edit, clear). Sources: pandacode:board, pandacode:goals.
  • Revisions [partial] Edit forks the thread before a prompt (new task) but no plan revision history is stored. Sources: pandacode:rew.
  • Amendment [supported] Steering and queued messages amend a running task; goals can be edited while running. Sources: pandacode:readme, pandacode:goals.
  • Restart persistence [supported] Tasks, conversations and provider session IDs persist in SQLite across restarts; interrupted runs are marked Interrupted. Sources: pandacode:readme, pandacode:crash.

Sources: pandacode:board, pandacode:appr, pandacode:goals, pandacode:arch, pandacode:rew, pandacode:chg, pandacode:gh, pandacode:readme, pandacode:crash

Agent cross-checked partial

Flat tasks with four board stages; architect teams create worker assignments inside a task and agents can create separate tasks, but there are no epics, arbitrary hierarchy or enforced blockers in the board.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • epics: No epic entity in docs/task-board.md or task tables; an architect team groups workers under one task.
  • hierarchy: Architect and workers form a two-level team inside a task; machine launches record a parent task; no arbitrary depth.
  • blockers: Schedules can start a task when another finishes (after trigger); no blocker field between board tasks.
  • readiness: No readiness computation on the board; Pending means not started.
  • enforcement: after schedules wait for a watched task; dependency ordering is otherwise left to the architect prompt.
  • state propagation: Architect summaries roll up worker changes; no parent or child status propagation on the board.
  • task to epic: A task cannot be promoted to an epic; users switch to Team or start separate tasks.
  • identity retention: Turning Team on keeps the same task and transcript; Codex requires a conversation transfer that replaces its native ID.
9 atomic answers
  • Epics [unsupported] No epic entity in docs/task-board.md or task tables; an architect team groups workers under one task. Sources: pandacode:board, pandacode:arch.
  • Hierarchy [partial] Architect and workers form a two-level team inside a task; machine launches record a parent task; no arbitrary depth. Sources: pandacode:arch, pandacode:mach.
  • Blockers [partial] Schedules can start a task when another finishes (after trigger); no blocker field between board tasks. Sources: pandacode:sched.
  • Readiness [unsupported] No readiness computation on the board; Pending means not started. Sources: pandacode:board.
  • Enforcement [partial] after schedules wait for a watched task; dependency ordering is otherwise left to the architect prompt. Sources: pandacode:sched, pandacode:arch.
  • State propagation [partial] Architect summaries roll up worker changes; no parent or child status propagation on the board. Sources: pandacode:chg, pandacode:arch.
  • Task to epic [unsupported] A task cannot be promoted to an epic; users switch to Team or start separate tasks. Sources: pandacode:board, pandacode:arch.
  • Identity retention [partial] Turning Team on keeps the same task and transcript; Codex requires a conversation transfer that replaces its native ID. Sources: pandacode:prov, pandacode:arch.
  • Workflow constraints [supported] No mandatory workflow: single agent, automatic team or classic architect and workers are chosen per task. Sources: pandacode:arch, pandacode:readme.

Sources: pandacode:board, pandacode:arch, pandacode:tools, pandacode:mach, pandacode:sched, pandacode:chg, pandacode:prov, pandacode:readme

Agent cross-checked partial

Per-project instructions and skills come from the harnesses (AGENTS.md, CLAUDE.md, skills), plus a PandaCode rules library, skill manager, bot memory and a compact app-tool instruction block. Switching harnesses or colleagues carries bounded text history only.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • project docs: Project docs are read by the agent from the checkout; moodboards store per-project reference images; no managed docs store.
  • memory: Bots keep editable bot_memories; coding tasks rely on harness memory and transcripts, with a recovered 200-message history after harness switches.
  • provenance: Bot memory is written by the bot or user and shown in the app; harness memory provenance is not tracked.
  • freshness: No stale-context controls for repository instructions or harness memory found in docs/ or src-tauri/.
  • takeover continuity: A new session receives bounded transcript history (newest 200 messages, 64 KiB); tool state and private reasoning do not transfer.
10 atomic answers
  • Instructions [supported] Repository instruction files are read by the harnesses; PandaCode imports rules once in Rules & library and turns them on per harness and scope. Sources: pandacode:skill, pandacode:readme.
  • Skills [supported] Skill manager lists and switches skills for all projects or one project and adds skills from skills.sh, GitHub, GitLab or a folder. Sources: pandacode:skill.
  • Project docs [partial] Project docs are read by the agent from the checkout; moodboards store per-project reference images; no managed docs store. Sources: pandacode:readme, pandacode:tools.
  • Memory [partial] Bots keep editable bot_memories; coding tasks rely on harness memory and transcripts, with a recovered 200-message history after harness switches. Sources: pandacode:bots, pandacode:prov.
  • Scope [supported] Rules and skills switch per harness and for all projects or one; bot memory is per bot. Sources: pandacode:skill, pandacode:bots.
  • Provenance [partial] Bot memory is written by the bot or user and shown in the app; harness memory provenance is not tracked. Sources: pandacode:bots.
  • Correction [supported] Bot memories can be edited and deleted in the app; skills can be turned off. Sources: pandacode:bots, pandacode:skill.
  • Deletion [supported] Reset bot deletes conversations, memories and schedules together; skills uninstall with Undo. Sources: pandacode:bots, pandacode:skill.
  • Freshness [unsupported] No stale-context controls for repository instructions or harness memory found in docs/ or src-tauri/. Sources: pandacode:prov.
  • Takeover continuity [partial] A new session receives bounded transcript history (newest 200 messages, 64 KiB); tool state and private reasoning do not transfer. Sources: pandacode:prov.

Sources: pandacode:skill, pandacode:prov, pandacode:bots, pandacode:cap, pandacode:readme, pandacode:tools

Agent cross-checked partial

Code search uses harness tools plus an optional local codedb CLI tool; no PandaCode index. Images, PDFs, documents, folders, moodboards and mockups can be attached; GitHub issue and PR context is pulled in with #N.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • code search: pandacode_code_search runs the third-party codedb CLI when installed (bounded, read-only); otherwise harness search tools.
  • indexing: No PandaCode-resident index or daemon: codedb keeps its own index; docs state no resident index service.
  • external knowledge: GitHub issues and PRs referenced with #N are attached as context (4 KiB excerpt); the built-in browser and harness web tools can fetch pages.
  • freshness: GitHub link states refresh on focus at most every five minutes; codedb index freshness belongs to that tool.
  • api retrieval: No API or architecture-doc retrieval service found in src-tauri/src or docs.
  • citations: No retrieval citations are produced by PandaCode.
  • conflict detection: No architecture conflict detection found in the repository.
  • review enforcement: No architecture or API consistency check at review time; only optional project checks after turns.
10 atomic answers
  • Indexing [unsupported] No PandaCode-resident index or daemon: codedb keeps its own index; docs state no resident index service. Sources: pandacode:tools.
  • External knowledge [partial] GitHub issues and PRs referenced with #N are attached as context (4 KiB excerpt); the built-in browser and harness web tools can fetch pages. Sources: pandacode:gh, pandacode:tools.
  • Multimodal inputs [supported] Attachments: images, PDFs, documents, folders, a mockup sketcher, moodboards and up to four screenshots per message; PDF, Word and spreadsheet viewers. Sources: pandacode:readme.
  • Freshness [partial] GitHub link states refresh on focus at most every five minutes; codedb index freshness belongs to that tool. Sources: pandacode:gh, pandacode:tools.
  • Access scope [supported] Task operations stay inside the current project; file previews validate paths against the task checkout. Sources: pandacode:tools.
  • Api retrieval [unsupported] No API or architecture-doc retrieval service found in src-tauri/src or docs. Sources: pandacode:tools.
  • Citations [unsupported] No retrieval citations are produced by PandaCode. Sources: pandacode:tools.
  • Conflict detection [unsupported] No architecture conflict detection found in the repository. Sources: pandacode:tools.
  • Review enforcement [unsupported] No architecture or API consistency check at review time; only optional project checks after turns. Sources: pandacode:set.

Sources: pandacode:tools, pandacode:gh, pandacode:readme, pandacode:set

Agent cross-checked partial

The board provides native tasks with title, description, harness, model, stage, linked GitHub issue/PR and comments, plus a sidebar filter. No labels, priorities, assignees, archive or epic views.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • discuss: Right panel shows comments from the user and agent alongside the transcript; single-user app.
  • prioritize: No priority field in docs/task-board.md or the sidebar sections.
  • label: GitHub labels are displayed on imported issues but tasks have no label field.
  • assign: Single local user; each task has a harness and model, not an assignee.
  • archive: The board has no archive: settled tasks stay on the board and in the Settled section.
  • project epic views: Project filter and per-project board; no epic view.
  • task session separation: Tasks are the sessions: one provider session lineage per task with forks creating new tasks.
  • metadata: Harness, model, approval mode, worktree and branch, linked issue and PR, stage.
17 atomic answers
  • Create [supported] Board adds a title and description without starting an agent. Sources: pandacode:board.
  • Edit [supported] Editing available once details have loaded for the selected task. Sources: pandacode:board.
  • Discuss [partial] Right panel shows comments from the user and agent alongside the transcript; single-user app. Sources: pandacode:board.
  • Prioritize [unsupported] No priority field in docs/task-board.md or the sidebar sections. Sources: pandacode:board.
  • Label [unsupported] GitHub labels are displayed on imported issues but tasks have no label field. Sources: pandacode:board, pandacode:gh.
  • Assign [unsupported] Single local user; each task has a harness and model, not an assignee. Sources: pandacode:board.
  • Filter [supported] Board follows a project filter; sidebar sorts Active, Finished, Settled and Pinned. Sources: pandacode:board, pandacode:readme.
  • Archive [unsupported] The board has no archive: settled tasks stay on the board and in the Settled section. Sources: pandacode:board.
  • Reopen [supported] Reopen returns a settled task to Review and clears completion time. Sources: pandacode:board.
  • Project epic views [partial] Project filter and per-project board; no epic view. Sources: pandacode:board.
  • Idea parking [supported] A Pending task saves an idea without starting an agent. Sources: pandacode:board.
  • Explicit dispatch [supported] Start sends the saved description with the selected harness, model and approval policy. Sources: pandacode:board.
  • Task session separation [partial] Tasks are the sessions: one provider session lineage per task with forks creating new tasks. Sources: pandacode:board, pandacode:rew.
  • Metadata [partial] Harness, model, approval mode, worktree and branch, linked issue and PR, stage. Sources: pandacode:board, pandacode:gh, pandacode:wt.
  • Durable states [supported] Four durable stages stored in SQLite. Sources: pandacode:board.
  • External tracker requirement [supported] No external tracker is required; GitHub is optional. Sources: pandacode:gh, pandacode:board.

