Skip to content

A personal assistant that lives on your desktop, remembers what matters, and can use your Mac for you.

TypeScript
165 KB
Raw Blame

Gentle Dot — always-on desktop and web assistant backed by Gentle Shell

Objective: a local, always-available assistant (floating Dot on macOS, menu bar, global shortcut, and browser) whose hidden engine is Gentle Shell with Engram. Design: docs/design.md. Branch: feat/gentle-dot · Delivery strategy: single-pr (user choice, L7; forecast ~3,500 authored lines, recorded as a size exception) · Test runner: Vitest (pnpm test), Playwright (pnpm e2e), cargo test for apps/desktop.

Specs

  • S1 — Always available on the Mac desktop. User: "un asistente que esté todo el tiempo disponible, que puedas levantarlo, por así decirlo, y ejecutarlo, por ejemplo, desde el escritorio, que esté flotando ahí en el escritorio, que lo puedas convocar también, no sé, en la barra de tareas". "Vamos a empezar con Mac, al principio."

    • Acceptance: a Tauri 2 macOS app shows a floating, always-on-top, draggable Dot (design §6). It also has a menu bar item with the menu listed in §6 and a global shortcut (default ⌥ Space) that expands and collapses the panel. The Dot position persists across restarts.
    • Checks: cargo test for config and position persistence; a manual macOS checklist (T9) with observed results.
  • S2 — Also from a web page. User: "también que lo puedas levantar desde una web, también podría ser, sin ningún tipo de problema".

    • Acceptance: the daemon serves the same React UI at http://127.0.0.1:4317. A browser with the token can chat, answer asks, and switch conversations exactly like the desktop panel.
    • Checks: Playwright against the daemon with the fake agent.
  • S3 — Gentle Shell runs behind it, unchanged. User: "la idea es que por detrás corra GentlyAI, ¿bien? Gentle Shell, en este caso." "el agente va a hacer todo como lo hace Gentle Shell, literalmente. Todo lo mismo, usando Engram, usando todo."

    • Acceptance: the daemon spawns gentle-shell --mode rpc with the Gentle extensions loaded and its existing configuration and auth. No Gentle Shell source is modified.
    • Checks: an integration test with the real binary asserts that get_state succeeds and that get_commands lists the Gentle commands. Tag it @real-agent; it is skipped when gentle-shell is absent.
  • S4 — The user does not know it is Gentle Shell. User: "pero que el usuario no lo sepa."

    • Acceptance: identity is appended through --append-system-prompt (design §5). The sanitizer removes Gentle, Pi, el Gentleman, ODD, and Engram names from every UI-bound event. gentle:* commands are never exposed, and asked "who are you", the agent answers as Gentle Dot.
    • Checks: unit tests for the sanitizer over recorded real events; one @real-agent test that asserts the identity answer contains "Gentle Dot" and none of the hidden names.
    • Assumption: the gentle-shell launcher forwards --append-system-prompt to Pi. T4 verifies it first; if it does not, T4 stops with Needs your decision.
  • S5 — Resume after a cut. User: "Lo que tú decías de que si se corta, después no puede volver, con Engram sí es que se puede volver. porque la gente lo primero que va a hacer es ir a consultar a Engram para continuar. o si no tiene los datos de la misma sesión donde él existe."

    • Acceptance: sessions persist in ~/.gentle-dot/sessions. After the agent process dies, the supervisor respawns it with backoff and reopens the same conversation. The UI shows "I was interrupted. Continue?" with a Continue action (design §7), and Engram stays enabled.
    • Checks: an integration test kills the fake agent mid-run and asserts the respawn, the same conversation id, the history intact, and the interruption prompt.
  • S6 — Local only, verified working. User: "No lo vamos a desplegar en HostGator, solamente lo vamos a hacer a nivel local para probar y demás, bien, que todo funcione. Es más, quiero que vos también te fijes si funciona, ¿sí? Que todo ande correctamente."

    • Acceptance: everything binds 127.0.0.1 with token auth (design §8). The agent itself runs the full verification: unit, integration, Playwright, and @real-agent tests, plus the desktop app build and launch. Every result is recorded in the Log; no deployment is performed.
    • Checks: T9.
  • S7 — VPS deployment explained, not executed. User: "si tengo que hacer un video para HostGator o algo explicando esto, podría decirle, lo podés, por ejemplo, subir en la VPS de HostGator, taca taca taca, ¿bien? sin ningún tipo de drama."

    • Acceptance: docs/deploy-vps.md covers the Dockerfile, docker compose service, reverse proxy with HTTPS and authentication, token handling, and the security warnings. It closes with a short step list suitable for a video.
    • Checks: docker build of the image succeeds locally if Docker is available; otherwise the build is reported as not run.
  • S8 — Design defined before implementation. User: "quiero que me hagas un diseño, que me hagas ya todas las especificaciones, todo y ya lo dejes en el ODD". "dejamos todo, todo definido, cosa que la implementación sea muy bonita."

    • Acceptance: docs/design.md holds the architecture, protocol v1, white-label layer, UX, visual system, recovery, and security, and this document holds the specs and tasks.
    • Checks: structural readback (done in T0).
  • S9 — The assistant is our rose, glowing. User: "para el diseño, trata de utilizar nuestra rosa. ¿Bien? Estaría buenísimo que tenga animaciones, que tenga un glow, por ejemplo, como está trabajando, cosas así. pero un glow, ojo con lo que te voy a decir, la rosa misma, no el círculo donde está contenida, sino la rosa, que tenga un glow. Es más, podría ser como una línea que va recorriendo, que va teniendo un glow y que va recorriendo la rosa en sí, algo así." Earlier: "tendríamos que tener uno flotante también para Mac, cosa que si lo quieren tener en el escritorio, así dando vueltas mientras uno trabaja y le puedes hacer clic en cualquier lugar".

    • Change (L24): user: "ponle un fondo negro y que sea un circulo". The floating Dot is now the glowing rose inside a black circle (the glow and the traveling light still belong to the rose strokes).
    • Acceptance (superseded part struck by L24): the floating Dot is the rose itself in a transparent window, with no circle or container. Its traced strokes glow, and per state a line of light travels along the rose (design sheet docs/brand/rose-design.html). The menu bar uses rose-glyph.svg as a template icon, and the app icon is the neon rose on a black squircle. Reduce motion is respected.
    • Checks: UI tests for state classes; a screenshot review of every state; tray and app icons regenerated.
  • S10 — Sign in from the assistant itself. User: "tenemos que tener la posibilidad de, desde lo que es el bot, poder hacer un login para las suscripciones y demás de que ya soporta API, bien, que ya soporta genteel Cell. Tenemos que tener esa capacidad. Eso es muy importante." "Puede ser alguna opción, hacer un barra login desde ahí mismo, etcétera. Te lo digo porque la idea es que esto es para las personas que no saben determinar eso, literal."

    • Change (L25): user: "y solo aparece para loguearse con claude, que justo es algo que no anda proque no esta soportado por pi de forma legal, tenemos que ofrecer todas las otras opciones, de ultima para todo lo que no es sub que sea por api". The Claude (Anthropic) subscription sign-in is not offered (Anthropic is API key only, and the daemon refuses an Anthropic OAuth login). The accounts screen shows every other subscription first, then API keys for everything else.
    • Change (L26): user: "entonces demos una mejor variante, elegir entre sub y api primero". The accounts screen first asks for the method (subscription or API key), then lists the providers for that method, like Pi's /login.
    • Acceptance: from the panel or web UI, without a terminal, the user can sign in to every subscription provider except Anthropic and to every API key that Gentle Shell supports. That works both from a visible option and by typing /login, with plain-language guidance. Credentials stay where Gentle Shell keeps them and never pass through logs or the UI after entry.
    • Checks: to be defined after exploring how Gentle Shell performs login in RPC mode.
  • S11 — Create and edit profiles from the assistant. User: "También la de crear profiles y modificarlos, como tenemos ahora, bien, pero desde el bot mismo."

    • Acceptance: the user can list, create, edit, and switch the same profiles Gentle Shell supports today, from the UI, without a terminal.
    • Checks: to be defined after exploration.
  • S12 — The assistant brings its own instance. User: "piensa que la persona puede ni siquiera tener GentleShell instalado, ¿eh? Por eso te lo digo. O sea, el bot tiene que tener su propia instancia con sus cositas. Y lo mismo que te digo, puede ser que la persona no tenga GentleShell instalado, por lo cual el login, los perfiles, todo eso, puede ser que no tenga nada, que venga pelado. Entonces, si tiene GentleShell, lo podríamos reutilizar, pero claro, si vos guardas un modelo con un syncing y tal para lo que es el bot, que quede solamente para el bot."

    • Acceptance: Gentle Shell ships as a dependency of the daemon (gentle-pi), and the daemon runs it with its own home ~/.gentle-dot/agent (forced through --home, so ~/.gentle-shell/config.json cannot redirect it) and its own profile store (GENTLE_PI_CONFIG_HOME=~/.gentle-dot/gentle-ai). Sign-ins, models, settings, and profiles saved by the assistant stay in the assistant. It works from a clean machine with nothing configured. Reusing an existing Gentle Shell setup is only ever an explicit user choice.
    • Checks: integration test of the spawn arguments and environment; a smoke run with a fresh data dir and no gentle-shell on PATH.
    • Decided (L20): profiles are the assistant's own, with a one-time "Import my Gentle Shell profiles" action when an existing setup is found.
  • S13 — Linux desktop app for Debian, Ubuntu, and Omarchy. User: "Para lo de Linux... la que use debian, ubuntu o omarchy" (answering whether to port for X11 only or also Wayland, L21).

    • Delivery (L56): from source, with a per-distro setup script; self-contained packages are deferred.
    • Acceptance: the desktop app builds and runs on Debian and Ubuntu (GNOME, Wayland by default) and Omarchy (Arch with Hyprland, Wayland). The floating rose, panel, tray, and toggle shortcut work on each. Where Wayland forbids it (self-positioning, always-on-top, app-registered global shortcuts), a documented fallback is used: XWayland for the GNOME session, Hyprland window rules plus a bind for Omarchy, and a toggle entry point a desktop shortcut can call. Packages: .deb for Debian/Ubuntu and AppImage (also usable on Arch).
    • Checks: Rust unit tests per platform branch, plus a Linux build of the app (.deb and AppImage) in a Debian container. GUI behavior on real desktops is verified by a community tester (L27), using docs/linux-testing.md: a step-by-step checklist per distro with expected results and a copy-paste report template.
  • S14 — Queued messages are visible. User (L30): "una cosa, en la app no muestra los queue messages, y me gustaria que la app tenga la misma visual que la pagina web de gentlemanprogramming.com el mismo estilo, tambien si uno empieza a escribir deberia de cerrar las opciones abiertas, para las partes de profiles no entiendo porque el icono de contraste". When the user sends while the assistant works, the queued message shows right away as pending (from Pi's queue_update event: complete steering and follow-up queues), and it turns into a normal message when the agent takes it.

  • S15 — Same visual style as gentlemanprogramming.com. Same quote. Measured from the live site (L30): background #1a1218, surfaces #20161e/#241822, borders #342230, text #f6eff3, muted #a78e9b, accent #f095c8 (primary button with 0 10px 30px rgba(240,149,200,.3) shadow), note green #b4e7c7, red #ff718f, yellow #e0c27a; Inter for text and an uppercase monospace eyebrow (Iosevka Term / JetBrains Mono, 12 px, letter-spacing .12em, accent color); radii 8/12/16 px and pills 999 px; a translucent header with 12 px blur and a bottom border; headings in 800 weight with a light-to-pink gradient. Applies to the panel and the web view (accounts, profiles, ask cards, composer, conversations); the rose Dot keeps its S9 look.

