AGENTS.md@desktop · git:20261009.142cc6b · 2026-10-09 · sha256 3835335dafc301e4
AGENTS.md@desktop git:20261009.142cc6bA
Immutable. This exact content never changes and is served at /api/v1/blob/3835335dafc301e4.
# Desktop Agent Notes ## Frontend And Host Compatibility - The desktop frontend and host are versioned and shipped together. Do not support a current frontend talking to an older desktop host. - Evolve desktop protocol payloads, the frontend, and the host in lockstep. Do not make fields optional, hide controls, or add capability detection solely for compatibility with a pre-change host unless the user explicitly asks for it. ## Development Builds - The supported development build is the packaged `.dev` application (`package/build.mjs` with the development project config, e.g. `dist/SeekMoonDev.app`). It runs the same packaged layout as a release — `seekmoon/web` for the frontend, the per-user runtime dir for state — and differs only by its identity suffix and the development-only behavior gated on it. - The unpackaged launch — `just desktop-dev` / `package/dev`, `proton dev`, `desktop-dev.html`, `@development.Layout`, `desktop/target/dev-state` — is deprecated and no longer supported. Do not design for it, review against it, or fix issues that only reproduce under it; code that exists only to serve it may be removed. ## UI Pull Requests - For PRs with visible Desktop UI changes, include screenshots in the PR description showing the affected UI from the changed build. Include before-and-after comparisons when useful, and a short recording when the change is best demonstrated through interaction. - Upload UI screenshots and recordings as PR attachments; do not commit them to Git. Keep local captures outside the repository or in an ignored directory. - Prefer demo data when capturing screenshots or recordings. Before uploading, remove or redact personal information, real conversations, private project names and paths, credentials, and other sensitive content, including content in sidebars, terminals, and notifications. - Inspect the final images or recording before uploading to confirm that sensitive content is removed and the UI change remains clear. If visual evidence cannot be captured, explain the limitation in the PR description. ## Child Processes - For a one-shot command that collects stdout and stderr but must not read input, use `@processx.collect_output_no_stdin` from `desktop/backend/internal/processx` instead of `@process.collect_output`. - `collect_output_no_stdin` gives the child a valid stdin handle that is already at EOF and always suppresses creation of a console window. Both are required for Windows GUI builds: inheriting a missing standard-input handle makes process creation fail, while launching a console application without suppression flashes a visible console window. - Do not use it for interactive commands, long-lived protocol processes, or commands that receive a real stdin pipe, file, or PTY. ## JSON Protocol Codecs - Never use `derive(FromJson)`, `derive(ToJson)`, or a combined `derive(FromJson, ToJson)` anywhere under `desktop/`. - Every type crossing a JSON boundary must provide explicit `FromJson` and `ToJson` implementations. The decoder must deliberately specify the behavior of required, missing, `null`, malformed, and unknown fields; the encoder must deliberately specify which optional fields are omitted. - When modifying an existing Desktop protocol type that still derives either JSON trait, replace that derived codec with explicit implementations in the same change. ## Desktop Trust And Security Model - A frontend connected and authenticated to a Desktop instance is trusted with the authority of the local user. This includes the Proton renderer, an authenticated relay frontend, and code the user runs in that frontend's DevTools. - The connected frontend can already invoke Desktop terminal and shell operations. Letting the same frontend select a working directory or send an app-server execution-policy field is therefore not, by itself, a privilege escalation or a Desktop security-boundary violation. - The security boundary is before a frontend obtains an authenticated Desktop command channel. Unauthenticated network peers, unrelated web origins, repository content, model output, and other displayed data are not trusted principals and must not acquire command authority merely by being processed or rendered. - Continue validating command payloads for type safety, protocol correctness, product invariants, and protection against accidental destructive actions. Workspace and path checks may enforce product scope, but they do not sandbox a trusted frontend that can already spawn a shell. - Prefer typed request and response models at Desktop command boundaries. They document the protocol, reject malformed data, and expose schema drift during development; do not justify them as a privilege boundary against the trusted frontend. - A security finding must name the less-trusted principal and the new capability it gains. Do not report a privilege escalation solely because a trusted frontend can craft fields that the normal UI does not expose. - The local user and whatever they choose to run are one principal. A branch the user checked out and built, its hooks and build scripts, any process running under their account, a development build, and the DevTools console all already act with the user's authority. A finding whose attacker is "a branch the developer checked out", "a file in the checkout", or "a process on the developer's machine" names no less-trusted principal and is not a security finding. - "Repository content is untrusted" means the host must not grant command authority to content it processes as data — rendered Markdown, diffs, file previews, tool output, model output. It does not apply to code the user runs on purpose; checking out and building a branch is running its code. - Release and development builds share this one model. A development build differs in the surfaces it exposes, not in who is trusted: a development-only surface such as the loopback console is judged by the same two-principal test. Loopback binding limits reachability, not identity — a web page from another origin can reach loopback — so such a surface fences web origins with a launch token, a session cookie, and an `Origin` check, and does not need to defend against processes already running as the user.