Sources: pandacode:board, pandacode:gh, pandacode:readme, pandacode:rew, pandacode:wt

Agent cross-checked partial

Agents launched by PandaCode can discover, create, start, follow up, steer, rename and archive tasks and set board stages through pandacode_* tools within the project; changes appear in the same board with the agent identified in comments. No documented issue CLI or claiming model.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • claim: No claim or lock operation on tasks found in agent-tools.md or the task tables.
  • assign: An agent can start a task on a specific harness and model, or launch a thread on a linked machine; no assignee field.
  • dependencies: Agents can schedule after-task wake-ups; no dependency field.
  • children: Machine launches record child threads; workers belong to architect assignments.
  • attribution: Schedules record created_by user or agent and the creating thread; task comments distinguish agent from user.
  • conflicts: Starting or following up the current running task is rejected; no optimistic concurrency on edits.
14 atomic answers
  • Shared ids [supported] Tools use the same task IDs as the board. Sources: pandacode:tools.
  • Read [supported] pandacode_read_task, project and task discovery and subagent listing. Sources: pandacode:tools.
  • Create [supported] Agents create and start tasks and worktrees in the current project. Sources: pandacode:tools.
  • Update [supported] Rename, board stages, archive; steer and queue messages. Sources: pandacode:tools.
  • Claim [unsupported] No claim or lock operation on tasks found in agent-tools.md or the task tables. Sources: pandacode:tools.
  • Assign [partial] An agent can start a task on a specific harness and model, or launch a thread on a linked machine; no assignee field. Sources: pandacode:tools, pandacode:mach.
  • State changes [supported] Board stages can be set by agents. Sources: pandacode:tools.
  • Dependencies [partial] Agents can schedule after-task wake-ups; no dependency field. Sources: pandacode:sched.
  • Children [partial] Machine launches record child threads; workers belong to architect assignments. Sources: pandacode:mach, pandacode:arch.
  • Results [supported] Followed threads return final replies; machine_wait returns each child final reply. Sources: pandacode:mach, pandacode:tools.
  • Human visibility [supported] Agent-created tasks are ordinary tasks in the sidebar and board. Sources: pandacode:tools, pandacode:mach.
  • Attribution [partial] Schedules record created_by user or agent and the creating thread; task comments distinguish agent from user. Sources: pandacode:sched, pandacode:board.
  • Permissions [supported] Host tools ask for approval except in Full access; Plan mode makes them unavailable; project scoped. Sources: pandacode:tools.
  • Conflicts [partial] Starting or following up the current running task is rejected; no optimistic concurrency on edits. Sources: pandacode:tools.

Sources: pandacode:tools, pandacode:board, pandacode:mach, pandacode:sched, pandacode:arch

Agent cross-checked unsupported

No living product specification feature: no persistent requirements store, revisions, approvals or drift detection beyond repository files the agents edit.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • persistent requirements: No spec or requirements object in docs/ or src-tauri/ (bounded search of task, board and rules docs).
  • human edits: No requirements editor.
  • agent edits: Agents can edit repository files but PandaCode has no spec store.
  • revisions: No spec revisions.
  • approvals: No spec approval flow.
  • implementation links: No requirement-to-task links.
  • later retrieval: No spec retrieval.
  • drift detection: No drift detection.
  • review integration: Review shows diffs only.
9 atomic answers
  • Persistent requirements [unsupported] No spec or requirements object in docs/ or src-tauri/ (bounded search of task, board and rules docs). Sources: pandacode:board, pandacode:readme.
  • Human edits [unsupported] No requirements editor. Sources: pandacode:board.
  • Agent edits [unsupported] Agents can edit repository files but PandaCode has no spec store. Sources: pandacode:tools.
  • Revisions [unsupported] No spec revisions. Sources: pandacode:board.
  • Approvals [unsupported] No spec approval flow. Sources: pandacode:board.
  • Later retrieval [unsupported] No spec retrieval. Sources: pandacode:tools.
  • Drift detection [unsupported] No drift detection. Sources: pandacode:tools.
  • Review integration [unsupported] Review shows diffs only. Sources: pandacode:chg.

Sources: pandacode:board, pandacode:tools, pandacode:readme, pandacode:chg

Coordination and automation

COORD-01

Parallel work

Agent cross-checked partial

Parallel tasks across projects and harnesses run concurrently (README), tiled up to six threads; docs mention a four-agent concurrency limit shared by goals and the app admission path. Architect workers run unlimited in parallel unless maxParallel is set.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • quotas: goals doc and validation notes refer to an existing four-agent limit for app-started runs; the code constant was not located in a bounded search of lib.rs, task_controls.rs and storage.rs, so the number is documented, not source-confirmed. Worker parallelism is a user setting (maxParallel, 0 = unlimited).
  • observed concurrency: No concurrency measurements by this catalog; ADEMark figures are not used.
5 atomic answers
  • Concurrent sessions [supported] README: parallel tasks across projects; up to six threads tiled side by side with live updates. Sources: pandacode:readme.
  • Concurrent tasks [supported] Architect workers run at once with no limit unless the user sets one; machine orchestration spreads threads across linked computers. Sources: pandacode:arch, pandacode:mach.
  • Quotas [partial] goals doc and validation notes refer to an existing four-agent limit for app-started runs; the code constant was not located in a bounded search of lib.rs, task_controls.rs and storage.rs, so the number is documented, not source-confirmed. Worker parallelism is a user setting (maxParallel, 0 = unlimited). Sources: pandacode:goals, pandacode:arch.
  • Observed concurrency [unknown] No concurrency measurements by this catalog; ADEMark figures are not used. Sources: pandacode:readme.
  • Status visibility [supported] Sidebar sections Active, Finished and Settled, board columns, running agents widget, menu bar item listing running, waiting and finished tasks, Agents flyout for workers. Sources: pandacode:readme, pandacode:arch.

Sources: pandacode:readme, pandacode:goals, pandacode:arch, pandacode:mach

Agent cross-checked partial

Tasks are assigned to a harness and model, not to people. Queues exist per task (message queue) and for workers; no claims, assignees or routing beyond optional Panda Router model choice. Guards stop starting or following up a task that is already running.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • human assignment: Single-user app; tasks have no assignee field.
  • claims: No claim operation.
  • duplicate prevention: Starting or following up the current running task is rejected; issue import marks already-linked issues On board.
  • ownership enforcement: Agents can change only threads their own task launched on other machines; no general ownership enforcement.
7 atomic answers
  • Human assignment [unsupported] Single-user app; tasks have no assignee field. Sources: pandacode:board.
  • Agent assignment [supported] Each task or worker assignment names a harness and model; architect assigns work to workers; machine_launch targets a computer. Sources: pandacode:arch, pandacode:mach.
  • Claims [unsupported] No claim operation. Sources: pandacode:tools.
  • Queues [supported] Per-task message queue delivers follow-ups after the run; headless host drains queues natively. Sources: pandacode:head, pandacode:tools.
  • Routing [supported] Panda Router picks a worker model locally (Cheap, Balanced, Best); machine auto picks the online computer with the fewest running threads. Sources: pandacode:route, pandacode:mach.
  • Duplicate prevention [partial] Starting or following up the current running task is rejected; issue import marks already-linked issues On board. Sources: pandacode:tools, pandacode:gh.
  • Ownership enforcement [partial] Agents can change only threads their own task launched on other machines; no general ownership enforcement. Sources: pandacode:mach.

Sources: pandacode:board, pandacode:tools, pandacode:arch, pandacode:route, pandacode:mach, pandacode:head, pandacode:gh

Agent cross-checked partial

Architect and workers (21 harnesses, mixed models and efforts) are a product-owned coordinator with live worker rows, saved worker conversations, corrections and retained results; agents can also launch and wait on threads on linked computers. Provider-native subagents are displayed. Workers always run full access.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • invocation: Natural-language only through the lead prompt and Team menu; no CLI or API invocation.
12 atomic answers
  • Roles [supported] Architect (lead) and workers; named team profiles; reviewer lenses in dispatch. Sources: pandacode:arch.
  • Model harness effort [supported] Each worker profile sets harness, model and effort; Panda Router can choose per assignment. Sources: pandacode:arch, pandacode:route.
  • Messages [supported] Architect sends corrections to existing worker contexts; shared notes board; machine_message steers or sends to launched threads. Sources: pandacode:arch, pandacode:mach.
  • Results return [supported] Worker results return to the architect, which reviews and summarises; machine_wait returns each child final reply up to 8 KB. Sources: pandacode:arch, pandacode:mach.
  • Cross harness [supported] The architect and each worker may use different harnesses. Sources: pandacode:arch.
  • Orchestration owner [supported] Product orchestration in architect.rs; provider subagents are shown separately in Subagent activity. Sources: pandacode:arch, pandacode:prov.
  • Activation [supported] Team button (automatic) or Classic team setup; machine orchestration off by default and needs sign-in and links. Sources: pandacode:arch, pandacode:mach.
  • Authentication [supported] Workers use the lead harness logins; machine tools require Connect sign-in and per-machine link access. Sources: pandacode:arch, pandacode:mach.
  • Live progress [supported] Workers list with tally and per-worker streamed response, tools and conversation. Sources: pandacode:arch.
  • Retained results [supported] Saved worker conversations and attributed changes roll up into the architect summary. Sources: pandacode:arch, pandacode:chg.
  • Followup [supported] Corrections to an existing worker context; failed workers can be replaced. Sources: pandacode:arch.
  • Invocation [partial] Natural-language only through the lead prompt and Team menu; no CLI or API invocation. Sources: pandacode:arch.

Sources: pandacode:arch, pandacode:mach, pandacode:tools, pandacode:route, pandacode:prov, pandacode:chg

Agent cross-checked partial

Built-in scheduler (once, interval, daily, weekly, after another task finishes) that sends ordinary messages to a task; agents can create wake-ups; bots have schedules and heartbeats. Runs only while PandaCode (desktop or headless host) is running; no external event or webhook triggers for coding tasks.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • events: after another task finishes is the only event trigger; bots react to channel messages.
  • completion conditions: max runs and onlyOnSuccess for after triggers; no goal-based completion condition.
  • retries: Three consecutive delivery failures turn a schedule off; no retry backoff policy.
  • client independence: Schedules run while the app process runs, including on a headless host; they stop when the desktop app is quit and fire once on the next start.
8 atomic answers
  • Schedules [supported] Once, interval (5 to 10080 min), daily, weekly. Sources: pandacode:sched.
  • Events [partial] after another task finishes is the only event trigger; bots react to channel messages. Sources: pandacode:sched, pandacode:bots.
  • Recurring tasks [supported] Recurring schedules with max runs up to 1000; agent schedules stop after 20 runs unless set. Sources: pandacode:sched.
  • Completion conditions [partial] max runs and onlyOnSuccess for after triggers; no goal-based completion condition. Sources: pandacode:sched.
  • Scheduler location [supported] Native clock loop inside the app every 15 s; also in the headless host. Sources: pandacode:sched, pandacode:head.
  • Overlaps [supported] Overdue runs after restart fire once; busy threads queue the message; an after schedule runs at most once per 30 s; loop detection refuses cycles. Sources: pandacode:sched.
  • Retries [partial] Three consecutive delivery failures turn a schedule off; no retry backoff policy. Sources: pandacode:sched.
  • Client independence [partial] Schedules run while the app process runs, including on a headless host; they stop when the desktop app is quit and fire once on the next start. Sources: pandacode:sched, pandacode:head.

Sources: pandacode:sched, pandacode:bots, pandacode:head

Agent cross-checked partial

Stop, steer while running, queue messages, approve from desktop or phone, pending-decision bar with Approve all, usage-limit holds and resume; Goals pause and resume. Errors stay visible with retry; terminal-only harness questions open the embedded console.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • pause: Codex goals pause and resume; schedules have a global Paused switch; no generic task pause.
  • notifications: Menu bar item, overlay, running-agents widget, phone push alerts; no desktop OS notifications found.
  • escalation: Pending approvals surface on phone and in the menu bar; no escalation policy.
  • chat terminal parity: Commands that belong to a harness terminal UI open the real harness in an embedded terminal and are not passed as prompts.
10 atomic answers
  • Pause [partial] Codex goals pause and resume; schedules have a global Paused switch; no generic task pause. Sources: pandacode:goals, pandacode:sched.
  • Stop [supported] Stop ends the turn and keeps history; cloud runs stop the remote process group. Sources: pandacode:cloud, pandacode:board.
  • Resume [supported] Follow-up messages from Review start another run; interrupted apps recover; goals resume deliberately. Sources: pandacode:board, pandacode:goals.
  • Steer [supported] Steering and message queueing while an agent runs. Sources: pandacode:readme.
  • Notifications [partial] Menu bar item, overlay, running-agents widget, phone push alerts; no desktop OS notifications found. Sources: pandacode:readme, pandacode:mob.
  • Pending decisions [supported] Requests wait in a slim bar above the prompt, grouped with Approve all; structured questions need an answer. Sources: pandacode:readme, pandacode:appr.
  • Escalation [partial] Pending approvals surface on phone and in the menu bar; no escalation policy. Sources: pandacode:mob, pandacode:readme.
  • Visible errors [supported] Errors and stops remain In progress with a retry action; crash notice; cloud stage and detail. Sources: pandacode:board, pandacode:crash, pandacode:cloud.
  • Chat terminal parity [partial] Commands that belong to a harness terminal UI open the real harness in an embedded terminal and are not passed as prompts. Sources: pandacode:goals.
  • Retained history [supported] Transcripts persist in SQLite; schedule run history keeps 200 runs per task. Sources: pandacode:readme, pandacode:sched.