  • S16 — Typing closes open options. Same quote. When the user starts typing in the composer, open menus and panels (conversations list, accounts, profiles) close; a sign-in that is waiting for the user's answer is kept, so its prompt is not lost.

  • S17 — Clear header icons. Same quote ("para las partes de profiles no entiendo porque el icono de contraste"). The header uses one consistent set of clear inline SVG icons with tooltips and accessible names (conversations, accounts, profiles with a sliders-style icon, new conversation, hide), replacing the text glyphs (☰ ⚿ ◐ + –).

  • S18 — Connectors (Gmail, Slack, Discord, Notion). User: "como podemos hacer para un apartado de conectores, por ejemplo gmail, slack, discord, etc" and, on the proposal, "dale me gsuta )". Accepted proposal (L32): a Connectors screen like Accounts, backed by MCP servers in the assistant's own engine home (<agentHome>/mcp.json, logins in mcp-auth.json). The daemon manages the catalog, enabling, disabling, and removal, plus each connector's sign-in (remote OAuth, a guided token, or the provider's own OAuth app), with the same secret rules as model sign-in, and restarts the agent when idle. Any outward action (send mail, post to Slack or Discord, write to Notion) needs an approval card with a preview; read-only actions do not. Each connector can be set to read-only or read-and-send. First set: Gmail, Slack, Discord, Notion, Linear, and Atlassian (L33). Sign-in per connector (L36): Notion, Linear, and Atlassian use their official remote MCP with one-click OAuth. Discord uses a community server with a guided bot token. Slack and Gmail use their official servers with a guided "bring your own app" setup: in-app steps to create the Slack app or Google Cloud project and consent screen, then paste the client credentials (stored only in the assistant's private home).

    • Open questions (being mapped, read-only): (1) whether the engine's MCP integration completes MCP OAuth in RPC mode or the daemon must run it (pi-mcp/oauth) and write mcp-auth.json; (2) which of these services offer an official remote MCP with OAuth today, versus community servers needing tokens or the user's own OAuth app; (3) how to classify tools as write actions, so only those require approval.
  • S19 — Add any other MCP by asking the assistant. User (L33): "Notion, Linear, Atlassian tmabien, esta bueno, y pon un apartado de conectar mcp extra, que directamente se lo pida al agente, creo que podria hacerlo no? es mas podriamos reconocer los mcps del user o no ?". The Connectors screen has an "Add another connector" option that starts a chat request. The assistant drafts the server configuration (name, command or URL, required credentials), but it never writes mcp.json itself: the daemon applies the draft only after the user approves a card that shows every detail. The same approval applies whenever the assistant tries to change connectors on its own initiative.

  • S20 — Recognize the user's existing MCP servers. Same quote. The assistant detects MCP configurations from common clients (for example Claude Desktop, Claude Code, Cursor, VS Code, and Gentle Shell), reads them read-only, and offers "Found these servers: which ones to import?". Import is an explicit per-server choice; secrets go only into the assistant's private home, and the source files are never modified.

  • S21 — Switching conversations while the assistant answers, and errors in the chat. User (L34): "y creo que encontre un bug, si mientras el agente esta respondiendo un mensaje, abro otra conversacion, da error y no dice que hubo error en el mismo chat".

    • Acceptance: opening another conversation or starting a new one while the assistant is answering asks first ("The assistant is still answering. Stop it and open the other conversation?" with Stop and switch, or Cancel). On confirm, the run is stopped and the switch happens, with no error shown for that intentional stop. Errors while answering appear inline in the conversation where they happened, in plain words, not only as a toast. A user-initiated stop shows a neutral "Stopped." note. History requested before the agent is ready waits for it instead of failing.
  • S22 — One continuous chat. User (L38): "por ahora entonces si es una sola sesion no pongas la capacidad de hacer nuevos chats, dejalo como esta pero ocultalo que sea un solo chat continuo te parece? mi unico miedo es que pasa cuando compacta mucchas veces y asi, si no sera mucho contexto".

    • Acceptance: the UI hides "New conversation" and the conversations list (the code and protocol stay, behind a flag, so they can return), and the header shows the assistant instead of a conversation title. The chat feels endless: the daemon rotates the underlying session transparently when it grows past a threshold (default 20 MB or 10 compactions), starting a fresh session seeded with a handoff summary, and the UI shows the previous session's recent messages above a subtle "Earlier messages" divider. The UI loads only recent messages, with "Show earlier" for more. Engram stays the durable memory, so compaction detail loss does not lose decisions.
  • S23 — Use the user's global Engram, always in project gentle-dot. User (L39): "y que pasa con engram? veo que no usa el engram total puede ser ? el global qiuero decir, estaria bueno que use el global y que siempre escriba en un proyecto llamado gentle-dots", then "gentle-dot perdon".

    • Acceptance: the engine's memory talks to the user's global Engram (the default server, port 7437, data dir ~/.engram) instead of a private one, and every memory the assistant saves goes to the project gentle-dot, whatever the workspace; searches and recall cover the user's whole memory (all projects) (corrected in L85: the user asked that it always writes to gentle-dot). The rest of the engine's HOME stays isolated (S12). The assistant no longer starts its own Engram server (which also removes the orphaned-server advisory from L31). If the global Engram is not running, the engine's memory plugin may start it as it does for Gentle Shell, with the user's data dir.
    • Mechanism (to verify in the implementation): ENGRAM_DATA_DIR=<real ~/.engram> in the child env so the plugin's instance-id check matches the global server; the project forced through the plugin's supported override (Engram has ENGRAM_PROJECT / process_override; check that gentle-engram honors it), or a workspace whose detected project is gentle-dot as the fallback.
  • S24 — Computer control (last item). User (L43): "un item mas al final de todo, que haga como codex y que permitas controlar tu computadora, que pueda manejar tus aplicaciones y tal".

    • Acceptance (original draft, refined below in L60): the assistant can see the screen and operate apps (screenshots, mouse, keyboard, opening and switching apps) after the user grants macOS Accessibility and Screen Recording permissions, with an in-app guide for granting them. Safety: a visible "in control" state on the rose, a global panic shortcut that stops control instantly, approval per control session plus mandatory confirmation for risky actions (send, pay, delete, submit), an optional allowlist of apps, and a clear note when the selected model cannot use images. Likely delivered as a built-in MCP connector behind the S18 approval guard. Scope (L59): macOS only for now (user: "lo de controlar la computadora por ahora solo para mac"); Linux is out of scope.
    • S24.1 Helper location. A native helper inside the Tauri app (Rust, apps/desktop/src-tauri), never in the daemon or the engine, so macOS attributes Accessibility and Screen Recording to Gentle Dot.app and every rule below lives outside the agent's reach. It serves MCP (JSON-RPC 2.0 over streamable HTTP, initialize, tools/list, tools/call) on 127.0.0.1:<random port>/mcp, requiring Authorization: Bearer <token> with a fresh random token per app launch. Off macOS the helper does not start.
    • S24.2 Tools. screenshot {}, click {x, y, button?: "left"|"right", count?: 1|2|3}, move {x, y}, drag {fromX, fromY, toX, toY}, scroll {x, y, dx, dy}, type {text}, key {combo} (for example cmd+shift+t, return), open_app {name} (name or bundle id), list_apps {}, wait {ms ≤ 5000}. Every action accepts an optional intent string. Coordinates are pixels of the latest screenshot; the helper maps them to logical points with the scale and origin of that screenshot. Screenshots of the display under the pointer are downscaled to a long edge of at most 1280 px (Anthropic: "pre downscale your screenshots before sending them to the API"; Pi resizes above 2000×2000), returned as an MCP image block (PNG or JPEG) plus a text block with {imageWidth, imageHeight, scale, frontApp}. Actions return a fresh screenshot. Gentle Dot's own windows are excluded from screenshots.
    • S24.3 Session grant. Any tool call (screenshots included) without an active session shows a native dialog owned by the app: "Allow Gentle Dot to see and control this Mac for 30 minutes?" (Allow / Deny). Deny returns a tool error and starts nothing. A session ends on timeout (30 min), the Stop button, the tray item "Stop computer control", or the panic shortcut ⌥⇧Esc; ending drops queued actions immediately, and the next call asks again.
    • S24.4 Guards enforced by the helper. A fixed blocklist: no action targets Gentle Dot's own windows, System Settings, SecurityAgent, loginwindow, Keychain Access, or known password managers (1Password, Bitwarden, Dashlane, LastPass, Apple Passwords); a target is the frontmost app for keyboard actions and the app owning the window under the point for pointer actions. Keyboard actions (type, key) go only to the app that was steadily in front during the latest screenshot (L130, #89); otherwise nothing is sent and the agent is told to take a screenshot first. Their recipient (the front app and, for a non-activating panel, the app owning the focused element) is fixed when the action is checked and confirmed, and rechecked before each chunk of text, each Return, and the key itself; a change stops the input and the error says how many characters were already sent ("the rest was not sent … Typing stopped after N of M characters"), never "nothing was sent" once some landed. After a confirmation, the app that was in front before the dialog must be back in front within 500 ms, or nothing is done. A rate limit of 10 actions per second. Risky actions need a native per-action confirmation showing the action and target: a click on an Accessibility element whose role is a button or menu item and whose title or description matches send, pay, buy, purchase, delete, remove, submit, confirm, transfer, place order, checkout, or their Spanish equivalents (enviar, pagar, comprar, eliminar, borrar, confirmar, transferir); return or cmd+return while the focused element is a default button or a send-like field; or any declared intent matching those words. The intent only adds checks, never removes them.
    • S24.5 Permissions guide. The app reports Accessibility (AXIsProcessTrusted) and Screen Recording (CGPreflightScreenCaptureAccess) status, can prompt for them, and deep-links to the System Settings panes. The connectors screen shows a built-in "Computer" entry with both statuses, grant buttons, a note that ad-hoc signed debug builds lose the grant on every rebuild (macOS ties it to the code signature), and a note when the current model does not accept images.
    • S24.6 Visible state. While a session is active, the rose shows an "in control" state, and the panel shows a banner with the remaining time and a Stop button.
    • S24.8 Speed (L66, user: "anda muy bien pero lento, bastante lento"; "hagamos todas si"). (a) A batch {steps: [{tool, ...args}], intent?} tool runs up to 20 steps of the action tools (click, move, drag, scroll, type, key, open_app, wait) in one call and returns one screenshot at the end; every step goes through the same checks as a single call (session, epoch/panic, rate limit, blocklist, risky confirmation, coordinate mapping against the screenshot taken before the batch); the batch stops at the first refused, declined, or failed step and reports which steps ran. The tool descriptions steer the model to batch predictable sequences. (b) The helper gets faster: dependencies that encode and resize images are optimized in debug builds, and the settle wait after an action drops from 300 ms to about 150 ms (open_app keeps its longer wait). (c) Screenshots get lighter (lower JPEG quality) without losing legibility of UI text. (d) Explain and fix the 18 s type seen in L66 if it is the helper's doing.
    • S24.9 Yolo mode (L67, user: "y mete un modo yolo"). User choices: "Solo las confirmaciones (recomendado)" and "Interruptor, se apaga solo (recomendado)". While on, risky actions (S24.4) run without the per-action confirmation; the session grant, the blocklist, the rate limit, the panic shortcut, and Stop stay as they are. A switch in Connectors → Computer and a tray item turn it on and off; turning it on asks a native confirmation owned by the app; it turns itself off when the app quits, after 1 hour, or on the panic shortcut (L74; Stop leaves it on), never persists to disk, and only the app (Tauri command or tray) can change it, never the engine or the WS. The rose and the session banner show when it is on.
    • S24.10 Prune old screenshots from the model's context (L95). User: "si lo pruebo desde pi funciona, en gentle dot esta MUY lento". Before each model call, only the most recent computer-control screenshot is sent as an image (L97, user: "no seria mejor solo enviar la ultima y de las anteriores mandar un resumen?"); each older one is replaced, in the request only, by a one-line text summary built from data already in the history (the tool's front app, action, and result), e.g. "[earlier screenshot omitted: Safari, after click at (412, 230)]", with no extra model calls. The saved session keeps every image. Applies to any model.
    • S24.11 Computer control on Windows (L131). The same helper, session and guards as on macOS, with Windows input, screenshots, app discovery and UI Automation; the details are in docs/design.md S24.11.
    • S24.7 Wiring. The desktop webview asks the app for the endpoint and registers it with the daemon (computer_register {url, token}); the daemon renders a built-in computer MCP server into the engine's mcp.json (direct exposure, no in-engine approval cards, since the helper enforces) and removes it on computer_unregister or when that client disconnects. The web UI (VPS) never offers computer control.
  • S25 — Connector enforcement outside the agent's reach (stage 2, before publishing to other users). Decided in L48. The daemon proxies MCP traffic: connector tokens live in the OS secure store (macOS Keychain, with an equivalent on Linux and the VPS), the engine sees only local proxy URLs, and enabled state, mode, and approvals are enforced by the daemon, not by files or an in-engine hook. Subagent child processes are covered too. Approvals must reach the user through a channel the agent cannot reach (L49: with the daemon access key, the agent could connect to the WebSocket and answer its own approval cards), for example a native OS dialog shown by the desktop app, or a per-window key never stored in a file the agent can read.