Sources: pandacode:readme, pandacode:board, pandacode:goals, pandacode:mob, pandacode:appr, pandacode:sched, pandacode:cloud, pandacode:crash

Agent cross-checked partial

Board and sidebar show every task by stage with harness and linked issue/PR icons and an Agents flyout for team workers; no epic, dependency or workload views, and no fixture-based fleet mode.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • fleet fixture: This catalog does not run products; no fleet fixture exists to check.
  • parent child: Team workers are listed under their architect; machine children recorded under the launching task.
  • dependencies: No dependency display on the board.
  • attention: Sidebar groups Active, Finished, Settled; menu bar lists waiting tasks; no explicit blocked or failed column.
  • workload: No workload view per harness or person.
  • navigation errors: Errors stay In progress with retry.
  • fixture mode: No fixture mode applies because the catalog performs no hands-on tests.
8 atomic answers
  • Fleet fixture [unknown] This catalog does not run products; no fleet fixture exists to check. Sources: pandacode:board.
  • Parent child [partial] Team workers are listed under their architect; machine children recorded under the launching task. Sources: pandacode:arch, pandacode:mach.
  • Dependencies [unsupported] No dependency display on the board. Sources: pandacode:board.
  • Attention [partial] Sidebar groups Active, Finished, Settled; menu bar lists waiting tasks; no explicit blocked or failed column. Sources: pandacode:readme, pandacode:board.
  • History [supported] Board cards and sidebar persist history and settled tasks. Sources: pandacode:board.
  • Workload [unsupported] No workload view per harness or person. Sources: pandacode:board.
  • Navigation errors [partial] Errors stay In progress with retry. Sources: pandacode:board.
  • Fixture mode [unknown] No fixture mode applies because the catalog performs no hands-on tests. Sources: pandacode:board.

Sources: pandacode:board, pandacode:arch, pandacode:readme, pandacode:mach

Agent cross-checked partial

Flexible: no mandatory preset; single agent, automatic team (Cheap, Balanced, Best) or classic architect and workers; natural-language instructions; per-worker model and effort; model routing and retries during the run; worker rows and histories remain after completion. Discoverable from the Team button.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • growing epic: Switching Team on in a task keeps the transcript but is not an epic; Codex needs a conversation transfer.
  • in run changes: Corrections to existing workers, replacement workers with escalation (at most 2 retries); changing team settings mid-run is guarded.
  • auth friction: Needs each harness logged in; workers run full access under the lead authority.
10 atomic answers
  • Quick task [supported] A plain task needs only a prompt and harness. Sources: pandacode:readme.
  • Growing epic [partial] Switching Team on in a task keeps the transcript but is not an epic; Codex needs a conversation transfer. Sources: pandacode:prov, pandacode:arch.
  • Presets [supported] Automatic teams with Cheap, Balanced, Best; named team profiles; Classic setup. Sources: pandacode:arch, pandacode:route.
  • Natural language [supported] The lead plans and delegates from the prompt; Lead guidance adds instructions. Sources: pandacode:arch.
  • Agent configuration [supported] Per-worker harness, model, effort; maxParallel. Sources: pandacode:arch.
  • In run changes [partial] Corrections to existing workers, replacement workers with escalation (at most 2 retries); changing team settings mid-run is guarded. Sources: pandacode:arch.
  • Live roles [supported] Workers list and Agents flyout. Sources: pandacode:arch, pandacode:readme.
  • Retained relationships [supported] Worker conversations and summaries are saved with the team task. Sources: pandacode:arch, pandacode:chg.
  • Discoverability [supported] Team button in the composer and Settings → Agent teams. Sources: pandacode:arch, pandacode:set.
  • Auth friction [partial] Needs each harness logged in; workers run full access under the lead authority. Sources: pandacode:arch, pandacode:prov.

Sources: pandacode:arch, pandacode:route, pandacode:readme, pandacode:prov, pandacode:chg, pandacode:set

Execution and reliability

Agent cross-checked partial

Default is local execution. Cloud agents run on machines in the user own account (Daytona, Docker, DigitalOcean, Vultr, Hetzner, RunPod, Cloudflare, AWS, Google Cloud, E2B, Modal and others) and keep going while the laptop sleeps; pandacode serve hosts an always-on machine; Connect continues from other devices. No PandaCode-managed cloud.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: local, own-server · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • managed cloud: Cloud agents run in the user own cloud accounts; the vendor operates only the Connect relay, which stores no task content.
  • resources: Provider defaults (for example Hetzner CPX22, RunPod RTX 3090, Docker ubuntu:24.04 image); no PandaCode resource sizing UI beyond provider options.
  • region: Provider options region or location with defaults (nyc1, ewr, fsn1).
  • expiry: keep_alive capped by each provider max unattended minutes (InstaBox at most 60); idle machines stop 15 minutes after a run; non-persistent machines are recreated and rehydrated.
12 atomic answers
  • Local [supported] Default: agents run on this computer. Sources: pandacode:readme.
  • Byo remote [supported] Cloud agents on the user own provider account and a headless host the user runs on a server or VPS. Sources: pandacode:cloud, pandacode:head.
  • Self hosted [supported] The whole stack can run on user infrastructure: pandacode serve hosts, and relay/README.md documents deploying the Connect relay to the user own Cloudflare account (wrangler deploy) and building the app with PANDACODE_RELAY_URL. Sources: pandacode:relay, pandacode:head.
  • Managed cloud [unsupported] Cloud agents run in the user own cloud accounts; the vendor operates only the Connect relay, which stores no task content. Sources: pandacode:cloud, pandacode:remote.
  • Hybrid [supported] Mirrors in the local sidebar with execution on another computer, headless host or cloud machine. Sources: pandacode:remote, pandacode:cloud.
  • Provisioning [supported] Each cloud task creates a machine named pc-<task id> and installs git, Node and the harness; cloud harness list claude, glm, mimo, codex, opencode. Sources: pandacode:cloud.
  • Payer [supported] The user pays the cloud provider; PandaCode charges nothing. Sources: pandacode:cloud, pandacode:readme.
  • Resources [partial] Provider defaults (for example Hetzner CPX22, RunPod RTX 3090, Docker ubuntu:24.04 image); no PandaCode resource sizing UI beyond provider options. Sources: pandacode:cloud.
  • Region [partial] Provider options region or location with defaults (nyc1, ewr, fsn1). Sources: pandacode:cloud.
  • Expiry [partial] keep_alive capped by each provider max unattended minutes (InstaBox at most 60); idle machines stop 15 minutes after a run; non-persistent machines are recreated and rehydrated. Sources: pandacode:cloud.
  • Client disconnect [supported] Closing the laptop pauses only the poller; the remote run continues and is re-attached at startup; quitting PandaCode leaves the remote run alone. Sources: pandacode:cloud.
  • Other device continuation [supported] A task on a host can be followed and steered from another signed-in computer or the phone through Connect. Sources: pandacode:remote, pandacode:mob.

Sources: pandacode:cloud, pandacode:head, pandacode:remote, pandacode:readme, pandacode:relay, pandacode:mob

Agent cross-checked partial

Per-task git worktree on a pandacode/ branch is the isolation unit; cloud agents use separate machines or containers; Tart VMs and Docker are optional machines; harness sandboxes apply per approval mode. Worktrees are not security sandboxes and Full access is the default.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • container: Docker, Podman, Apple Containers and LXD containers are cloud-agent targets, one container per task; local tasks do not use containers.
  • vm: Local headless macOS and Linux VMs on Apple silicon (Tart) and cloud VMs are available as machines agents can drive; local coding tasks are not run in a VM by default.
  • os sandbox: Harness sandboxes apply per mode (Codex workspace sandbox in Standard, danger-full-access in Full); PandaCode adds none of its own.
  • filesystem: Tasks use their checkout; PandaCode validates paths for its tools; harnesses can read elsewhere in Full access.
  • process: No process isolation between tasks on the local machine beyond separate harness processes.
  • network: Codex Standard mode sandbox limits network; otherwise unrestricted.
  • secrets: Provider secrets stored in a 0600 file or OS credential store and sent only to the provider; harness logins are inherited by processes.
  • security boundary: chg doc states separate worktrees are needed for absolute isolation; approvals doc states Full access runs with no sandbox; no hostile-code containment claim.
12 atomic answers
  • Checkout [supported] Tasks default to Local (main checkout) or Worktree. Sources: pandacode:wt.
  • Branch [supported] Worktree branches are named pandacode/<name>. Sources: pandacode:wt.
  • Worktree [supported] Created from the branch chip, removed with guards. Sources: pandacode:wt.
  • Container [partial] Docker, Podman, Apple Containers and LXD containers are cloud-agent targets, one container per task; local tasks do not use containers. Sources: pandacode:cloud.
  • Vm [partial] Local headless macOS and Linux VMs on Apple silicon (Tart) and cloud VMs are available as machines agents can drive; local coding tasks are not run in a VM by default. Sources: pandacode:readme, pandacode:cloud.
  • Os sandbox [partial] Harness sandboxes apply per mode (Codex workspace sandbox in Standard, danger-full-access in Full); PandaCode adds none of its own. Sources: pandacode:appr.
  • Filesystem [partial] Tasks use their checkout; PandaCode validates paths for its tools; harnesses can read elsewhere in Full access. Sources: pandacode:appr, pandacode:tools.
  • Process [unsupported] No process isolation between tasks on the local machine beyond separate harness processes. Sources: pandacode:appr.
  • Network [partial] Codex Standard mode sandbox limits network; otherwise unrestricted. Sources: pandacode:appr.
  • Secrets [partial] Provider secrets stored in a 0600 file or OS credential store and sent only to the provider; harness logins are inherited by processes. Sources: pandacode:readme.
  • Code separation [supported] Separate worktrees per task; attributed per-turn change capture. Sources: pandacode:wt, pandacode:chg.
  • Security boundary [unsupported] chg doc states separate worktrees are needed for absolute isolation; approvals doc states Full access runs with no sandbox; no hostile-code containment claim. Sources: pandacode:chg, pandacode:appr.

Sources: pandacode:wt, pandacode:cloud, pandacode:appr, pandacode:chg, pandacode:readme, pandacode:tools

Agent cross-checked partial

Transcripts, sessions and provider session IDs persist in SQLite; crash guard restarts the app and marks interrupted agents Interrupted; cloud runs re-attach with log replay; schedule and queue survive restarts; context recovery carries 200 messages to a new harness session.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • worker crash: Architect watchdog fails hung workers; stalled architect gets one Continuing the plan turn and a second stall pauses the team; replacement workers are dispatched.
  • expiry: Non-persistent cloud machines are recreated and rehydrated from the mirror worktree and saved harness session files.
  • process reconnect: Cloud runs are re-attached by replaying the log; local harness processes are not reattached after a crash (agents start only for active turns).
  • state reconstruction: Context recovery sends bounded history; goals recover paused.
  • context loss: Private reasoning, tool state and image bytes do not transfer on harness switches; unsent drafts kept 30 days plain text.
  • duplicate actions: Cloud inbox lines already written are not rewritten on re-attach; overdue schedules fire once.
  • manual repair: Worktree missing offers Recreate; failed cloud preparation offers retry; app hang reports exist as issues.
  • multi day continuity: Tasks and cloud mirror worktrees persist; machines idle-stop and rehydrate; no multi-day epic model.
12 atomic answers
  • Client disconnect [supported] Cloud runs continue; mirrors re-subscribe and copy H messages after reconnect. Sources: pandacode:cloud, pandacode:remote.
  • Worker crash [partial] Architect watchdog fails hung workers; stalled architect gets one Continuing the plan turn and a second stall pauses the team; replacement workers are dispatched. Sources: pandacode:arch.
  • Service restart [supported] Crash guard relaunches PandaCode after about 1.5 s, three crashes in ten minutes stop restarts; interrupted agents are marked. Sources: pandacode:crash.
  • Expiry [partial] Non-persistent cloud machines are recreated and rehydrated from the mirror worktree and saved harness session files. Sources: pandacode:cloud.
  • Transcript reload [supported] Conversations persist across restarts. Sources: pandacode:readme.
  • Process reconnect [partial] Cloud runs are re-attached by replaying the log; local harness processes are not reattached after a crash (agents start only for active turns). Sources: pandacode:cloud, pandacode:readme.
  • Task resume [supported] Saved provider session IDs resume the harness session; Review follow-ups start another run. Sources: pandacode:readme, pandacode:board.
  • State reconstruction [partial] Context recovery sends bounded history; goals recover paused. Sources: pandacode:prov, pandacode:goals.
  • Context loss [partial] Private reasoning, tool state and image bytes do not transfer on harness switches; unsent drafts kept 30 days plain text. Sources: pandacode:prov, pandacode:crash.
  • Duplicate actions [partial] Cloud inbox lines already written are not rewritten on re-attach; overdue schedules fire once. Sources: pandacode:cloud, pandacode:sched.
  • Manual repair [partial] Worktree missing offers Recreate; failed cloud preparation offers retry; app hang reports exist as issues. Sources: pandacode:wt, pandacode:cloud.
  • Multi day continuity [partial] Tasks and cloud mirror worktrees persist; machines idle-stop and rehydrate; no multi-day epic model. Sources: pandacode:cloud, pandacode:board.

Sources: pandacode:crash, pandacode:cloud, pandacode:readme, pandacode:prov, pandacode:remote, pandacode:arch, pandacode:board, pandacode:goals, pandacode:sched, pandacode:wt

Agent cross-checked partial

Cloud agents install git, Node and the harness idempotently; pandacode-run.sh scripts start dev servers; local ports are discoverable to agents; no dependency cache, snapshot or port-allocation manager for parallel local worktrees.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • dependencies: Cloud runtime ensure_runtime installs git, Node and the harness; local projects use the user environment and pandacode-run.sh.
  • caches: No dependency cache management found in docs/ or src-tauri/.
  • snapshots: InstaBox machine snapshots and harness session save/restore for cloud machines.
  • services: Run script starts foreground dev servers in a managed Run terminal; Docker/VM machines for remote services.
  • ports: Agent tools list local ports; no port allocation between worktrees.
  • secrets: Cloud runs copy harness logins (0600) and GH_TOKEN; optional provider environment variables.
  • duplicate installs: No deduplication of installs between worktrees.
  • collisions: No port-collision avoidance beyond user scripts.
  • automatic repair: Cloud retry action, Worktree recreate; no automatic dependency repair.
10 atomic answers
  • Dependencies [partial] Cloud runtime ensure_runtime installs git, Node and the harness; local projects use the user environment and pandacode-run.sh. Sources: pandacode:cloud, pandacode:run.
  • Caches [unsupported] No dependency cache management found in docs/ or src-tauri/. Sources: pandacode:run.
  • Snapshots [partial] InstaBox machine snapshots and harness session save/restore for cloud machines. Sources: pandacode:cloud.
  • Services [partial] Run script starts foreground dev servers in a managed Run terminal; Docker/VM machines for remote services. Sources: pandacode:run, pandacode:cloud.
  • Ports [partial] Agent tools list local ports; no port allocation between worktrees. Sources: pandacode:tools.
  • Secrets [partial] Cloud runs copy harness logins (0600) and GH_TOKEN; optional provider environment variables. Sources: pandacode:cloud.
  • Duplicate installs [unsupported] No deduplication of installs between worktrees. Sources: pandacode:run.
  • Collisions [unsupported] No port-collision avoidance beyond user scripts. Sources: pandacode:run.
  • Remote setup [supported] Idempotent runtime setup with version markers on cloud machines. Sources: pandacode:cloud.
  • Automatic repair [partial] Cloud retry action, Worktree recreate; no automatic dependency repair. Sources: pandacode:cloud, pandacode:wt.

Sources: pandacode:cloud, pandacode:run, pandacode:tools, pandacode:wt

Agent cross-checked partial

Limits are mostly the harness and provider own: usage-limit holds and auto-resume with account pools, Codex goal token budgets, two-minute project checks, schedule max runs. No PandaCode turn, token or spend caps for general tasks.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • time limits: Project checks have a two-minute limit; cloud keep_alive capped; no per-task time cap.
  • turn limits: No per-task turn limit found.
  • token limits: Codex /goal supports an optional token budget.
  • spend limits: No spend caps; usage panel is an estimate only.
  • timeouts: Host tools have per-call deadlines (45 s); architect watchdog fails hung workers.
  • retries: Architect retries replaced workers at most twice; schedules disable after three failures; failed tasks show Retry.
  • provider failures: Account pools move to the next account on a usage limit; failed workers move to the same model on another harness; models that keep failing pause.
  • enforcement: Limits are enforced by PandaCode only for its own checks, schedules and watchdogs; spend is external billing.
8 atomic answers
  • Time limits [partial] Project checks have a two-minute limit; cloud keep_alive capped; no per-task time cap. Sources: pandacode:set, pandacode:cloud.
  • Turn limits [unsupported] No per-task turn limit found. Sources: pandacode:prov.
  • Token limits [partial] Codex /goal supports an optional token budget. Sources: pandacode:goals.
  • Spend limits [unsupported] No spend caps; usage panel is an estimate only. Sources: pandacode:usage.
  • Timeouts [partial] Host tools have per-call deadlines (45 s); architect watchdog fails hung workers. Sources: pandacode:mach, pandacode:arch.
  • Retries [partial] Architect retries replaced workers at most twice; schedules disable after three failures; failed tasks show Retry. Sources: pandacode:arch, pandacode:sched, pandacode:board.
  • Provider failures [partial] Account pools move to the next account on a usage limit; failed workers move to the same model on another harness; models that keep failing pause. Sources: pandacode:prov, pandacode:readme.
  • Enforcement [partial] Limits are enforced by PandaCode only for its own checks, schedules and watchdogs; spend is external billing. Sources: pandacode:set, pandacode:usage.

Sources: pandacode:prov, pandacode:goals, pandacode:set, pandacode:sched, pandacode:cloud, pandacode:usage, pandacode:mach, pandacode:arch, pandacode:board, pandacode:readme

Agent cross-checked partial

Transcripts with tool activity, per-turn attributed changes, usage panel with limit meters and estimated API cost for Claude Code and Codex, task states Pending/In progress/Review/Settled, truthful interrupted state after crashes. No traces export or per-agent success statistics.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • logs: Conversation transcripts and crash reports on disk; no central log viewer beyond transcripts.
  • traces: Tool traces appear in turns; no distributed trace export.
  • failure reasons: Errors and stops shown on tasks; crash notice with report.
  • usage export: No usage export found in agent-usage.md.
  • task stats: Task usage recorded locally (task_usage.rs) and Codex goal token usage; cost estimates are per agent and period.
  • project stats: No per-project statistics found.
  • performance method: Performance numbers come only from ADEMark; validation.md contains vendor benchmarks that this catalog does not adopt.
  • responsiveness: The catalog measures no responsiveness.
13 atomic answers
  • Logs [partial] Conversation transcripts and crash reports on disk; no central log viewer beyond transcripts. Sources: pandacode:readme, pandacode:crash.
  • Traces [partial] Tool traces appear in turns; no distributed trace export. Sources: pandacode:readme.
  • Failure reasons [partial] Errors and stops shown on tasks; crash notice with report. Sources: pandacode:board, pandacode:crash.
  • Usage export [unsupported] No usage export found in agent-usage.md. Sources: pandacode:usage.
  • Task stats [partial] Task usage recorded locally (task_usage.rs) and Codex goal token usage; cost estimates are per agent and period. Sources: pandacode:usage, pandacode:goals.
  • Agent stats [supported] Usage panel per agent with limit windows, cost share and calendar. Sources: pandacode:usage.
  • Project stats [unsupported] No per-project statistics found. Sources: pandacode:usage.
  • Time stats [supported] Week, month and year cost and daily chart. Sources: pandacode:usage.
  • Failure status [supported] Failed runs remain In progress with retry; Interrupted state after crash. Sources: pandacode:board, pandacode:crash.
  • Approval status [supported] Waiting status and approval bar. Sources: pandacode:readme.
  • Reconnect consistency [supported] Mirrors take the host status at the next subscription; approval cards follow the host pending list. Sources: pandacode:remote.
  • Performance method [unknown] Performance numbers come only from ADEMark; validation.md contains vendor benchmarks that this catalog does not adopt. Sources: pandacode:usage.
  • Responsiveness [unknown] The catalog measures no responsiveness. Sources: pandacode:usage.

Sources: pandacode:usage, pandacode:chg, pandacode:board, pandacode:crash, pandacode:readme, pandacode:goals, pandacode:remote

Agent cross-checked partial

Usage panel shows estimated API-equivalent cost for Claude Code and Codex from local logs priced from the LiteLLM list, plus account limit meters; other harnesses show Cost not tracked. Estimates are not billed amounts and not broken down per task, child or epic.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • task cost: Cost is shown per agent and period, not per task.
  • session cost: No per-session cost.
  • child cost: Subagent logs count toward cost but are not shown per child.
  • epic cost: No epic entity or cost.
  • harness aggregation: Overview aggregates agents (Claude Code and Codex) in one estimate; others not tracked.
  • tokens: Token and cache counts feed the estimate; Codex goals show token usage.
  • provider reconciliation: No reconciliation with provider invoices.
  • compute cost: Cloud provider charges are not tracked; docs note droplets bill while they exist.
  • refresh delay: Limits refresh at most every five minutes on focus or on demand; prices daily.
  • alerts: Meters turn amber at 70% and red when full; no budget alerts.
  • export: No export.
  • drilldown: Overview to agent tab and model or day breakdown.
14 atomic answers
  • Task cost [unsupported] Cost is shown per agent and period, not per task. Sources: pandacode:usage.
  • Session cost [unsupported] No per-session cost. Sources: pandacode:usage.
  • Child cost [unsupported] Subagent logs count toward cost but are not shown per child. Sources: pandacode:usage.
  • Epic cost [unsupported] No epic entity or cost. Sources: pandacode:usage.
  • Harness aggregation [partial] Overview aggregates agents (Claude Code and Codex) in one estimate; others not tracked. Sources: pandacode:usage.
  • Billed vs estimated [supported] Doc states the estimate answers what tokens would cost through the API, not what a subscription bills. Sources: pandacode:usage.
  • Tokens [partial] Token and cache counts feed the estimate; Codex goals show token usage. Sources: pandacode:usage, pandacode:goals.
  • Subscription allowance [supported] Limit windows (5-hour and weekly) as percentage used with reset times. Sources: pandacode:usage.
  • Provider reconciliation [unsupported] No reconciliation with provider invoices. Sources: pandacode:usage.
  • Compute cost [unsupported] Cloud provider charges are not tracked; docs note droplets bill while they exist. Sources: pandacode:cloud.
  • Refresh delay [partial] Limits refresh at most every five minutes on focus or on demand; prices daily. Sources: pandacode:usage.
  • Alerts [partial] Meters turn amber at 70% and red when full; no budget alerts. Sources: pandacode:usage.
  • Export [unsupported] No export. Sources: pandacode:usage.
  • Drilldown [partial] Overview to agent tab and model or day breakdown. Sources: pandacode:usage.