    • Refined in L109 (map: L109; user choices: "Mac primero, Linux documentado (recomendado)", "Motor con otro usuario + PIN (recomendado)"). Trust boundary: the agent is a same-user process with a shell, so files and environment variables are never secrets. What it cannot reach on macOS: Keychain items owned by the signed app (other readers trigger a SecurityAgent prompt, which computer control blocks), the app's native dialogs, and a secret handed from the app to the daemon over an inherited pipe and closed before any child spawns.
    • S25.1 App channel. The desktop app passes one end of a Unix socketpair to a daemon it launched as an inherited file descriptor (never a file, env var, or listening socket); the daemon keeps it from every child process (L111: the connection itself authenticates, replacing the per-launch key first drafted here). A daemon the app merely attached to (already running) gets no key and none of the privileges below.
    • S25.2 Privileged actions only through the app. Answering a connector approval, changing a connector's mode, enabling, connecting, or removing a connector, registering the computer helper, and storing or reading connector secrets are accepted only on a channel authenticated with the app key. The WebSocket access key (<data>/token) keeps chat, history, accounts, and read-only views; approval cards on the WebSocket become informational.
    • S25.3 Native approvals. Connector approvals and mode changes are confirmed in a native dialog shown by the desktop app (as computer control does), with the action's full preview; declining or no answer refuses; if the app or the channel is gone, the action is refused (fail closed).
    • S25.4 MCP proxy. The engine's mcp.json lists only http://127.0.0.1:<port>/mcp/<id> entries with a per-engine-launch key; the daemon fronts every connector (remote HTTP with tokens it injects upstream, stdio servers it runs itself), filters tools/list by mode, refuses hidden tools on tools/call, filters resources and prompts the same way, and asks S25.3 before every sending action. Subagents reach connectors only through the proxy, so they can be enabled again (GENTLE_PI_AGENTS).
    • S25.5 Secrets in the Keychain (macOS). Connector tokens (OAuth, bot tokens, client secrets, imported headers) live in Keychain items created by the signed app (stable identity from S29.2) and reach the daemon in memory over the app channel; OAuth sign-in and refresh move into the daemon (pi-mcp OAuth with a store backed by the app). Existing secrets in mcp-auth.json, connectors.json, and mcp.json are migrated once and removed from the files.
    • S25.6 Integrity. connectors.json carries an HMAC with a key kept in the Keychain; a file changed while the daemon was down is refused and reported instead of loaded.
    • S25.7 Linux desktop (documented limit). The proxy, the app key, and native approvals work as on macOS; secrets are stored with the Secret Service (encrypted at rest), and the docs state plainly that a same-user agent could read unlocked Secret Service items.
    • S25.8 VPS / web. In the container the engine runs as a different OS user than the daemon, so the daemon's files and secrets are outside its reach; secrets are encrypted at rest with a key from the daemon's environment only; approvals and mode changes in the web UI require a PIN the user sets, which the agent never sees (assumption: set on first use in the web UI, stored only as a salted hash, with a lockout after repeated failures).
    • Accepted limits: the agent can still kill the app or the daemon (everything then fails closed); third-party stdio servers hold their own token in their process; a daemon started from Terminal inherits Terminal's macOS permissions.
  • S26 — Hide the rose and full-screen chat (L68). User: "ah y pon un ocultar la rosa, que la puedas abrir bajo demanda y un full screen para el chat cuando quieras".

    • S26.1 Hide the rose. A "Hide the rose" action (tray item, checkable, and a panel header option) hides the floating Dot window; the choice is remembered across restarts in the desktop config. While hidden, the panel still opens on demand with the global shortcut and the tray "Open" item, and the tray offers "Show the rose" to bring it back. Computer control's in-control and yolo states stay visible in the tray and the panel banner while the rose is hidden (assumption: the tray icon reflects in-control).
    • S26.2 Full-screen chat. A header button (and ⌘⇧F while the panel has focus) expands the chat panel to fill the current display's visible area, and the same button or Esc returns it to its previous size and position; the full-screen layout gives the conversation a readable centered column. Assumption: it fills the screen in place rather than opening a separate macOS fullscreen Space, since the panel is an accessory window; the web UI already fills the browser window.
  • S27 — Act without the user's mouse (L69–L70). DROPPED (L71). User: "y se le puede poner su propio mouse? en vez de utilizar el mio?"; after the explanation (one system cursor; options ranked), "que sea como un glow rosa", taken as approval of the recommended combination with the agent's pointer shown as a pink glow.

    • S27.1 Accessibility first. New tools to read an app's element tree (ui_tree {app?, maxDepth}) and act on elements without moving the cursor or taking focus (press {elementId}, set_value {elementId, text}, pick_menu {path}), with the same session grant, blocklist, rate limit, risky confirmation (by the element's role and title), yolo, and panic as pointer actions. Tool descriptions steer the model to prefer them, falling back to screenshots and coordinates when an app exposes no usable tree.
    • S27.2 Keyboard to an app. type and key can target a named app without bringing it to the front when that app accepts it; otherwise they fall back to the frontmost app as today, with the same checks.
    • S27.3 Real mouse as the last resort. Pointer actions restore the user's cursor position right after the action.
    • S27.4 Pink glow. The agent's pointer is shown as a pink glow (accent #f095c8) at the point where it acts, with a pulse on click, in a click-through overlay above all windows that is never in screenshots; it fades when the agent is idle and disappears when the session ends.
    • Out of scope: a truly independent second cursor (needs a virtual machine or another user session).
  • S28 — Pink glow where the agent acts (L72). User: "lo del glow rosa si eso si para saber que esta haciendo el agente y solo mientras este interactuando". During computer control, each pointer action (click, move, drag, scroll) shows a pink glow (accent #f095c8) at the point where the agent acts, with a pulse on click and a trail along a drag; keyboard actions show it briefly at the focused element when its position is known. It is visible only while the agent is interacting: it fades out shortly after the last action and never shows when idle or after the session ends. It lives in a click-through overlay above all windows (it never blocks the user's clicks) and is never in screenshots. The rest of S27 stays dropped.

  • S29 — Installers (L79–L80). User: "anda increible!! ahora lo unico que te falta... es hacerle un instalador a la app". Choices: "Solo para mí y gente de confianza"; "macOS y Linux".

    • S29.1 Self-contained. The installed app runs on a clean machine with nothing preinstalled: it carries its own Node runtime, the compiled daemon, the bundled engine (gentle-pi) and its dependencies, and Engram; no build-machine path (today build.rs:39-41, daemon.rs:30-36) or user binary (~/go/bin/engram) is needed. Source builds keep working as today.
    • S29.2 macOS. A DMG for Apple Silicon, not notarized (on macOS 15 and later the first open goes through System Settings → Privacy & Security → "Open Anyway", since Apple removed the Control-click override, developer.apple.com/news/?id=saqachfa, Aug 2024; the docs say how). Signing keeps a stable identity across updates where possible so the Accessibility and Screen Recording grants survive updates (assumption to verify: a self-signed certificate keeps the TCC designated requirement stable, unlike ad-hoc signing).
    • S29.5 Runtime layout (L81, the contract between units). The installed app carries runtime/ as a Tauri resource: runtime/node/ (official Node 24 distribution with bin/node, bin/npm, bin/npx), runtime/bin/ (engram, a relocatable pi wrapper), runtime/daemon/cli.mjs (the daemon and protocol bundled; gentle-pi, @earendil-works/*, ws external) with runtime/daemon/identity.md, runtime/daemon/extensions/approval-guard.ts, and runtime/daemon/node_modules/ (a flat install for the target platform), and runtime/ui/ (the web UI). When runtime/daemon/cli.mjs exists, the app launches runtime/node/bin/node runtime/daemon/cli.mjs with PATH=<runtime>/bin:<runtime>/node/bin:/usr/bin:/bin:/usr/sbin:/sbin and GENTLE_DOT_RUNTIME=<runtime>; otherwise it keeps today's build-machine launch; GENTLE_DOT_NODE and GENTLE_DOT_DAEMON_SCRIPT still override. With GENTLE_DOT_RUNTIME set, the daemon serves runtime/ui, and resolves Engram as: ENGRAM_BIN, then an engram on the user's login PATH or in ~/go/bin, /opt/homebrew/bin, /usr/local/bin, ~/.local/bin, then runtime/bin/engram; ENGRAM_DATA_DIR=~/.engram and project gentle-dot stay.
    • S29.3 Linux. Self-contained packages for Debian/Ubuntu (.deb) and Arch/Omarchy (a package or AppImage), replacing build-from-source for end users; the setup scripts stay for development.
    • S29.4 Docs and checks. An install guide for both platforms, a smoke check that the installed app starts its daemon and engine on a clean user (temp HOME / container), and one command per platform that builds the installer.
    • Out of scope (L80): notarization, a Developer ID, auto-updates, and public distribution; S25 stays the gate before any public release.
  • S30 — Talk to the assistant (L86). User: "te pido una cosa mas?... posibilidad de audio a chat, cosa que puedas "hablar" con el agente, y que esto use la sub o api que ya tenes conectada". Choices: "Conversación por voz" (the user speaks, it is transcribed into the chat, and the assistant also answers aloud, with a way to mute it); "Botón de micrófono" (a button in the composer: tap to start, tap to stop).