Sources: pandacode:usage, pandacode:goals, pandacode:cloud

Agent cross-checked partial

No compute-aware scheduling: concurrency is an app admission limit and a worker maxParallel setting; queues are per task; no CPU or RAM limits, host awareness (beyond fewest running threads for machine auto), check reuse or fairness.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • concurrency: Documented four-agent limit; worker maxParallel (0 unlimited).
  • cpu ram: No CPU or RAM limits found.
  • backpressure: Busy threads queue messages; limits hold queues for failed tasks.
  • host awareness: machine auto picks the computer with the fewest running threads.
  • check reuse: No check reuse.
  • check deduplication: No check deduplication.
  • bounded tests: Project checks run with a two-minute limit and skip plan mode.
  • fairness: No fairness policy.
  • recovery: Watchdog and replacement workers.
  • shared limits: Goals share the app admission limit.
12 atomic answers
  • Concurrency [partial] Documented four-agent limit; worker maxParallel (0 unlimited). Sources: pandacode:goals, pandacode:arch.
  • Queues [supported] Per-task message queue and queue drain; schedules coalesce. Sources: pandacode:sched, pandacode:head.
  • Cpu ram [unsupported] No CPU or RAM limits found. Sources: pandacode:arch.
  • Backpressure [partial] Busy threads queue messages; limits hold queues for failed tasks. Sources: pandacode:sched, pandacode:head.
  • Host awareness [partial] machine auto picks the computer with the fewest running threads. Sources: pandacode:mach.
  • Check reuse [unsupported] No check reuse. Sources: pandacode:set.
  • Check deduplication [unsupported] No check deduplication. Sources: pandacode:set.
  • Bounded tests [partial] Project checks run with a two-minute limit and skip plan mode. Sources: pandacode:set.
  • Cancellation [supported] Stop cancels runs; a cancelled call ends host tools. Sources: pandacode:mach, pandacode:cloud.
  • Fairness [unsupported] No fairness policy. Sources: pandacode:arch.
  • Recovery [partial] Watchdog and replacement workers. Sources: pandacode:arch.
  • Shared limits [partial] Goals share the app admission limit. Sources: pandacode:goals.

Sources: pandacode:arch, pandacode:mach, pandacode:set, pandacode:sched, pandacode:goals, pandacode:head, pandacode:cloud

Human collaboration and control

Agent cross-checked partial

Single-user desktop app. Collaboration is limited to the same GitHub account on several devices (view, control or full grant per paired device) and shared bot channels where several chat participants can message one bot. No presence, comments between users or co-prompting ordering.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • visibility: The same account devices see mirrored tasks; no sharing with other users or read-only links.
  • discussion: No multi-user task discussion: board comments are between the user and the agent.
  • comments: No review comments between users.
  • presence: No presence indicators; one socket per device.
  • coprompting: A bot in a shared Slack, Discord or Telegram channel takes messages from several chat participants as separate conversation threads; coding tasks have one user.
  • ordering: Per-task queue orders follow-ups; no multi-user ordering model.
  • conflicts: No conflict handling between users.
  • attribution: Bot activity log records inbound messages and decisions per channel participant.
  • permissions: Per-device View, Control and Full grants.
11 atomic answers
  • Visibility [partial] The same account devices see mirrored tasks; no sharing with other users or read-only links. Sources: pandacode:remote.
  • Discussion [unsupported] No multi-user task discussion: board comments are between the user and the agent. Sources: pandacode:board.
  • Comments [unsupported] No review comments between users. Sources: pandacode:board.
  • Presence [unsupported] No presence indicators; one socket per device. Sources: pandacode:remote.
  • Coprompting [partial] A bot in a shared Slack, Discord or Telegram channel takes messages from several chat participants as separate conversation threads; coding tasks have one user. Sources: pandacode:bots.
  • Ordering [partial] Per-task queue orders follow-ups; no multi-user ordering model. Sources: pandacode:sched.
  • Conflicts [unsupported] No conflict handling between users. Sources: pandacode:remote.
  • Attribution [partial] Bot activity log records inbound messages and decisions per channel participant. Sources: pandacode:bots.
  • Permissions [partial] Per-device View, Control and Full grants. Sources: pandacode:remote.
  • Plan restrictions [supported] No plans; all features free. Sources: pandacode:readme.
  • Owner online [supported] Remote tasks require the host computer to be on and hosting; a headless host stays up. Sources: pandacode:remote, pandacode:head.

Sources: pandacode:remote, pandacode:bots, pandacode:mob, pandacode:board, pandacode:sched, pandacode:readme, pandacode:head

Agent cross-checked partial

No organisations, projects or task roles, no SSO or provisioning. Access control is per paired device (View, Control, Full) and per machine link (read, ask, full), plus bot channel pairing.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • organization roles: No organisation concept in README, docs/ or relay/ (device list per GitHub account only).
  • project roles: No project roles.
  • task roles: No task roles.
  • membership: Devices join an account by GitHub sign-in; up to 10 per account; revoke removes links and tokens.
  • sso: GitHub device flow is identity only; no SAML, OIDC or SCIM in docs or source.
  • provisioning: No provisioning.
9 atomic answers
  • Organization roles [unsupported] No organisation concept in README, docs/ or relay/ (device list per GitHub account only). Sources: pandacode:remote, pandacode:relay.
  • Project roles [unsupported] No project roles. Sources: pandacode:board.
  • Task roles [unsupported] No task roles. Sources: pandacode:board.
  • Membership [partial] Devices join an account by GitHub sign-in; up to 10 per account; revoke removes links and tokens. Sources: pandacode:remote, pandacode:mach.
  • Sso [unsupported] GitHub device flow is identity only; no SAML, OIDC or SCIM in docs or source. Sources: pandacode:remote.
  • Provisioning [unsupported] No provisioning. Sources: pandacode:remote.
  • Unauthorized access [supported] Methods enforce device grants; secret files are refused below a Full grant; relay checks tokens against its own OAuth app. Sources: pandacode:remote, pandacode:relay.
  • Revocation [supported] DELETE device and account endpoints; linked machine grants removed on sign-out. Sources: pandacode:remote, pandacode:mach.
  • Plan scope [supported] No plans or tiers. Sources: pandacode:readme.

Sources: pandacode:remote, pandacode:mach, pandacode:bots, pandacode:relay, pandacode:board, pandacode:readme

Agent cross-checked partial

Per-task permission modes (Standard, Accept edits, Approve for me, Plan, Full access) map to each harness native policy; approvals are answerable from desktop or phone; host tools ask unless Full access; deploy and cloud tools always ask. New tasks default to Full access and team workers always run Full access.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • network: Codex Standard sandbox limits network; Plan mode allows listed web lookups; otherwise harness-defined.
  • secret: Secret-looking paths ask in Plan mode lookups for Claude; remote file reads refuse secret files below Full.
  • approver identity: Single user; phone approvals carry the device identity in Connect.
  • lifetime: allow_once grants only; lasting grants (allow_always) are never granted automatically to workers; Cline or others may keep allow_always in the harness.
  • revoke: Reduced parent mode applies on retry; harness always-allow rules are managed in the harness.
  • tunnels: Bot web chat (127.0.0.1:47321) and the WhatsApp webhook can be exposed through a user-run tunnel (cloudflared, tailscale serve/funnel, SSH); PandaCode does not start or revoke tunnels, and requests still need the channel token or signature.
  • enforcement: Enforced through native flags or ACP modes; Pi through a PandaCode extension; unsupported modes stop the turn; Antigravity policy remains the server own.
14 atomic answers
  • Command [supported] Command approvals through the harness policy per mode, with Approve all for several requests. Sources: pandacode:appr, pandacode:readme.
  • Tool [supported] Host callbacks request approval in Standard and Auto, unavailable in Plan, none in Full. Sources: pandacode:tools.
  • Network [partial] Codex Standard sandbox limits network; Plan mode allows listed web lookups; otherwise harness-defined. Sources: pandacode:appr.
  • Secret [partial] Secret-looking paths ask in Plan mode lookups for Claude; remote file reads refuse secret files below Full. Sources: pandacode:appr, pandacode:remote.
  • Delivery [supported] Push and open pull request, Merge when ready and deploy CLI runs are approval-gated actions. Sources: pandacode:gh, pandacode:readme.
  • Approver identity [partial] Single user; phone approvals carry the device identity in Connect. Sources: pandacode:remote.
  • Lifetime [partial] allow_once grants only; lasting grants (allow_always) are never granted automatically to workers; Cline or others may keep allow_always in the harness. Sources: pandacode:arch, pandacode:appr.
  • Allow [supported] Approve and Approve all. Sources: pandacode:readme.
  • Deny [supported] Decline requests; workers decline lasting-only requests. Sources: pandacode:arch.
  • Revoke [partial] Reduced parent mode applies on retry; harness always-allow rules are managed in the harness. Sources: pandacode:appr.
  • Tunnels [partial] Bot web chat (127.0.0.1:47321) and the WhatsApp webhook can be exposed through a user-run tunnel (cloudflared, tailscale serve/funnel, SSH); PandaCode does not start or revoke tunnels, and requests still need the channel token or signature. Sources: pandacode:bots, pandacode:head.
  • Defaults [supported] Full access is the default for new tasks, board tasks, worktrees and the quick launcher. Sources: pandacode:appr.
  • Safer modes [supported] Standard, Accept edits, Approve for me and Plan modes. Sources: pandacode:appr.
  • Enforcement [partial] Enforced through native flags or ACP modes; Pi through a PandaCode extension; unsupported modes stop the turn; Antigravity policy remains the server own. Sources: pandacode:appr.

Sources: pandacode:appr, pandacode:tools, pandacode:mach, pandacode:arch, pandacode:readme, pandacode:remote, pandacode:gh, pandacode:bots, pandacode:head

Agent cross-checked partial

Local per-task transcripts and attributed change summaries, schedule runs (200 per task) and a bot activity log. Single-user; no audit log, export or integrity model.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • actions: Transcripts record tool calls; schedules record created_by user or agent.
  • approvals: Bot activity records approvals and rule decisions; coding approvals show in the transcript.
  • discussion: Board comments between user and agent.
  • decisions: Bot decisions logged; no decision records for coding tasks.
  • cross user persistence: No other users.
  • export: No transcript or history export found in docs/ or settings.
  • mutability: Users can delete tasks and messages; saved patches are immutable; bot activity is kept.
  • retention: Patch bodies 128 MiB budget, schedule runs 200 per task, drafts 30 days; transcripts have no stated limit.
  • integrity: No integrity or tamper protection.
11 atomic answers
  • Actions [partial] Transcripts record tool calls; schedules record created_by user or agent. Sources: pandacode:sched, pandacode:readme.
  • Approvals [partial] Bot activity records approvals and rule decisions; coding approvals show in the transcript. Sources: pandacode:bots.
  • Edits [supported] Attributed per-turn file changes with saved patches. Sources: pandacode:chg.
  • Discussion [partial] Board comments between user and agent. Sources: pandacode:board.
  • Decisions [partial] Bot decisions logged; no decision records for coding tasks. Sources: pandacode:bots.
  • Cross user persistence [unsupported] No other users. Sources: pandacode:remote.
  • Cross session persistence [supported] SQLite persistence across restarts. Sources: pandacode:readme.
  • Export [unsupported] No transcript or history export found in docs/ or settings. Sources: pandacode:set.
  • Mutability [partial] Users can delete tasks and messages; saved patches are immutable; bot activity is kept. Sources: pandacode:chg, pandacode:bots.
  • Retention [partial] Patch bodies 128 MiB budget, schedule runs 200 per task, drafts 30 days; transcripts have no stated limit. Sources: pandacode:chg, pandacode:sched, pandacode:crash.
  • Integrity [unsupported] No integrity or tamper protection. Sources: pandacode:chg.