    • Speech to text and text to speech use the subscription or API key already connected in the app when that account offers them; credentials stay in the daemon, never in the UI. Which accounts can do it, and the fallback when none can, are decided after the read-only map (L86).
    • L88 finding and choice. The only connected account is OpenAI "Sign in with ChatGPT" (OAuth), and OpenAI documents that this flow does not support the transcription API or audio input (developers.openai.com/siwc/token-sharing-open-source/preview-limitations). User choice: "Voz de macOS, y API key si la conectas (recomendado)".
    • S30.1 Desktop speech to text (macOS). The Tauri app recognizes speech natively (Speech framework, on-device when available; microphone captured natively), asking for Microphone and Speech Recognition permissions with usage strings in Info.plist. Partial text shows live while speaking; the final text is the message.
    • S30.2 OpenAI API key, when connected. If an OpenAI API key is connected in Accounts, the daemon transcribes recorded audio with OpenAI (/v1/audio/transcriptions, a gpt-4o transcribe model) and speaks replies with OpenAI text to speech; the key never leaves the daemon. On the desktop the key is preferred when present; in the web UI it is the only way to talk (the web shows the mic only when a key is connected, over HTTPS or localhost). Recordings are capped (about 60 s) to fit the WebSocket limit.
    • S30.3 Spoken replies. Without a key, replies are spoken with the system voices (speechSynthesis in the UI, matching the reply language when possible). Only replies to a spoken message are read aloud; a mute toggle in the panel header silences them (remembered), and starting to talk or pressing Stop cuts the speech.
    • S30.5 Optional local model (L89). User: "no seria mejor sino tener un modelo de ia local chiquito? Parakeet es muy muy bueno", then "pero que sea opcional para descargarlo si solo quiere". Voice settings offer a "Download local voice model" action for NVIDIA Parakeet TDT 0.6B v3 (25 European languages incl. Spanish, huggingface.co/nvidia/parakeet-tdt-0.6b-v3), run through the sherpa-onnx Rust crate (offline transducer, int8 model; works on macOS and Linux). Nothing downloads unless the user asks; the download shows size and progress, can be cancelled, verifies the files, and can be removed. When present, it is the preferred speech-to-text engine on that machine (fully offline); otherwise S30.1/S30.2 apply. On Linux it is the only desktop voice engine besides an API key.
    • S30.6 Apple Speech gets the microphone's own audio (L108). On macOS without the local model, the recognizer receives the input node's native buffers (its own sample rate and channel layout, or a high-quality AVAudioConverter mono conversion), not our 16 kHz box-filtered stream, which Parakeet keeps using. The user reported that Apple Speech transcribed "cualquier cosa" in both languages while Parakeet, fed the same capture, transcribed correctly (L101-L102).
    • S30.4 Mic button. A mic button in the composer: tap to start, tap to stop; while recording it shows the level and elapsed time, Esc or ✕ cancels, and the final transcript is left in the input, after any typed text, for the user to confirm with Enter (L103); a remembered "Send when I stop" switch in the recording strip sends it on stop instead. A sent message that included a dictation has its reply read aloud. When voice is unavailable the button explains why (missing permission, unsupported, web without a key).
  • S31 — Send files to the assistant (L91). User: "y una cosa mas :) la posibilidad de pasar archviso por chat, eso estaria genial". Choice: "Solo yo al agente".

    • S31.1 Attach. In the panel and the web UI the user attaches files to a message with an attach button, drag and drop onto the chat, or paste (images); attachments show as removable chips before sending and as chips in the sent message.
    • S31.2 Delivery. Files upload to the daemon over HTTP (same access key and Origin rules as the WebSocket; not through the 1 MiB WebSocket), are stored under the assistant's workspace (<data>/workspace/uploads/<message>/<sanitized name>, private modes), and the message tells the agent each file's name and path. Images (PNG, JPEG, GIF, WebP) also go to the model as images when it accepts them (the engine's prompt takes images); otherwise they are only saved, with a note.
    • S31.3 Limits and safety. A per-file size limit (assumption: 25 MB) and a per-message limit (assumption: 10 files, 50 MB); names are sanitized (no path traversal, no hidden or reserved names); nothing is executed or opened on upload; refused files explain why.
    • Out of scope (L91): the agent sending files back to the user.
  • S32 — Change the model from the chat (L93). User: "una cosa mas, da la posibilidad de cambiar los modelos en el chat mismo". A model picker in the chat (panel and web) shows the current model and, on open, the models of the connected accounts grouped by account (and the saved profiles), with the thinking level when the model supports it; choosing one switches the running engine live (set_model, set_thinking_level, already used by profile apply, profiles.ts:670-698) for the next message, without leaving the chat. While a reply is running the switch applies to the next message. The image note (S24.5) and voice capability follow the chosen model. Assumption: the choice persists like today's model setting.

  • S33 — Choose the shortcut in the app (L106). User: "sacalo y creo que deberiamos de cambiar por otra cosa, que uno pueda configurar su atajo no siempre space"; placement choice "Nueva pantalla de ajustes (recomendado)". Today the shortcut is read once from ~/.gentle-dot/config.json "shortcut" (config.rs:55) and changes need a restart.

    • S33.1 Settings screen. A gear icon in the panel header opens "Settings" (desktop panel only; the web UI shows no gear for now). It holds "Keyboard shortcut" and is where later options will gather.
    • S33.2 Record a shortcut. "Change" starts recording: the user presses a combination, shown with macOS symbols on macOS (⌘ ⌥ ⌃ ⇧) and names elsewhere; Esc cancels recording without hiding the panel; a combination needs ⌘, ⌥, or ⌃ unless it is a function key (L107: Shift alone would capture typing); the computer-control panic shortcut ⌥⇧Esc and combinations the system or another app already holds are refused with a clear reason. "Reset to default" restores ⌥ Space.
    • S33.3 Applies at once and persists. The app registers the new shortcut and unregisters the old one immediately (no restart), keeps the old one if the new one cannot be registered, saves it to config.json preserving every other key (private mode), and updates the tray's "Open" item.
  • S35 — Community bug fixes by root cause (L121). User: "2" (triage the public repo's issues and PRs first, fix the bugs we caused, and review PRs for the user to decide). Fix at the root, one fix per cluster, each issue closed against named tests:

    • S35.1 Immutable bundle (#8, #28). Loading gentle-pi's NaN provider through jiti writes a transpile cache into the nearest node_modules/.cache, inside the signed app bundle; the daemon must write nothing inside the runtime tree.
    • S35.2 The daemon lives as long as its app (#22). A daemon launched with an app channel stops when that channel closes, so a relaunched app always starts a daemon with a channel instead of attaching to an orphan with connectors locked; a locked state names its exit.
    • S35.3 The engine owns its home on first start (#13). A fresh default agentHome is not taken by gentle-shell as a foreign home because the daemon wrote files there first, so the engine's first-run setup (companion packages) runs on a clean install, and installs already stuck recover.
    • S35.4 The composer keeps what was not sent (#21). A message the client could not send (disconnected) keeps its text and attachments.
    • S35.5 Voice (#10, #25). A macOS audio configuration change while the engine runs is not an interruption; the mic re-checks voice status when the local model is installed or removed.

Tasks

ID Specs Route Status Commit / evidence Scope
T0 S8 inline done 5ec596c readback OK Design (docs/design.md) and this plan
T1 S3, S6 inline done 605b6c2 typecheck, lint, 2 tests green Scaffold the pnpm workspace (packages/protocol, packages/daemon, packages/ui, apps/desktop placeholder), TypeScript strict, Vitest, lint, README
T2 S3, S5 inline done c199fef RED (missing modules) → GREEN 14 daemon tests, 3 consecutive runs Agent supervisor: spawn, strict JSONL framing, id correlation, fake agent for tests, restart with backoff, reopen last session
T3 S2, S6 inline done 64e25fb RED (missing daemon) → GREEN 39 tests; node packages/daemon/src/cli.ts smoke: health OK, token 0600 Gateway: HTTP static + /health, WebSocket protocol v1, token + Origin checks, Pi RPC mapping (design §4)
T4 S4 inline done 194721d RED (missing module) → GREEN 57 tests; pnpm test:real 2/2 White-label: verify --append-system-prompt passthrough, identity.md, event sanitizer, tool-kind mapping, error neutralization
T5 S2 inline (parallel with delegated T7) done f871c01, fix 4ca84d3 (3 visual defects, RED → GREEN) Web UI: chat, streaming Markdown, activity rows, conversations, composer with Stop/steer, visual tokens
T6 S2, S4, S5 inline done f871c01, e2e dc425c0 + 5b466a8: Playwright 7/7 Ask cards (select/confirm/input/editor with timeout), toasts, interrupted-run Continue prompt, reconnect with seq gap recovery
T7 S1 delegated writer (parallel units T5/T6 and T7, disjoint surfaces) done 81d70c3 cargo test 29/29 (parent re-ran), tauri build --debug produced Gentle Dot.app; GUI not launched yet (T9) Tauri 2 app: tray menu, floating Dot window, expand panel, ⌥ Space shortcut, daemon spawn/attach, position persistence, launch at login
T8 S7 inline done ee74148 docker build OK, container smoke: health idle, companion setup 3/3 Dockerfile, compose.yaml, docs/deploy-vps.md, security checklist, video step list
T10 S9 delegated writer (context backstop; parallel with T12) done a5e1c5b RED → GREEN: UI 9 failing → 13/13; Rust 17 compile errors → 34/34; tauri build OK; parent re-ran both Apply the rose design: Dot surface (transparent rose window ~64×84, traced strokes, glow, traveling light per state), tray template glyph with frames, app icon from the neon rose
T11 S10 explore (delegated) then inline done 43a512d RED → GREEN: 12 daemon auth tests (fake runtime + real offline ModelRuntime), 9 UI tests, e2e 8/8 incl. /login + API key connect/disconnect Login from the UI for subscriptions and API keys (/login and a visible option)
T12 S11 explore (delegated) then delegated writer (context backstop; parallel with T10) done see L23; RED → GREEN per layer; parent re-ran 179 tests + typecheck Profiles: list, create, edit, switch from the UI
T13 S12 inline done 43a512d config + spawn tests; real smoke without gentle-shell on PATH: own home created, 42 providers, none configured Bundle gentle-pi, own --home and profile store, first-run timeout, no PATH dependency
T14 S13 delegated writer (context backstop) done (desktop verification pending the community tester) 3c002e3; parent re-ran cargo 53/53, 374 tests, typecheck; bundled-pi fix see L58 Linux port of the desktop app: per-platform window effects, tray icon, Wayland fallbacks, AppImage/deb bundles
T15 S6, S7, S11, S12, S9 delegated writers (parallel: daemon/isolation vs. Dot circle) done 8626a7d, 05000c8, ac86f64; limited recheck PASS (L31) One correction batch for verifier blockers B1–B4 and the S9 circle change, then a recheck limited to them
T16 S10 inline done 7337b52 RED 4 failing → GREEN 14/14 Accounts in two steps: method first, then provider
T17 S14–S17, S21 delegated writer (context backstop) done b0ee7ac; parent re-ran 227 tests, typecheck, lint Queue display, gentlemanprogramming.com style, typing closes options, header icons
T18 S18–S20 explore (delegated) → T18a core + T18b guides/drafts/import, each verified independently done e49bd9d, 8651bba (T22), 994abce, 63c8e9c; see L46–L55 Connectors: map the engine's MCP integration and the providers first
T19 S22 (+ design §4/§6/§7 refresh) delegated writer (context backstop) done 225596d; parent re-ran 260 tests, typecheck, lint + real-engine handoff probe Single continuous chat: hide conversations, transparent session rotation, recent-only history
T20 S23 delegated writer, after T17 (the runtime refused it in parallel: test surface overlap with T17's packages/daemon/test/**) done 823363e, follow-up 400ec73; parent re-ran 235 tests, typecheck, e2e 12/12 Global Engram, project gentle-dot
T21 S24 (macOS only) explore (delegated, read-only) → spec → implement in progress map done (L60) Computer control with permissions guide, panic stop, and approvals
T21a S24.1–S24.5 (app side), S24.6 events delegated writer (parallel with T21b, disjoint surface apps/desktop/src-tauri/**) done (verified, real Mac pending) 176f547, fixes 3e29886, 3ebfe5c, N1/N2 below; RED 82 failing → GREEN; parent re-ran cargo test 145/145 Rust helper: MCP server, session grant, guards, permissions, panic
T21b S24.5 UI, S24.6, S24.7 delegated writer (parallel with T21a, surface packages/**, docs/**) done 5b59c0e; RED 20 failing → GREEN; parent re-ran 400 passed / 5 skipped Protocol, daemon registration, connectors entry, rose state, banner
T24 S24.8, S24.9 delegated writer (context backstop), then independent verifier (high: touches the S24.4 checks) done (verified, real Mac pending) 952830e, panic-off de50aa9; RED 15 failing → GREEN; parent re-ran cargo test 193/193, pnpm test 407 passed Computer control speed: batch, faster helper, lighter screenshots
T25 S26 delegated writer (context backstop), parallel with the read-only T24 verifier; surface excludes computer/** done (real Mac pending) 8b274ed; RED 30 compile errors + 8 UI failing → GREEN; parent re-ran cargo test 203, pnpm test 417 Hide the rose; full-screen chat
T26 S27 — dropped (L71) —
T27 S28 delegated writer (context backstop), then independent verifier (high: owner/element resolution under S24.4) done (verified, real Mac pending) f01dfdf, A1 fix below; RED 51 compile errors + UI import → GREEN; 2 mutation checks; parent re-ran cargo test 217, pnpm test 426 Pink glow while the agent interacts
T28 S29 explore (delegated, read-only) → plan → implement done (second-Mac and real-desktop checks pending) T28a–g Self-contained installers: macOS DMG, Linux .deb and Arch
T28a S29.1, S29.5 (app launch) delegated writer, parallel with T28b (surface apps/desktop/src-tauri/src/daemon.rs) done 84a38a4; RED 3 failing → GREEN; parent re-ran cargo test 222 Launch the bundled runtime when present
T28b S29.1, S29.5 (daemon) delegated writer, parallel with T28a (surface packages/daemon/**) done ea52273; RED 7 failing → GREEN; parent re-ran pnpm test 440 Bundled daemon build, runtime-aware paths, Engram resolution, relocatable pi wrapper
T28c S29.1, S29.4 delegated writer (context backstop) done see L83; parent re-ran pnpm test 454 scripts/package/stage-runtime.mjs per target; clean-HOME smoke, read-only tree smoke, size
T28d S29.1, S29.2, S29.4 delegated writer, parallel with T28f done 546aa92 Tauri macOS resources + DMG + pnpm package:mac; launch from /Applications copy
T28e S29.2 inline; user consented (L84) done 5ec1d25; two signed builds both certificate leaf = H"4611e0a8…", verify ok Self-signed identity; compare the designated requirement across two builds
T28f S29.3, S29.4 delegated writer, parallel with T28d done see L87 Linux .deb (arm64, amd64) and Arch PKGBUILD in containers; clean install smoke
T28g S29.4 inline done see L87 docs/install.md; Hyprland snippets point at /usr/bin/gentle-dot
T29 S30 explore (delegated, read-only) → spec → implement in progress map done (L88) Voice conversation: mic button, transcription, spoken replies
T29a S30.1 delegated writer, parallel with T29b (surface apps/desktop/src-tauri/**) done (real Mac pending) 93ee637; RED 39 compile errors → GREEN; parent re-ran cargo test 259 Native macOS speech recognition commands and events
T29b S30.2–S30.4 delegated writer, parallel with T29a (surface packages/**) done (real Mac pending) fd7c25f; RED → GREEN; parent re-ran pnpm test 520 Daemon OpenAI-key transcription/TTS, mic button, spoken replies
T29c S30.5 delegated writer (Rust: engine, download, commands), parallel with T30; UI hook after T30 done (real Mac pending) Rust efc9e66, UI 4848d24 Optional Parakeet local model via sherpa-onnx: download, verify, remove, prefer
T30 S31 delegated writer, parallel with T29c (surface packages/**); independent verifier (high: authenticated upload writing to disk) done (verified, real Mac pending) 8a79b48, drop handler 2e9b087; RED → GREEN; 2 mutation checks; parent re-ran pnpm test 562 Attach files to messages: upload, workspace storage, images to the model
T31 S32 delegated writer, parallel with T29c done (real Mac pending) d0b63ce; RED 16 failing → GREEN; parent re-ran pnpm test 579
T32 S24.10 inline, after T31 done (real NaN check pending) d9f0fa8; RED 2 files failing → GREEN; pnpm test 584 Model picker in the chat
T33 S33 delegated writer (context backstop; Rust + UI) done (real Mac pending) 2d1f5e4; RED 51 compile errors + UI import → GREEN; cargo test 314, pnpm test 609 Settings screen with a recordable, live-applied shortcut
T34 S30.6 delegated writer (Rust, apps/desktop/src-tauri/src/voice/**), parallel with the S25 map in progress — Apple Speech fed with native buffers
T23a S25.4 delegated writer (daemon proxy), parallel with T23b done (verified, no blockers) 6016b43; RED 14+8 failing → GREEN; mutation check; pnpm test 636 MCP proxy in the daemon: list/call filtering, hidden-tool refusal, resources, upstream HTTP and stdio
T23b S25.5 (app side), S25.3 (dialogs) delegated writer (Rust), parallel with T23a done (B1 fixed in 8ac11aa, rechecked PASS) 1c345b5; RED 22 failing → GREEN; 2 mutation checks; cargo test 348 Keychain store and native approval dialog in the app
T23c S25.1–S25.3 delegated writer (Rust + TS), parallel with the T23a/b verifier done (verified; B2 fixed 4d1eafb; real Mac pending) 8ac11aa; RED across protocol, daemon, Rust, UI → GREEN; 4 mutation checks; pnpm test 658, cargo test 366 Private app channel (socketpair on fd 3); privileged actions and native approvals only through it
T23d S25.5 (daemon side) delegated writer (Rust + TS; context backstop) done (verified, no blockers; A1 fixed, L117) 8b29bab, e2e c6d7c6f; RED → GREEN; pnpm test 669 (8 clean full runs), cargo test 369, e2e 17 OAuth in the daemon with the app-backed store; migration from files
T23e S25.6 delegated writer (TS; context backstop) done (verified; A1/A3 fixed, L120) 9925c39; RED 17 failing → GREEN; mutation check; pnpm test 688 ×2, e2e 17 connectors.json HMAC
T23f S25.4 (subagents) delegated writer, parallel with T23g done (verified, B1 fixed and rechecked) e9869d4, b6e235d Enable subagents again behind the proxy; real-engine probe
T23g S25.7 delegated writer (Rust Linux + containers), parallel with T23f done (verified, B1/B2 fixed and rechecked) 837a42d, 920455b Linux: Secret Service store and documented limit
T23h S25.8 delegated writer (context backstop), after T23f done (verified; B1 fixed; full re-verification no blockers) 35b942b, 162473c, 4d1035e VPS: engine as another OS user, encrypted secrets, PIN approvals
T35 S35.1 inline in progress — No jiti cache inside the bundle
T36 S35.2, S35.3 delegated writer (daemon + Rust daemon spawn), parallel with T37/T38 in progress — Daemon lifetime and engine home ownership
T37 S35.4, S35.5 (#25) delegated writer (UI), parallel in progress — Composer send result; voice re-check
T38 S35.5 (#10) delegated writer (Rust voice), parallel in progress — Configuration change is not an interruption
T22 S18 (hardening, L47 B1/B2) delegated writer done (accepted with documented gaps, L52) 8651bba; parent re-ran 328 tests + typecheck Option A: authoritative in-memory connector state with revert, credential reads blocked, path-independent bash checks, full approval preview, redirect manual, private Engram stop, onboarding overlap fix
T23 S25 explore (delegated, read-only) → spec → implement (stage 2) in progress map L109; units T23a–T23h Daemon-side MCP proxy with Keychain tokens
T9 S1–S6, S9–S12 delegated verifier (high risk: process supervision, auth) done L24 + L31 Full verification: unit, integration, Playwright, @real-agent, tauri build + launch, manual macOS checklist with observed results

Environment notes: Node v24.14.1 and pnpm are present. cargo is on PATH but rustc is not, so T7 must first install or repair the Rust toolchain (rustup) and asks before installing. Xcode is present.