Sources: pandacode:chg, pandacode:sched, pandacode:bots, pandacode:readme, pandacode:board, pandacode:remote, pandacode:set, pandacode:crash

Agent cross-checked partial

Review per turn: Changes summary with unified and split diffs, word-level highlighting, Changes inspector with staging, a GitHub view for PR review with Review, Fix failing checks and Address review comments; inline review comments to requests exist only through the agent chat. Phone shows diffs read-only.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • evidence navigation: Files mentioned in responses are clickable; Open diff from the footer; Search tab.
  • comments: Design canvas comments on HTML elements; GitHub PR comments handled by the agent through Address review comments; no inline diff comments.
  • shipping guidance: Push and open pull request and Merge into main guidance with guards.
  • effort: No review-effort estimate or measurement found; the catalog does not time reviews.
  • missed evidence: Partial captures labelled; unreported writes may be omitted.
  • narrow screen: Board panel overlays on narrow windows; mobile is a companion app.
  • roles: No reviewer roles.
13 atomic answers
  • Pending reviews [supported] Board Review stage and GitHub Pull requests tab list work awaiting review. Sources: pandacode:board, pandacode:gh.
  • Change understanding [supported] Turn cards summarise steps; file summaries with added and deleted lines; diff viewer with word-level changes. Sources: pandacode:chg, pandacode:readme.
  • Evidence navigation [partial] Files mentioned in responses are clickable; Open diff from the footer; Search tab. Sources: pandacode:readme, pandacode:chg.
  • Comments [partial] Design canvas comments on HTML elements; GitHub PR comments handled by the agent through Address review comments; no inline diff comments. Sources: pandacode:readme, pandacode:gh.
  • Request changes [supported] Follow-up message from Review starts another run; Address review comments. Sources: pandacode:board, pandacode:gh.
  • Shipping guidance [partial] Push and open pull request and Merge into main guidance with guards. Sources: pandacode:gh, pandacode:wt.
  • Effort [unknown] No review-effort estimate or measurement found; the catalog does not time reviews. Sources: pandacode:chg.
  • Missed evidence [partial] Partial captures labelled; unreported writes may be omitted. Sources: pandacode:chg.
  • Keyboard [supported] Keyboard navigation in sidebar, board and Settings; shortcuts for tiles. Sources: pandacode:readme, pandacode:set.
  • Narrow screen [partial] Board panel overlays on narrow windows; mobile is a companion app. Sources: pandacode:board, pandacode:mob.
  • Context switching [supported] Tiled threads, quick overlay and menu bar. Sources: pandacode:readme.
  • Roles [unsupported] No reviewer roles. Sources: pandacode:board.
  • Retained context [supported] Saved patches and transcripts. Sources: pandacode:chg.

Sources: pandacode:chg, pandacode:gh, pandacode:readme, pandacode:board, pandacode:wt, pandacode:set, pandacode:mob

Agent cross-checked partial

Not a team product: another person cannot take over a task. The same account on another device can continue it through Connect; architect workers can be taken over by the lead. Context transfer between harnesses is bounded text.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • ongoing takeover: Another device on the same GitHub account can steer and approve a running task on the host.
  • decisions: Approval cards follow the host pending list.
  • artifacts: Changes and files read through host methods.
  • permissions: Device grants View, Control and Full limit what the device can do.
  • runtime context: Mirrors cannot touch the local checkout; checkout operations refuse with This task runs on <name>.
  • ownership: No ownership transfer to another user.
  • attribution: No user attribution.
  • context loss: Private reasoning and tool state do not transfer on harness switches.
12 atomic answers
  • Ongoing takeover [partial] Another device on the same GitHub account can steer and approve a running task on the host. Sources: pandacode:remote.
  • Same task [supported] Mirrors are the same task on the host. Sources: pandacode:remote.
  • Session history [supported] Messages are copied from the host after each run and on subscription. Sources: pandacode:remote.
  • Decisions [partial] Approval cards follow the host pending list. Sources: pandacode:remote.
  • Artifacts [partial] Changes and files read through host methods. Sources: pandacode:remote.
  • Permissions [partial] Device grants View, Control and Full limit what the device can do. Sources: pandacode:remote.
  • Runtime context [partial] Mirrors cannot touch the local checkout; checkout operations refuse with This task runs on <name>. Sources: pandacode:remote.
  • Owner disconnect [supported] A host with remote control on keeps running; offline hosts end local mirrored turns as stopped. Sources: pandacode:remote.
  • Ownership [unsupported] No ownership transfer to another user. Sources: pandacode:remote.
  • Attribution [unsupported] No user attribution. Sources: pandacode:remote.
  • Context loss [partial] Private reasoning and tool state do not transfer on harness switches. Sources: pandacode:prov.
  • Harness transfer [supported] A task can change harness with a bounded text transfer. Sources: pandacode:prov.

Sources: pandacode:remote, pandacode:prov, pandacode:arch

Verification and delivery

Agent cross-checked partial

Per-turn diffs and file previews, project documents, browser screenshots and a visual-progress timeline, design canvas artboards, run and preview, simulator tools, project check output, and worker histories are retained locally in SQLite; no gallery or recordings.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • documents: PDF, Word and spreadsheet viewers for files; design canvas artboards; no generated documents store.
  • recordings: No recording feature found in README or docs.
  • test outputs: Turn checks queue an editable fix message on failure; output appears in the transcript.
  • reviewer outputs: Architect reviewer lenses return findings in the team summary.
  • gallery: No artifact gallery.
  • revision links: Per-turn captures; branches and PR links.
  • prompt leakage: No statement on instruction or prompt leakage into artifacts found.
13 atomic answers
  • Diffs [supported] Saved patches per turn with unified and split viewer. Sources: pandacode:chg.
  • Documents [partial] PDF, Word and spreadsheet viewers for files; design canvas artboards; no generated documents store. Sources: pandacode:readme.
  • Screenshots [supported] Visual progress timeline of annotated screenshots and up to four screenshots per message. Sources: pandacode:readme.
  • Recordings [unsupported] No recording feature found in README or docs. Sources: pandacode:readme.
  • Previews [supported] Run starts pandacode-run.sh and opens the preview; HTML opens in the Browser panel. Sources: pandacode:run, pandacode:readme.
  • Test outputs [partial] Turn checks queue an editable fix message on failure; output appears in the transcript. Sources: pandacode:readme, pandacode:set.
  • Reviewer outputs [partial] Architect reviewer lenses return findings in the team summary. Sources: pandacode:arch.
  • Persistence [supported] SQLite with 128 MiB patch budget; summaries survive eviction. Sources: pandacode:chg.
  • Child findings [supported] Worker histories and architect rollup. Sources: pandacode:arch, pandacode:chg.
  • Prompt leakage [unknown] No statement on instruction or prompt leakage into artifacts found. Sources: pandacode:tools.

Sources: pandacode:chg, pandacode:readme, pandacode:run, pandacode:set, pandacode:arch, pandacode:gh, pandacode:tools

Agent cross-checked partial

Optional project checks after clean turns (npm test or cargo check, two-minute limit, failure queues one fix message), architect lead verification of routed results, GitHub Fix failing checks and recipes with check steps. No required-check gates, spec checks or enforced policy.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • lint: Project checks run the test script or cargo check only; lint is not a separate stage.
  • types: cargo check for Rust checkouts; no TypeScript check stage.
  • ci: Fix failing checks and Merge when ready consult GitHub checks through gh.
  • independent review: Architect review of worker results; reviewer lenses; GitHub Review starts a review thread.
  • spec checks: No spec checks.
  • required checks: Checks are opt-in and off by default.
  • failure blocking: A failed check queues a fix message but does not block settling.
  • requirement conflicts: No requirement conflict detection.
  • configured policy: Settings toggle and recipes check steps.
  • enforced policy: No enforced policy; settle implies no commit or merge.
11 atomic answers
  • Tests [supported] Project checks run the npm test script. Sources: pandacode:set.
  • Lint [unsupported] Project checks run the test script or cargo check only; lint is not a separate stage. Sources: pandacode:set.
  • Types [partial] cargo check for Rust checkouts; no TypeScript check stage. Sources: pandacode:set.
  • Ci [partial] Fix failing checks and Merge when ready consult GitHub checks through gh. Sources: pandacode:gh.
  • Independent review [partial] Architect review of worker results; reviewer lenses; GitHub Review starts a review thread. Sources: pandacode:arch, pandacode:gh.
  • Spec checks [unsupported] No spec checks. Sources: pandacode:set.
  • Required checks [unsupported] Checks are opt-in and off by default. Sources: pandacode:set.
  • Failure blocking [partial] A failed check queues a fix message but does not block settling. Sources: pandacode:set, pandacode:board.
  • Requirement conflicts [unsupported] No requirement conflict detection. Sources: pandacode:set.
  • Configured policy [partial] Settings toggle and recipes check steps. Sources: pandacode:set, pandacode:recipes.
  • Enforced policy [unsupported] No enforced policy; settle implies no commit or merge. Sources: pandacode:board.

Sources: pandacode:set, pandacode:arch, pandacode:gh, pandacode:recipes, pandacode:board

Agent cross-checked partial

Git is first-class: per-task worktrees and branches, diffs, staging and commits, local Merge into main, branch switching, push and open pull request through gh, PR review, fix-check and address-comments threads, Jujutsu support. No automatic rebase or conflict resolution UI.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • comments: Address review comments starts a thread; no inline comment UI.
  • rebase: No rebase action found in worktrees.md or github.md.
  • conflicts: Merge into main is local and guarded; conflict handling is left to git and the agent.
9 atomic answers
  • Branches [supported] Branch chip creates, checks out and switches branches. Sources: pandacode:wt, pandacode:readme.
  • Worktrees [supported] Per-task worktrees with guarded cleanup. Sources: pandacode:wt.
  • Commits [supported] Staging and commits in the Changes inspector. Sources: pandacode:readme.
  • Pull requests [supported] Push and open pull request with title from the task and Closes #N. Sources: pandacode:gh.
  • Comments [partial] Address review comments starts a thread; no inline comment UI. Sources: pandacode:gh.
  • Ci response [supported] Fix failing checks thread. Sources: pandacode:gh.
  • Rebase [unsupported] No rebase action found in worktrees.md or github.md. Sources: pandacode:wt, pandacode:gh.
  • Conflicts [partial] Merge into main is local and guarded; conflict handling is left to git and the agent. Sources: pandacode:readme, pandacode:wt.
  • Product state [supported] Chips show #N open and PR #N merged states refreshed on focus. Sources: pandacode:gh.

Sources: pandacode:wt, pandacode:gh, pandacode:readme

Agent cross-checked partial

Humans merge: Merge into main (local) and Merge when ready (GitHub, only when checks pass, approved and clean) are user-triggered actions; deploys to Vercel, Cloudflare, Railway, Netlify, AWS and Google Cloud run as approval-gated tasks. No exact-revision approvals or authority policy.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • approve: PR approval happens on GitHub; Approve for tool actions in PandaCode.
  • release: No release action found.
  • exact revision: No exact-revision binding of approvals.
  • authority enforcement: Merge when ready checks checks, approval and mergeability; no role-based authority.
6 atomic answers
  • Approve [partial] PR approval happens on GitHub; Approve for tool actions in PandaCode. Sources: pandacode:gh.
  • Merge [supported] Local Merge into main and GitHub Merge when ready. Sources: pandacode:gh, pandacode:readme.
  • Release [unsupported] No release action found. Sources: pandacode:readme.
  • Deploy [supported] Start a deploy task from the action dock with approval-gated CLI runs. Sources: pandacode:readme.
  • Exact revision [unsupported] No exact-revision binding of approvals. Sources: pandacode:gh.
  • Authority enforcement [partial] Merge when ready checks checks, approval and mergeability; no role-based authority. Sources: pandacode:gh.

Sources: pandacode:gh, pandacode:readme, pandacode:wt

Agent cross-checked partial

Tasks link to branch, issue and PR with state chips; settling records completion only; no build or deployment links, rollback or revert helpers, and no stale-approval invalidation.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • commit links: Attributed changes and branch are stored; no commit link surface.
  • build links: No build link.
  • deployment links: No deployment links stored.
  • delivery status: PR state chips (open, merged).
  • retained evidence: Patches and transcripts.
  • rollback: No rollback action.
  • revert: No revert action.
  • stale approval invalidation: No approval-to-revision binding.
9 atomic answers
  • Delivery status [partial] PR state chips (open, merged). Sources: pandacode:gh.
  • Retained evidence [partial] Patches and transcripts. Sources: pandacode:chg.
  • Rollback [unsupported] No rollback action. Sources: pandacode:readme.
  • Revert [unsupported] No revert action. Sources: pandacode:readme.
  • Stale approval invalidation [unsupported] No approval-to-revision binding. Sources: pandacode:gh.