Log

  • L1 (2026-10-08, user, original request): "quiero hacer algo, nuestro propio https://www.copilotkit.ai/opendots pero con gentle ai, se podra hacer algo asi que por detras use gentle shell ? con nuestro propio diseño y tambien evalua si pi durable vale la pena o engram ya esta bien"
  • L2 (2026-10-08, user): "Lo que yo quiero es un asistente, ¿bien? Es un asistente que esté todo el tiempo disponible, que puedas levantarlo, por así decirlo, y ejecutarlo, por ejemplo, desde el escritorio, que esté flotando ahí en el escritorio, que lo puedas convocar también, no sé, en la barra de tareas, algo así, ¿bien? Vamos a empezar con Mac, al principio. Entonces, esa sería la idea principal. Si no, también que lo puedas levantar desde una web, también podría ser, sin ningún tipo de problema, ¿bien? Que lo puedas levantar, desplegar, por ejemplo, en HostGator, se podría hacer sin ningún tipo de drama. y lo puedas también ejecutar desde ahí. Entonces puedes acceder a esa gente desde cualquier lugar. Bien, una cosa sí podríamos hacer. No sé qué te parece. Y la idea es que por detrás corra GentlyAI, ¿bien? Gentle Shell, en este caso. pero que el usuario no lo sepa. Entonces, el agente va a hacer todo como lo hace Gentle Shell, literalmente. Todo lo mismo, usando Engram, usando todo. Lo que tú decías de que si se corta, después no puede volver, con Engram sí es que se puede volver. porque la gente lo primero que va a hacer es ir a consultar a Engram para continuar. o si no tiene los datos de la misma sesión donde él existe."
  • L3 (2026-10-08, user): "No, no, lo que vamos a hacer es lo siguiente, mirá. No lo vamos a desplegar en HostGator, solamente lo vamos a hacer a nivel local para probar y demás, bien, que todo funcione. Es más, quiero que vos también te fijes si funciona, ¿sí? Que todo ande correctamente. Quiero que me crees un diseño y después yo, si tengo que hacer un video para HostGator o algo explicando esto, podría decirle, lo podés, por ejemplo, subir en la VPS de HostGator, taca taca taca, ¿bien? sin ningún tipo de drama. Entonces, yo lo que quiero ahora es que me hagas un diseño, que me hagas ya todas las especificaciones, todo y ya lo dejes en el ODD, porque vamos a tener que hacer con el ODD, es largo esto, ¿bien? y nada, dejamos todo, todo definido, cosa que la implementación sea muy bonita. ¿Te parece?"
  • L4 (2026-10-08, user decisions): Repository: "Repo nuevo ~/work/gentle-assistant (Recomendado)". Name: "Gentle Dot". Desktop: "Tauri 2 (Recomendado)".
  • L5 (2026-10-08, evidence): gentle-shell --mode rpc --no-session returned get_state success (model reported) and get_commands with 82 commands including gentle:*, and emitted extension_ui_request setStatus/setWidget/notify at startup, so branding reaches RPC clients and S4 needs the sanitizer. Pi RPC docs: custom() returns undefined in RPC mode, while select/confirm/input/editor/notify/setStatus/setWidget/setTitle are forwarded. extensions/ask-user-choice.ts already falls back to ctx.ui.select + ctx.ui.input for RPC hosts. Pi CLI has --append-system-prompt and --session-dir.
  • L6 (2026-10-08, decision): Pi Durable deferred. Recovery relies on persisted Pi sessions plus Engram, per L2; it is reconsidered only for unattended or scheduled work.
  • L7 (2026-10-08, user): "Un solo pr, instala lo que necesites y dale nomás" — plan approved, delivery single-pr with size exception, installs authorized (Rust stable via rustup).
  • L8 (2026-10-08, evidence): Rust stable 1.99.0 installed via rustup (Homebrew); the nix cargo precedes it on PATH, so Tauri commands prepend /opt/homebrew/opt/rustup/bin. Toolchain versions: TypeScript 7, Vitest 5, Biome 2, pnpm 11.5. T2 risk tier: medium (process supervision covered by integration tests with a fake agent; the fake delays crash exit because macOS pipes are asynchronous).
  • L9 (2026-10-08, T3 evidence): 18 gateway tests cover the token (4401), protocol (4400), Origin (403), seq numbering, streaming, broadcast, requestId dedupe, tool activity, asks and late joiners, conversations and history, id traversal refusal, malformed frames, interruption after restart, static UI with traversal block, and health. Exception: parseClientMessage was written before its unit test (the test passed on first run). Design §4 updated: send while busy steers, plus user_message, ask_resolved, conversations, interrupted payloads. Risk tier: high (auth on a local socket); independent verification in T9.
  • L10 (2026-10-08, T4 evidence): the launcher forwards unknown args to Pi (runtime/gentle-shell-launcher.mjs:153,971), so --append-system-prompt works; the S4 assumption holds. pnpm test:real (opt-in through GENTLE_DOT_REAL_AGENT=1 so ordinary runs cost no model calls) passed 2/2. Answer: "I’m Gentle Dot, a personal assistant that lives on your desktop and in your browser. I can work with your files, run tasks, remember things across conversations, and bring in focused helpers when needed." No hidden names appeared in any UI-bound message. Info toasts are hidden as harness chatter.
  • L11 (2026-10-08, T5/T6 evidence): Route: T5/T6 inline and T7 delegated in parallel (Writer trigger (a): disjoint surfaces packages/ui and apps/desktop/src-tauri, contract fixed first in design §9). RED observed for components, client, and Dot (missing modules); exception: the store reducer was written before its tests ran (tests passed on the first run). Screenshots exposed 3 defects that the tests missed: lost user line breaks, start/end activity ids split, and a title stuck at "New conversation"; each got a RED test, then a fix (4ca84d3). Totals: vitest 88 passed + 2 opt-in skipped; Playwright 7/7 (chat, ask card, conversations, interruption + Continue, reload, rejected key, missing key).
  • L12 (2026-10-08, T8 evidence): HostGator cPanel/WHM Apache owns ports 80/443, so the default compose binds 127.0.0.1:${DOT_PORT} behind an Apache include (docker/apache-gentle-dot.conf, Basic Auth + WebSocket proxy), and Caddy is an opt-in profile for a VPS without a web server. The safety policy blocked creating .env.example; the variables are documented in docs/deploy-vps.md step 3 instead. Image fixes found by the smoke run: gentle-pi installs its gentle-ai binary as root with mode 0700 (chown to dot), and companion setup needs curl plus pi on PATH (pinned to gentle-pi's bundled version). Final smoke: health idle, companion packages installed, verification 3 passed. A chat inside the container was not run because it needs a model key.
  • L13 (2026-10-08, user): "Me haces un diseño de cómo se podría ver el iconito que vamos a cliquear para levantar a la gente y demás. Acordate que te dije que tendríamos que tener uno flotante también para Mac, cosa que si lo quieren tener en el escritorio, así dando vueltas mientras uno trabaja y le puedes hacer clic en cualquier lugar, tenemos que tener esa representación también." A violet orb design was produced (9e8cecc), then superseded by L15.
  • L14 (2026-10-08, user): "Genérame un cloud de design, por favor, y pásame el link." Options offered (a Claude Design prompt, a Higgsfield file link, a Higgsfield site); the user cancelled and redirected in L15. Nothing was published.
  • L15 (2026-10-08, user): S9 quote. Evidence: the rose source is gentle-pi/assets/gentle-logo-only.png (1254×1254 neon PNG), copied to docs/brand/rose-source.png. docs/brand/tools/trace_rose.py (isolated venv with numpy, Pillow, and scikit-image; threshold 120, closing disk(3), skeletonize) traced 178 strokes and 45 sparkles into rose-lines.svg. Traced strokes are illegible at 18 pt, so a hand-drawn rose-glyph.svg is used for the menu bar. Sheet: rose-design.html/.png; video: rose-animation.webm (d4672e3). T7 integrated (81d70c3) after removing unused Android/iOS/Windows icons.
  • L16 (2026-10-08, user): S10 and S11 quotes (login and profiles from the bot).
  • L17 (2026-10-08, explore evidence): /login is a built-in TUI command that RPC does not execute, and /gentle:profiles//gentle:models use ctx.ui.custom (undefined in RPC). The daemon can sign in through ModelRuntime.create({authPath, modelsPath}).login(providerId, type, {prompt, notify, signal}, {getDeviceId}), exported from @earendil-works/pi-coding-agent (verified by the parent at dist/core/model-runtime.d.ts:56-108), with AuthStorage file locking. New credentials apply on the next request, but the model list needs a restart. Profiles live in ${GENTLE_PI_CONFIG_HOME:-~/.pi/gentle-ai}/profiles.json; switching them globally exists only inside extensions/gentle-ai.ts.
  • L18 (2026-10-08, user): S12 quote. Decision: the assistant owns its instance (bundled [email protected] with pinned pi-coding-agent, pi-ai, and pi-tui 1.0.4). pnpm 11 allowBuilds: esbuild and gentle-pi allowed; @google/genai and protobufjs denied.
  • L19 (2026-10-08, T13/T11 evidence): A smoke run with a fresh data dir and gentle-shell removed from PATH started the bundled engine in <data>/agent: companion setup passed 3/3, the agent was idle, and auth_list returned 42 providers (9 with subscription sign-in: Claude Pro/Max, ChatGPT Plus/Pro and subscription, GitHub Copilot, xAI, Kimi, Meta, OpenRouter, Radius) with none configured. Defects found by the e2e and fixed test-first: (1) "Back to accounts" kept the finished flow on screen; (2) a fresh ModelRuntime with refreshOnCreate: false did not see stored credentials (verified with a direct probe), so it now refreshes on create, still offline. The real API key flow is one secret prompt with no network check, and auth.json is written with mode 0600. The Tauri capability adds opener:default for sign-in pages; cargo build + 29 tests pass. Dockerfile updated for the bundled engine (chown /app to dot; global pi pinned to the bundled version). Risk tier: high (credentials, item 2); independent verification in T9. Not verified: a real OAuth subscription sign-in (needs the user's account in a browser).
  • L20 (2026-10-08, user decision): "Propios del bot, con importar (Recomendado)": the assistant keeps its own profiles in ~/.gentle-dot. When the person has Gentle Shell, an "Import my Gentle Shell profiles" button copies them once, and changes made in the assistant never touch Gentle Shell. /gentle:profiles takes no arguments (panel only, extensions/gentle-ai.ts:3606), so the daemon must manage profiles itself; a read-only map of what applying a profile writes is running. The Docker image with the bundled engine was verified (8d5a135): health idle, setup 3/3.
  • L21 (2026-10-08, user): "Ve evaluando que tan fácil es portarlo para Linux y actualiza el todo". Read-only evaluation: the daemon, UI, protocol, auth, and Docker image are already platform-neutral (the container is Linux). The desktop shell has macOS-only points (src/shell.rs): HudWindow vibrancy (macOS/Windows only), the accessory activation policy (already cfg(target_os = "macos")), icon_as_template, accept_first_mouse, the macos-private-api feature/macOSPrivateApi, Info.plist LSUIElement, and .cargo/config.toml pinning /usr/bin/cc. Linux constraints: on Wayland, apps cannot position their own windows or reliably stay on top, and global shortcuts need the GlobalShortcuts portal, so edge snapping, the floating rose, and Alt+Space are X11-only; on GNOME the tray needs the AppIndicator extension; transparency needs a compositor. Estimate: X11 parity is small (about 1–2 days); Wayland needs fallbacks (a normal window instead of the free-floating rose, a tray-or-panel launcher, a shortcut set by the user in desktop settings). Recorded as T14, pending a go-ahead.
  • L22 (2026-10-08, T10 evidence): The rose is an inline SVG built from the traced strokes: 178 static outline strokes, with light traveling along the 60 longest (at 64 px the shorter ones read as flicker). Finding: macOS system WebKit does not draw CSS drop-shadow on <g> inside inline SVG, so the rose is three stacked <svg> layers (base, light, sparkles), each with its own glow; verified with qlmanage system-WebKit renders. The Dot window is 64×84 with initial_dot_position using the real size. There are four tray template glyphs at 18/36 px (working is a static dashed variant, not two frames), plus a neon-rose squircle app icon and the favicon. Reduce motion hides the light layer and keeps a steady glow. Risks for T9: live animation in WKWebView is not yet observed, and the working tray glyph is subtle at 18 pt. The unused icons/tray.png was removed.