Sources: pandacode:gh, pandacode:board, pandacode:chg, pandacode:wt, pandacode:readme

Agent cross-checked partial

High-volume review is limited to the board Review column, GitHub Pull requests tab and architect summaries; no batch approvals, risk filters or dependency queues.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • grouping: Project filter and board columns.
  • risk size filters: PR rows show size and checks; no filter.
  • findings: Reviewer lens findings in team summaries.
  • artifacts: Saved patches per turn.
  • exact revision approval: No exact-revision approval.
  • dependency queues: No dependency queues.
  • next actions: Per-row actions (Review, Fix, Merge when ready).
  • stale approval invalidation: No approvals to invalidate.
  • check deduplication: No check deduplication.
12 atomic answers
  • Review queue [supported] Board Review stage and GitHub PR tab. Sources: pandacode:board, pandacode:gh.
  • Summaries [supported] Turn cards and architect summaries. Sources: pandacode:arch, pandacode:chg.
  • Grouping [partial] Project filter and board columns. Sources: pandacode:board.
  • Risk size filters [partial] PR rows show size and checks; no filter. Sources: pandacode:gh.
  • Findings [partial] Reviewer lens findings in team summaries. Sources: pandacode:arch.
  • Artifacts [partial] Saved patches per turn. Sources: pandacode:chg.
  • Request changes [supported] Follow-up from Review. Sources: pandacode:board.
  • Exact revision approval [unsupported] No exact-revision approval. Sources: pandacode:gh.
  • Dependency queues [unsupported] No dependency queues. Sources: pandacode:board.
  • Next actions [partial] Per-row actions (Review, Fix, Merge when ready). Sources: pandacode:gh.
  • Stale approval invalidation [unsupported] No approvals to invalidate. Sources: pandacode:gh.
  • Check deduplication [unsupported] No check deduplication. Sources: pandacode:set.

Sources: pandacode:board, pandacode:gh, pandacode:arch, pandacode:chg, pandacode:set

Ownership, privacy and economics

Agent cross-checked partial

Open source under the PandaCode License: MIT with an added attribution requirement for forks used commercially or released publicly. Covers desktop, relay and mobile in one repository; source public at release v0.3.129.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • hosted extensions: The vendor Connect relay is hosted by the maintainer, its source is in relay/; Panda Router profiler model is downloaded from a pinned GitHub commit under MIT.
  • hosted parity: The hosted relay is the same code as relay/ (deployed by the maintainer); parity of the live deployment cannot be checked.
6 atomic answers
  • Client license [supported] PandaCode License (MIT plus a fork attribution clause); GitHub reports NOASSERTION. Sources: pandacode:license, pandacode:repo.
  • Server license [supported] relay/ and the headless host live in the same repository under the same license. Sources: pandacode:license, pandacode:relay.
  • Hosted extensions [partial] The vendor Connect relay is hosted by the maintainer, its source is in relay/; Panda Router profiler model is downloaded from a pinned GitHub commit under MIT. Sources: pandacode:relay, pandacode:route.
  • Source visibility [supported] Public repository including docs, tests, relay and mobile. Sources: pandacode:repo.
  • Source revision [supported] Pinned to v0.3.129 (commit f7961a0). Sources: pandacode:rel.
  • Hosted parity [partial] The hosted relay is the same code as relay/ (deployed by the maintainer); parity of the live deployment cannot be checked. Sources: pandacode:relay.

Sources: pandacode:license, pandacode:repo, pandacode:rel, pandacode:relay, pandacode:route

Agent cross-checked partial

Runs locally with local SQLite storage; the only vendor-operated piece is the optional Connect relay, which can be self-deployed. Cloud agents run in the user own accounts. No export or import features for tasks; account deletion removes relay state.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • vendor dependencies: Optional Connect relay and GitHub sign-in; GitHub releases for updates; model providers for inference.
  • export: No export of tasks or conversations found; data lives in SQLite in the app data folder (readable by the user).
  • import: conversation_import reads Claude and Codex conversations (src-tauri/src/conversation_import); recipes and skills import from files.
  • formats: SQLite database and plain-text crash reports; patches in SQLite.
  • history portability: History stays in local SQLite and the harness own session files; no export format.
8 atomic answers
  • Self hosting [supported] Everything runs on user machines; relay deployable with wrangler. Sources: pandacode:relay, pandacode:head, pandacode:readme.
  • Vendor dependencies [partial] Optional Connect relay and GitHub sign-in; GitHub releases for updates; model providers for inference. Sources: pandacode:remote, pandacode:readme.
  • Export [unsupported] No export of tasks or conversations found; data lives in SQLite in the app data folder (readable by the user). Sources: pandacode:readme.
  • Import [partial] conversation_import reads Claude and Codex conversations (src-tauri/src/conversation_import); recipes and skills import from files. Sources: pandacode:skill.
  • Formats [partial] SQLite database and plain-text crash reports; patches in SQLite. Sources: pandacode:readme.
  • Cancellation [supported] No subscription; DELETE /v1/account deletes all relay state. Sources: pandacode:remote.
  • Account independence [supported] No PandaCode account needed. Sources: pandacode:readme.
  • History portability [partial] History stays in local SQLite and the harness own session files; no export format. Sources: pandacode:readme, pandacode:prov.

Sources: pandacode:readme, pandacode:relay, pandacode:head, pandacode:cloud, pandacode:remote, pandacode:skill, pandacode:prov

Agent cross-checked partial

Local-first: projects, conversations and settings in local SQLite; prompts go only to the chosen coding CLI or service; no PandaCode telemetry; one switch disables supported agent telemetry; documented network table; Connect relay sees only encrypted frames and device metadata.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • logs: Crash reports local; drafts unencrypted in web storage up to 30 days.
  • retention: Drafts 30 days; patch budget 128 MiB; no general retention policy for transcripts.
  • training: README: no telemetry and prompts go to the provider under its terms; PandaCode itself does no training.
  • residency: Local storage; relay on Cloudflare Worker without a stated region.
  • subprocessors: Optional services listed in README network table: GitHub, Cloudflare relay, APNs and FCM, cloud providers, OpenRouter, TypeSafe, Hugging Face and others.
  • consent prompt: Connect and cloud agents are off until the user signs in or connects; no telemetry prompt exists because none is collected.
15 atomic answers
  • Code data [supported] Code stays on the machine; cloud agents send the checkout to the chosen provider machine. Sources: pandacode:readme, pandacode:cloud.
  • Prompts [supported] Prompts go to the chosen harness or provider only; Jev decisions optional to OpenRouter or TypeSafe. Sources: pandacode:readme.
  • Credentials [supported] Keys in OS credential store or 0600 secrets.json, sent only to that provider; cloud runs copy harness logins to the machine. Sources: pandacode:readme, pandacode:cloud.
  • Logs [partial] Crash reports local; drafts unencrypted in web storage up to 30 days. Sources: pandacode:crash, pandacode:readme.
  • Artifacts [supported] Patches and attachments local; cloud and bots as configured. Sources: pandacode:chg.
  • Retention [partial] Drafts 30 days; patch budget 128 MiB; no general retention policy for transcripts. Sources: pandacode:crash, pandacode:chg.
  • Deletion [supported] Delete task, reset bot, DELETE /v1/account. Sources: pandacode:remote, pandacode:bots.
  • Training [unsupported] README: no telemetry and prompts go to the provider under its terms; PandaCode itself does no training. Sources: pandacode:readme.
  • Residency [partial] Local storage; relay on Cloudflare Worker without a stated region. Sources: pandacode:relay.
  • Subprocessors [partial] Optional services listed in README network table: GitHub, Cloudflare relay, APNs and FCM, cloud providers, OpenRouter, TypeSafe, Hugging Face and others. Sources: pandacode:readme.
  • Telemetry disclosure [supported] README states no telemetry and lists every automatic network check with frequency. Sources: pandacode:readme.
  • Telemetry defaults [supported] No telemetry by default; version checks every 6 hours, price list daily, updates unless Off. Sources: pandacode:readme.
  • Telemetry controls [supported] Updates Off or Notify me; Turn off agent telemetry in Settings → Agents; Keep unsent drafts switch. Sources: pandacode:readme.
  • Provider separation [supported] Provider keys are sent only to that provider; prompts reach only the selected harness. Sources: pandacode:readme.

Sources: pandacode:readme, pandacode:remote, pandacode:relay, pandacode:cloud, pandacode:crash, pandacode:chg, pandacode:bots

Agent cross-checked supported

Free: no price, subscription, seats, credits or trial. Users pay their coding-agent providers and any cloud machines. Distributed through GitHub; Connect relay free with documented limits.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
12 atomic answers
  • Currency [supported] No price: README states PandaCode is free and open source, no account or license key. Sources: pandacode:readme.
  • Interval [supported] No recurring charge. Sources: pandacode:readme.
  • Seats [supported] No seats. Sources: pandacode:readme.
  • Seat minimum [supported] No minimum. Sources: pandacode:readme.
  • Workspace fees [supported] None. Sources: pandacode:readme.
  • Included usage [supported] No usage included; provider plans are separate. Sources: pandacode:readme.
  • Credits [supported] No credits. Sources: pandacode:readme.
  • Trial [supported] No trial needed. Sources: pandacode:readme.
  • Overages [supported] None; relay limits listed (10 devices). Sources: pandacode:remote.
  • Taxes [supported] Not applicable because nothing is sold. Sources: pandacode:readme.
  • Model cost [supported] Paid to harness providers; PandaCode includes none. Sources: pandacode:readme.
  • Compute cost [supported] Paid to cloud providers for cloud agents. Sources: pandacode:cloud.

Sources: pandacode:readme, pandacode:remote, pandacode:cloud

Agent cross-checked partial

PandaCode adds no cost for one, two or five users; costs are model subscriptions or API usage and optional cloud machines, which PandaCode does not forecast.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • task assumptions: The docs give no scenario assumptions; the catalog defines none.
  • tokens: Estimated API cost panel for Claude Code and Codex.
  • compute hours: Cloud providers bill while machines exist; PandaCode does not total hours.
  • variable cost: Provider usage and cloud machines vary.
  • unknown cost: Subscription marginal cost is not shown; Cost not tracked for other harnesses.
11 atomic answers
  • One person [supported] Free software; user pays harness and cloud costs. Sources: pandacode:readme.
  • Two people [supported] Each user installs and logs in separately; no cost. Sources: pandacode:readme.
  • Five people [supported] No shared plan; no cost. Sources: pandacode:readme.
  • Task assumptions [unknown] The docs give no scenario assumptions; the catalog defines none. Sources: pandacode:usage.
  • Concurrency [supported] No fee tied to concurrency. Sources: pandacode:readme.
  • Tokens [partial] Estimated API cost panel for Claude Code and Codex. Sources: pandacode:usage.
  • Compute hours [partial] Cloud providers bill while machines exist; PandaCode does not total hours. Sources: pandacode:cloud.
  • Fixed cost [supported] Zero. Sources: pandacode:readme.
  • Variable cost [partial] Provider usage and cloud machines vary. Sources: pandacode:cloud.
  • Included cost [supported] None. Sources: pandacode:readme.
  • Unknown cost [partial] Subscription marginal cost is not shown; Cost not tracked for other harnesses. Sources: pandacode:usage.