  • L23 (2026-10-08, T12 evidence): The writer stalled once on a bash command and was resumed in its own session with timeouts on every long command. The daemon reimplements the gentle-pi.agent_model_profiles v1 store: atomic 0600 writes through one queue. Applying a profile writes the four files in order, restores all four on failure, never touches agent .md files, then sends set_model and set_thinking_level (deferred until agent_settled while busy). Models forward only provider, id, name, and reasoning (a gateway test asserts that headers never reach a client). The import is read-only, skips active, renames collisions to -imported[-N], and reports missing sign-ins. Extras beyond the request: editing the active profile re-applies it, identical imports are skipped, and Delete is disabled for the active profile. SettingsManager was not used because it cannot delete defaultThinkingLevel; a settings.json.lock mkdir lock (stale after 10 s) mimics proper-lockfile. Verification: 179 tests passed (2 opt-in skipped), typecheck OK, biome OK, e2e 9/9 (the e2e test was written after the implementation, so no RED). Open for T9: set_model against the real engine, lock compatibility with proper-lockfile, and one flaky supervisor.test.ts:54 (busy after settle), seen once in four runs.
  • L24 (2026-10-08, T9 evidence and user feedback): Independent verifier (scratch copy, temp HOME): runners all green (179 tests, typecheck, lint, e2e 9/9, cargo 34/34, test:real 2/2, supervisor ×10 = 10/10). Verdicts: pass S2, S3, S4, S5, S8, S10; not verifiable S1, S9; FAIL S6, S7, S11, S12. BLOCKERS: B1 the token printed by cli.ts:21 lands in ~/.gentle-dot/daemon.log (0644, dir 0755), readable by other local users. B2 the first run writes outside the instance (~/.gentle-shell/config.json provisioned mark, ~/.pi/agent/pi-pretty/fff/*.mdb via a fixed $HOME, ~/.gentle-ai/state.json, and ~/.gentle-ai/backups). B3 with the default workspace ~, the engine reads ~/.pi/gentle-ai/models.json as project config and writes it into <data>/agent/agents/*.md frontmatter, and an applied profile that omits a role does not clear that frontmatter. B4 deploy-vps step 3 (docker compose run … gentle-shell) cannot work. Advisories: generic error with no account; a busy race in supervisor.ts:118-123; stale ready.model after set_model; auth_logout accepts any providerId; auth_reply has no owner check; symlinked session files are accepted; rollback leaves 0600; internal role ids in the payload; unpinned engram@latest; the token in docker logs. App launched on the user's Mac: CGWindowList shows the Dot window floating (layer 5) but at 59×76 and about 106 px from the edge, against a saved 64×84 at a 12 pt margin; screenshots are not possible (no screen-recording permission). User: menu bar icon and menu OK; clicking the rose or ⌥ Space opens the panel beside it; on the rose: "ponle un fondo negro y que sea un circulo".
  • L25 (2026-10-08, user): S10 change quote. Reproduced: the accounts list is alphabetical with 42 rows, so the only subscription button on the first screen is "Sign in with Anthropic (Claude Pro/Max)", and ChatGPT, Copilot, Grok, Kimi, Meta, OpenRouter, and Radius sit far below among the API keys. Fix: a daemon policy that drops and refuses Anthropic OAuth (sent to the T15 daemon writer), and an accounts UI with a "Use your subscription" section first and "Use an API key" below (parent, inline).
  • L26 (2026-10-08, user): "entonces demos una mejor variante, elegir entre sub y api primero. Para lo de Linux... la que use debian, ubuntu o omarchy". The user had shared screenshots of Pi's /login showing the method choice and the 9 subscription providers (Anthropic, GitHub Copilot, Kimi For Coding, Meta, OpenAI, OpenAI Codex (legacy), OpenRouter, Radius, xAI), which matches the assistant's list; rows now use provider names with the plan underneath (7998aea). Linux targets all default to Wayland (Debian/Ubuntu GNOME, Omarchy Hyprland), so S13 includes Wayland fallbacks.
  • L27 (2026-10-08, user): "puedo poner a alguien de la comunidad a probarlo en linux no pasa nada". T14 will deliver a tester checklist (docs/linux-testing.md) and the build artifacts. Sharing them with the tester (push, release, or file transfer) stays the user's decision.
  • L28 (2026-10-08, T15 design part evidence): The Dot window is 72×72 pt, with a 66 pt near-black disc and a faint pink rim (subtle state cues on the rim: pink while working, amber with the "!" when it needs the user, grey when restarting or failing). The rose is cropped to 90 78 1040 1040 (its minimal enclosing circle). RED → GREEN: UI 3 failing → 72/72; Rust 6× E0425 → 39/39; tauri build OK; the parent re-ran both. Window bug: all placement now uses logical points, converting each rectangle with its own scale factor (tao 0.37.1 on macOS uses per-window or per-monitor backingScaleFactor, and set_outer_position/set_inner_size are async). The size is re-applied after positioning, on each snap, and on ScaleFactorChanged, and positions persist as dot_points (the legacy dot key is ignored once). The root cause of 59×76 is not proven without the GUI; the CGWindowList recheck on the user's Mac should read 72×72 with x + 72 + 12 equal to the work-area width.
  • L29 (2026-10-08, T15 daemon part evidence): Commit 05000c8. B1: data dir 0700, token 0600, the token printed only on a TTY, Tauri log 0600. B2: the child gets its own HOME=<data>/home plus XDG dirs, GIT_CONFIG_GLOBAL pointing at the real gitconfig, and its own Engram on port 7438 (gentle-engram rejects the user's server as foreign under another HOME); probe: seeded HOME hash identical before and after a first start. B3: workspace defaults to <data>/workspace, and frontmatter sync is guarded by realpath with rollback of bytes and modes. B4: the deploy doc signs in from Accounts. A1–A6 fixed, the L25 Anthropic OAuth refusal added (auth_unsupported). 196 tests, e2e 9/9, cargo 39/39, supervisor ×10 = 11/11 each. Incident: the writer's pkill -n -f "engram serve" may have restarted the user's Engram (same instance id, healthy, verified by the parent). Parent follow-ups ac86f64: "Back to accounts" returns to step 1 (RED → GREEN), access-key texts no longer mention a printed link, the unused container memory server removed, design §3/§9/§10 updated, and stray packages/daemon/src/*.d.ts untracked and ignored. On the user's Mac the Dot window now measures 72×72 at x = 1836 (1836 + 72 + 12 = 1920), confirming the window fix; the restarted app uses its own home, workspace, and Engram on 7438. The limited recheck is running.
  • L30 (2026-10-08, user): "una cosa, en la app no muestra los queue messages, y me gustaria que la app tenga la misma visual que la pagina web de gentlemanprogramming.com el mismo estilo, tambien si uno empieza a escribir deberia de cerrar las opciones abiertas, para las partes de profiles no entiendo porque el icono de contraste". Style evidence was measured with Playwright from the live https://gentlemanprogramming.com (computed styles and :root variables; the values are in S15). The site is dark-only.
  • L31 (2026-10-08, limited recheck): An independent verifier (scratch copy, temp HOMEs, ports away from the user's 4317/7437/7438) passed B1, B2, B3, B4, L25, and L26 at ac86f64, and each failure reproduced at baseline deeb33a. B2 probe: the seeded HOME hash is identical before and after a first start (b5f7056…). L25: 42 providers, 8 OAuth, none Anthropic, and Anthropic OAuth is refused with no runtime call. Regression: 197 tests, typecheck, lint, e2e 9/9, cargo 39/39. No blockers remain. Advisories for later: (1) the assistant's Engram server outlives the daemon (orphaned after stop); (2) deploy-vps telemetry opt-out line likely does not work in the container (pre-existing); (3) old daemon.log files from earlier builds still hold the token (now 0600; the token was not rotated); (4) e2e does not re-check the method buttons after "Back to accounts" (UI test covers it); (5) a small TOCTOU window between realpath and write in the frontmatter sync. With L24 (S1/S9 confirmed on the user's Mac: tray, menu, panel, 72×72 window), T9 is complete for S1–S12 except the items not verifiable here (live animation in WKWebView, real OAuth subscription sign-in).
  • L32 (2026-10-08, user): "como podemos hacer para un apartado de conectores, por ejemplo gmail, slack, discord, etc". The parent verified that the engine reads <agentHome>/mcp.json (mcpServers) and keeps mcp-auth.json and mcp-cache.json there, and that @earendil-works/pi-mcp provides stdio and Streamable HTTP transports plus an MCP OAuth client (pi-mcp/oauth; it does not open a browser or choose storage). The proposal was accepted: "dale me gsuta )".
  • L33 (2026-10-08, user): "Notion, Linear, Atlassian tmabien, esta bueno, y pon un apartado de conectar mcp extra, que directamente se lo pida al agente, creo que podria hacerlo no? es mas podriamos reconocer los mcps del user o no ?". Decision recorded: the assistant drafts and the daemon applies after user approval (prompt-injection guard). Import of the user's MCP servers is read-only, per server, and explicit. The explorer was asked to extend its map.
  • L34 (2026-10-08, user): "y creo que encontre un bug, si mientras el agente esta respondiendo un mensaje, abro otra conversacion, da error y no dice que hubo error en el mismo chat". Reproduced from the user's ~/.gentle-dot/daemon.log (error lines only): 12:17:05 assistant error: This operation was aborted. Pi's switch_session aborts the active run (agent-session-runtime.js teardownCurrent → session.abort()); the bridge turned the aborted message_end into a generic error toast outside the chat. Also, command get_history failed: Agent is not ready repeats at every startup (history requested before the agent is ready). Both were added to T17 (same files: bridge.ts and UI).
  • L35 (2026-10-08, T18 map + web verification): MCP is a built-in pi-coding-agent 1.0.4 extension that reads <agentDir>/mcp.json only at session_start and stores OAuth in mcp-auth.json (keys mcp__<ns>|<url>, locked writes); new tokens are picked up on the next turn. Headless OAuth runs as a separate process (pi mcp login <name> with the assistant's agent dir and HOME), because /mcp login in RPC sends the URL as an info notice that the sanitizer hides. The assistant disables codemode, so each server needs exposure: "direct"|"deferred". Approval is enforced by an extension loaded with -e that hooks tool_call (it can block, covers nested calls, and asks through ctx.ui.confirm → ask card). Pi documents agents running pi mcp add through bash, so the daemon must be the only writer of mcp.json and must validate it against the approved set (hash) before each start; propose_connector returns drafts only. Tool annotations are kept, so read-only means readOnlyHint AND a curated list (fail closed). Providers, from official docs on 2026-10-08: Notion https://mcp.notion.com/mcp (OAuth + PKCE); Linear's official remote server (OAuth 2.1 with dynamic client registration, or an API key); Atlassian Rovo https://mcp.atlassian.com/v2/mcp (OAuth 2.1, or an API token; the explorer's v1 URL was outdated); Slack's official hosted server requires a registered Slack app with client_id + client_secret; Gmail's official https://gmailmcp.googleapis.com/mcp/v1 is in Developer Preview and needs the user's own Google Cloud project and consent screen; Discord has only community servers (bot token). The import map (S20) found configs for Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, OpenCode, and Gentle Shell on this Mac (key names only).
  • L36 (2026-10-08, user decision): Slack and Gmail: "Guía paso a paso ahora (Recomendado)" — each user creates their own Slack app or Google Cloud project with an in-app guide; no Gentle Dot-owned OAuth apps for now.
  • L37 (2026-10-08, user): "ah no puede atender mas de una conversacion a la vez?" The assistant answers one conversation at a time (one RPC agent process with one active session). An agent pool (one process per active conversation, capped at 3) was proposed with its costs; user: "dejemos uno solo por ahora". Decision: keep a single active conversation; no spec added.
  • L38 (2026-10-08, user): "por ahora entonces si es una sola sesion no pongas la capacidad de hacer nuevos chats, dejalo como esta pero ocultalo que sea un solo chat continuo te parece? mi unico miedo es que pasa cuando compacta mucchas veces y asi, si no sera mucho contexto". Evidence (pi-coding-agent docs/compaction.md:20-45, session-format.md "Context Building" and "CompactionEntry"): auto-compaction triggers when context > window - reserveTokens (16384) and keeps about 20k recent tokens plus the latest summary, so the model context stays bounded. Detail degrades with repeated summaries (Engram mitigates that), and the session file keeps every entry, so it grows forever on disk. Hence S22 rotation.