Sources: pandacode:readme, pandacode:cloud, pandacode:usage

Agent cross-checked partial

Actively released open-source beta (v0.3.x, several releases a week); Windows and Linux are previews; support through GitHub issues; no SLA or status page; deprecation unknown.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • release stage: Version 0.3.x; Windows and Linux previews; iPhone beta.
  • support: GitHub issues and CONTRIBUTING; no support desk found.
  • response commitments: No response commitments in repository docs.
  • status history: No status page found.
  • deprecation: No deprecation policy.
  • export options: No export.
8 atomic answers
  • Release stage [partial] Version 0.3.x; Windows and Linux previews; iPhone beta. Sources: pandacode:platform, pandacode:rel, pandacode:mob.
  • Access restrictions [supported] Public downloads and source; iPhone needs TestFlight. Sources: pandacode:readme.
  • Support [partial] GitHub issues and CONTRIBUTING; no support desk found. Sources: pandacode:repo.
  • Response commitments [unsupported] No response commitments in repository docs. Sources: pandacode:repo.
  • Release activity [supported] v0.3.125, v0.3.126 and v0.3.129 within three days. Sources: pandacode:rel.
  • Status history [unsupported] No status page found. Sources: pandacode:repo.
  • Deprecation [unknown] No deprecation policy. Sources: pandacode:repo.
  • Export options [unsupported] No export. Sources: pandacode:readme.

Sources: pandacode:rel, pandacode:repo, pandacode:platform, pandacode:issues, pandacode:mob, pandacode:readme

No sourced public finding not-published

No publishable sourced finding is available for this criterion. This does not mean the product lacks the capability.

Agent cross-checked partial

The whole app is free with no card, caps or trial; agent subscriptions and cloud costs are separate.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
  • credits trial: No trial credits.
  • paid only: No paid-only features.
  • early access: iPhone TestFlight beta; Windows and Linux previews.
  • usage caps: Relay free limits (10 devices, rate limits); none for local use.
10 atomic answers
  • Ongoing free [supported] Free and open source forever per README. Sources: pandacode:readme.
  • Trial [supported] No trial needed. Sources: pandacode:readme.
  • Credits trial [unsupported] No trial credits. Sources: pandacode:readme.
  • Self host cost [supported] Self-hosting the relay costs the user Cloudflare charges; the app is free. Sources: pandacode:relay.
  • Paid only [unsupported] No paid-only features. Sources: pandacode:readme.
  • Early access [partial] iPhone TestFlight beta; Windows and Linux previews. Sources: pandacode:platform, pandacode:mob.
  • Free scenarios [supported] Local use with any logged-in harness. Sources: pandacode:readme.
  • Card required [supported] No card. Sources: pandacode:readme.
  • Usage caps [partial] Relay free limits (10 devices, rate limits); none for local use. Sources: pandacode:remote.
  • Paid evaluation [supported] No paid evaluation needed. Sources: pandacode:readme.

Sources: pandacode:readme, pandacode:remote, pandacode:relay, pandacode:platform, pandacode:mob

Agent cross-checked supported

Transparent: nothing is sold; provider and cloud costs are outside the product.

Scope
Surfaces: PandaCode desktop app (Tauri 2, Rust and TypeScript; macOS, Windows and Linux builds) · Version: PandaCode release v0.3.129 (tag v0.3.129, commit f7961a0, published 2026-11-16); all public repos of the aicodeking account checked, only aicodeking/pandacode contains the app, relay and mobile client · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
9 atomic answers

Sources: pandacode:readme, pandacode:cloud

Sources

pandacode:rel

PandaCode v0.3.129 release

Release v0.3.129 published 2026-11-16 with installers for macOS arm64 and x64, Windows x64 and arm64, Linux x86_64 and aarch64 AppImage plus checksums.

Kind
documented
Locator
Release page and assets
Source date
2026-11-16
Retrieved
2026-11-16
pandacode:readme

PandaCode README

Describes the desktop workspace for 20+ coding CLIs, board, teams, schedules, bots, Connect, headless server, cloud agents, privacy table (no account, no telemetry), install matrix and requirements.

Kind
source-inspected
Locator
Features, Run it anywhere, Privacy, Install and Use sections
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
README.md
pandacode:pkg

package.json

Version 0.3.129, license field "SEE LICENSE IN LICENSE", scripts for tauri desktop, vite build, node tests and Playwright screenshots; frontend is plain TypeScript with vite.

Kind
source-inspected
Locator
name, version, scripts, license field
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
package.json
pandacode:platform

Desktop platform support

Six release targets (macOS arm64/x64, Windows x64/arm64, Linux x86_64/aarch64 AppImage); Windows and Linux are previews with incomplete interactive validation; uses system webviews, no Electron.

Kind
source-inspected
Locator
Target table, Native runtimes and dependencies
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/platform-support.md
pandacode:board

Project board

Four durable stages Pending, In progress, Review, Settled; tasks saved without starting an agent; comments panel; no archive; settling implies no commit or merge.

Kind
source-inspected
Locator
Whole file
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/task-board.md
pandacode:remote

Remote access (PandaCode Connect)

Opt-in GitHub device-flow sign-in; Cloudflare Worker relay forwards end-to-end encrypted frames; tasks run on another signed-in computer or a headless host; limits are 10 devices per account.

Kind
source-inspected
Locator
Pieces, Mirrors, Accounts, Limits (free), End-to-end encryption, Privacy
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/remote-access.md
pandacode:head

Headless server

pandacode serve runs the same binary without a window as a Connect host; needs a display (xvfb) on Linux; secrets stored in a 0600 file; no updater.

Kind
source-inspected
Locator
What the server runs, First run, Linux server
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/headless.md
pandacode:mob

PandaCode mobile

Expo app for iOS and Android that views and controls tasks on computers through Connect; the phone never runs tasks; build from source.

Kind
source-inspected
Locator
Whole file
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
mobile/README.md
pandacode:arch

Architect and workers

Architect delegates assignments to workers across 21 harnesses; workers run full access in parallel; worker list, retained histories, corrections, model routing, stall watchdog.

Kind
source-inspected
Locator
Whole file
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/architect-workers.md
pandacode:prov

Providers and harnesses

Harness list, sign-in paths, named connections, account pools for Claude Code and Codex, harness switch with conversation transfer, custom OpenAI-compatible endpoints.

Kind
source-inspected
Locator
How providers relate to agents, Named connections, Account pools, Conversation continuity, per-harness sections
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/providers.md
pandacode:goals

Goals and harness command workflows

Codex /goal with token budget and a shared four-agent concurrency limit; harness slash commands.

Kind
source-inspected
Locator
Goals, Command controls
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/goals-and-command-workflows.md
pandacode:wt

Worktrees

Per-task git worktree beside the project on a pandacode/<name> branch; removal guarded; git worktree add/remove run with argument arrays.

Kind
source-inspected
Locator
Creating a worktree, Cleaning up, Native behavior
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/worktrees.md
pandacode:gh

GitHub

Uses gh CLI only; GitHub view with threads, PRs, issues; one-shot issue import to Pending tasks; push and open pull request; no background sync; one issue and one PR linked per task.

Kind
source-inspected
Locator
GitHub view, Importing issues, Linked work, Pull requests
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/github.md
pandacode:cloud

Cloud agents

Tasks can run on machines in the user own account at 20 cloud providers; checkout sent as git bundle; run detached under a runner and re-attached after sleep or quit; supported harnesses claude, glm, mimo, codex, opencode.

Kind
source-inspected
Locator
How it works, State, Provider settings, Privacy
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/cloud-agents.md
pandacode:set

Settings

Settings sections incl. Schedules, Tools, Connections; project checks after turns run the npm test script or cargo check with a two-minute limit.

Kind
source-inspected
Locator
Section table, Harnesses, Providers
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/settings.md
pandacode:mach

Machine orchestration

Off by default; agents can link, launch, message, wait on, read and stop threads on the user other computers with per-machine access levels read, ask and full.

Kind
source-inspected
Locator
Links, Agent tools
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/machine-orchestration.md
pandacode:tools

Agent tools and integrations

pandacode_* app tools for task discovery, create, start, follow up, steer, board stages, worktrees, browser and code search; approval boundary for host tools.

Kind
source-inspected
Locator
App tools, host callbacks
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/agent-tools.md
pandacode:bots

Bots

Persistent named agents reachable from the desktop, a local web chat, Telegram, Slack, Discord and WhatsApp with schedules, memory, approval rules and an activity log.

Kind
source-inspected
Locator
What each product contributes, Architecture
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/bots.md
pandacode:run

Run and preview

Run button starts the checkout pandacode-run.sh in a managed terminal tab and opens a preview.

Kind
source-inspected
Locator
Whole file
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/run-project.md
pandacode:skill

Skill manager

Lists skills installed for any harness, switches them per project or globally, adds skills from skills.sh, GitHub, GitLab or a folder.

Kind
source-inspected
Locator
Behaviour
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/skill-manager.md
pandacode:cap

Capabilities

Read-only inventory of slash commands, MCP server names, hooks, plugin names and skills across harnesses; does not activate integrations.

Kind
source-inspected
Locator
Whole file
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/capabilities.md
pandacode:recipes

Workflow recipes

JSON recipes of prompt, slash, check and approval steps stored per project or globally.

Kind
source-inspected
Locator
Whole file
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/recipes.md
pandacode:tauri

Tauri configuration

Tauri 2 app ai.pandacode.desktop version 0.3.129.

Kind
source-inspected
Locator
productName, version, identifier, bundle
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
src-tauri/tauri.conf.json
pandacode:appr

Approval behaviours

Per-task permission modes (Standard, Accept edits, Approve for me, Plan, Full access) mapped to each harness native flags; new tasks start in Full access; Pi enforced through a PandaCode extension.

Kind
source-inspected
Locator
Mode table, Full access mappings, Plan mode
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/approvals.md
pandacode:chg

Changes after a turn

Attributed per-turn file change summaries with immutable saved patches in SQLite; limits on paths, bytes and retention.

Kind
source-inspected
Locator
What is recorded, Resource limits
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/turn-changes.md
pandacode:rew

Edit and rewind a sent prompt

Edit forks the conversation before a message into a new task; files are not rewound.

Kind
source-inspected
Locator
Whole file
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/rewind.md
pandacode:crash

Crash recovery

Guard process restarts the app after a crash; interrupted agents are marked Interrupted; reports stay local; drafts kept as plain text for up to 30 days.

Kind
source-inspected
Locator
What you will see, Overview
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/crash-recovery.md
pandacode:sched

Task schedules

Native scheduler with once, interval, daily, weekly and after-task triggers; agent tools to schedule wake-ups; clock loop ticks every 15 s inside the app; paused setting; three consecutive failures disable a schedule.

Kind
source-inspected
Locator
Commands, Firing, Agent tools, Settings
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/schedules.md
pandacode:route

Model routing

Local Panda Router (MIT) picks a worker model in about 10-25 ms with no task data leaving the machine.

Kind
source-inspected
Locator
How it works
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/model-routing.md
pandacode:relay

PandaCode Connect relay

Cloudflare Worker with a Durable Object per account; deploy steps with a GitHub OAuth app; stores no task content; push payloads carry counts and a deep link only.

Kind
source-inspected
Locator
Deploy, Hardening notes, Push
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
relay/README.md
pandacode:usage

Agent usage panel

Account limit meters per harness and estimated API cost from Claude Code and Codex logs priced from the LiteLLM list; other harnesses show Cost not tracked.

Kind
source-inspected
Locator
Estimated API cost, Sources and tradeoffs
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
docs/agent-usage.md
pandacode:repo

aicodeking/pandacode repository

1,271 stars on 2026-11-16, last push 2026-11-16, default branch main, GitHub reports the license as NOASSERTION because the license text is custom; latest release v0.3.129 with 33 assets.

Kind
documented
Locator
Repository header and GitHub API (stars, license, pushed_at, releases)
Source date
Not stated
Retrieved
2026-11-16
pandacode:license

PandaCode License

MIT License text with one added attribution condition for forks used commercially or released publicly.

Kind
source-inspected
Locator
Whole file
Source date
Not stated
Retrieved
2026-11-16
Commit
f7961a0a903a6cd6a0978018ae3dcab1b7e35c09
Path
LICENSE

Comparisons with PandaCode

Compare PandaCode with any other product on adecomparison.com.