  • L39 (2026-10-08, user): "y que pasa con engram? veo que no usa el engram total puede ser ? el global qiuero decir, estaria bueno que use el global y que siempre escriba en un proyecto llamado gentle-dots" then "gentle-dot perdon". The parent found that Engram's Go code reads ENGRAM_DATA_DIR and ENGRAM_PORT and has ENGRAM_PROJECT with a process_override project source, and that the Pi plugin reads ENGRAM_URL, ENGRAM_PORT, and ENGRAM_BIN, with project override support unverified. T20 was launched in parallel with T17.
  • L40 (2026-10-08, T17 evidence): RED → GREEN in two batches (18 failing → 125/125; S21 10 failing → 136/136), plus two bridge races found and fixed: a send right after new_conversation was wiped by the late history (reproduced 5/5 with a Playwright WS capture, 4/4 after the fix), and stale history is re-read up to twice. Exceptions: no Playwright RED for the new e2e, and the gateway race test was written after the fix. Final: 227 tests, typecheck, e2e 12/12 in two consecutive runs; lint is green after the parent excluded test-results and playwright-report in biome.json. Decisions: the primary button text is #1a1218 on #f095c8 (8.60:1; white is 2.13:1, the site's #f6eff3 1.89:1); the daemon does abort-then-switch with a 10 s bounded wait; errors and stops appear inline (message_done.error/stopped) and the error toast was removed. Open: design.md §4/§6 still describe the old payloads and violet tokens; a residual race (a get_history requested before new_conversation arriving after the new history). The user's app was rebuilt and restarted with T17.
  • L41 (2026-10-08, T20 decision): The writer found that the Pi memory plugin has no per-client project override, and ENGRAM_PROJECT applies to the engram server process, so it would hijack every client of the global server. The supported per-directory override is <workspace>/.engram/config.json {"project_name":"gentle-dot"} (Engram detection step 0, source config). Parent decision: the agent's cwd is always the daemon-owned workspace that holds this config. A user-configured workspace becomes a preferred-folder instruction (an extra --append-system-prompt) instead of the cwd, so nothing is written into user folders and the project is always gentle-dot. Approved editing of the env-assertion block in gateway.test.ts (T17 finished).
  • L42 (2026-10-08, T20 evidence): The child gets ENGRAM_DATA_DIR=<real ~/.engram> (gentle-engram 0.3.0 accepts a server only when <data dir>/.instance-id matches /health, index.ts:580-649), and the user's ENGRAM_URL/ENGRAM_PORT are kept. The project comes from <data>/workspace/.engram/config.json {"project_name":"gentle-dot"} (Engram detection step 0). A custom workspace becomes preferredFolder (an extra --append-system-prompt, never the cwd, nothing written in it). GENTLE_DOT_ENGRAM=private keeps the old private mode. RED 9 → GREEN 49/49. Probe against a stand-in global server: observation saved under gentle-dot, HOME hash identical, no private .engram created. Parent follow-ups (400ec73): the identity now pins memory to project gentle-dot (an explicit project in mem_save would otherwise win), and the stale GENTLE_DOT_WORKSPACE was removed from the Dockerfile, entrypoint, and e2e serve (sh -n OK, e2e 12/12). Docker build not re-run because Docker Desktop is stopped. On the user's Mac: the app restarted, the old private Engram (PID 93883, ~/.gentle-dot/home/.engram) was stopped by PID with nothing listening on 7438, and the global server resolves ~/.gentle-dot/workspace to project gentle-dot (source config). Memories saved earlier in the private store were not migrated (test data from today; the DB file is left in place).
  • L43 (2026-10-08, user): "un item mas al final de todo, que haga como codex y que permitas controlar tu computadora, que pueda manejar tus aplicaciones y tal". Recorded as S24/T21, the last item, after T19, T18, and T14.
  • L44 (2026-10-08, T19 evidence): The UI hides conversations behind ready.features.conversations (default off; GENTLE_DOT_CONVERSATIONS=1), and the daemon refuses new_conversation/open_conversation with conversations_off when off. Rotation (rotation.ts) triggers on 20 MB or 10 compactions, only when idle. The handoff is a hand-written Pi v3 session file (header with parentSession plus one custom_message, display:false, customType gentle-dot.handoff) followed by switch_session, with a verification of get_state and get_messages and a new_session fallback; formats are copied from session-manager.js with paths cited. History now reads the session files (including pre-compaction messages) with 100-message pages, hasEarlier, and get_earlier across the chain persisted in state.json. Finding: Pi emits message_end on stdout before writing the entry to the file, so history reads do a get_state round trip first. The Stop pill is centered (inline-flex, no vertical padding; measured 7/8 px). Design §4/§6/§7/§9/§10 and the README were updated. RED → GREEN for protocol, UI, and the daemon module; e2e written after the code. 260 tests, typecheck, lint, e2e 14/14. Parent real-engine probe (bundled engine, temp HOME/data, private Engram on a temp port, stopped by PID): switch_session to the hand-written file returned cancelled:false, get_state.sessionFile equals the file, and get_messages holds one custom message of type gentle-dot.handoff with the probe text. Open: the tray still shows "New conversation" (apps/desktop follow-up); with the flag on, conversation switches combined with the chain are untested.
  • L45 (2026-10-08, T18a finding): The T18a writer found rotation.test.ts flaky in full runs (4/4 runs on a clean git archive HEAD copy, each time a different rotation test timing out; 3/3 in isolation). The likely cause is a T19 race in bridge.ts: when the prompt response and agent_settled arrive in one stdout chunk, rotateWhenIdle runs while sending is still 1, so rotation is skipped until the next run settles. In production this only delays rotation to the next run, but it is a real defect. Decision: fix in T18a, test first, with a single in-flight rotation flag and at least 3 full runs. The parent's two earlier green full runs were luck.
  • L46 (2026-10-08, T18a evidence): Commit e49bd9d. Catalog: Notion, Linear https://mcp.linear.app/mcp (verified), and Atlassian v2, with curated read-only lists (fail closed). The daemon is the only writer of mcp.json (exposure: direct, read-only mode as toolExposure {"*":"hidden", curated:"direct"}, validated with the engine's own validateMcpServerConfig), restored by hash in a new supervisor prepareSpawn hook before every spawn. The guard policy JSON travels in the engine env (GENTLE_DOT_CONNECTOR_POLICY); missing or malformed means everything asks. Sign-in runs the engine CLI mcp login as a child process; the daemon completes a pasted redirect only after validating origin, path, and state, and cancel kills the CLI by PID. The approval guard (-e, tool_call hook) asks with a preview and blocks writes to control files; the real-engine probe found and fixed a /var vs /private/var bypass (realpath compare). Tray "New conversation" only with conversations on. The L45 rotation race was fixed test-first (rotationDeferred, 14/14 ×3; the writer reported 6 green full runs). RED → GREEN per layer; e2e written after the code. Parent re-ran: 312 tests twice, typecheck, cargo 42/42. Risks reported: bash blocking is textual, the agent can read mcp-auth.json, separate sign-in queues for models and connectors, and the browser may open twice. Parent-observed UI defect: the onboarding banner overlaps the first chat message. Risk tier high (security) → an independent adversarial verifier is running.
  • L47 (2026-10-08, T18a adversarial verification): 2 BLOCKERS, reproduced with the real engine (scratch copy, fake model/MCP/CLI, temp HOME). B1: with one bash call from <data>/workspace, the agent edited ../agent/mcp.json and ../connectors.json (enabled, recomputed sha256 or deleted hash), ran kill $PPID, and the supervisor restarted it; prepareSpawn trusted the altered registry and notion-search ran without a card on a connector the user had disabled (catalog-only; sends still ask after the mode flip). B2: read ../agent/mcp-auth.json (the guard checks only write/edit) and cat returned the token to the model. Root cause: textual bash blocking (approval-guard.ts:113-119) and an integrity hash stored in the same agent-writable file. Passing: write asks/decline blocks/no UI blocks; read-only hides write tools; curated-list fail-closed; malformed policy asks; sign-in redirect validation against all variants; one flow at a time; clean logs; mcp.json valid and restored; rotation 3/3 at HEAD vs 2/3 failing at baseline; typecheck, lint, e2e 16/16, cargo 42/42. Advisories: gentle-pi subagent children do not load the guard (subagent tools not exposed today); project .pi/mcp.json and session_start re-reads bypass the restore; the preview shows at most 6 fields × 160 chars; orphaned private Engram; fetch follows redirects; the guard loads only with agentHome. Fundamental limit: the agent runs as the user with unrestricted bash, so file-based protections cannot be airtight.
  • L48 (2026-10-08, user decision): "A ahora, B después (Recomendado)". T22 hardens now (residual risk documented: an agent tricked into using unrestricted bash can still read what the user's account can read). T23/S25 adds the daemon-side MCP proxy with Keychain tokens as stage 2, before publishing to other users.
  • L49 (2026-10-08, T22 finding): The T22 writer found that <data>/token (the daemon access key) was readable by the agent. With it, the agent could open the WebSocket like a window and answer its own approval asks (ui_response confirmed:true), a self-approval bypass. T22 blocks reading it (best-effort, like the other file checks); a real fix is in S25's scope (the approval channel outside the agent's reach).
  • L50 (2026-10-08, T22 finding + user decision): The real-engine probe showed that the assistant's engine DOES offer subagent_* tools (and orchestrator_*). Children are spawned as pi --mode rpc --tools <agent tools> with only gentle-pi's internal child extensions, without the approval guard, and they inherit env, so a child's bash/read is unguarded and it could use connector tools without approval. Forwarding the guard to children has no supported hook. User: "Desactivar subagentes hasta S25". T22 sets GENTLE_PI_AGENTS=0 in the engine env, with a probe assertion; S25 must cover children before subagents return. This narrows S3's "everything like Gentle Shell" until stage 2.
  • L51 (2026-10-08, T22 evidence): Commit 8651bba. B1: connector state is in memory (loaded once), prepareSpawn renders from memory, the self-referential hash was removed, and changes to connectors.json/mcp.json are reverted while a <workspace>/.pi/mcp.json is removed (symlink-safe). The watcher uses fs.watch, a 1 s poll (macOS FSEvents missed early writes), a check before each spawn, and one after each settle; the UI shows "A change to your connectors was blocked.". B2: the guard blocks read/grep/find/ls/write/edit on auth.json, mcp-auth.json, models.json, mcp.json, connectors.json, <data>/token, and the guard file (realpath, dev:ino, case, @), and bash/powershell by basename after unquoting (documented best-effort). A3: full preview up to 8000 chars. A5: redirect: manual. A4: the private Engram is stopped by verified PID. A1: subagents disabled (GENTLE_PI_AGENTS=0; the probe asserts no subagent_*/orchestrator_*). The onboarding banner sits in the chat flow. Real-engine probe: before (base copy) notion-search ran with 0 cards and the files stayed altered; after, no MCP calls, files restored, notice shown. 328 tests ×3, typecheck, lint, e2e 17/17. Residuals documented for S25: edits while the daemon is down, edit-and-kill races, obfuscated bash, presentText normalizing whitespace in previews, A4 not exercised with a real Engram. The user's app was rebuilt and restarted with T22. Limited recheck running.
  • L52 (2026-10-08, limited recheck + user decision): The recheck was partial: the verifier's environment safety classifier blocked its own attack harness. Verified: the repo's real-engine probe of the original B1 attack passes (connector stays disabled, files restored, no MCP calls, notice shown); no subagent_* tools; 328 tests ×2, typecheck, lint, e2e 17/17. Not independently exercised: B1 variants (symlinked .pi, rename, python/node writers, the enforce→read race; code reading shows a second barrier, since the guard denies servers that are disabled in the in-memory policy) and B2 at runtime (covered only by the author's tests). The parent removed the scratch copies and confirmed that no probe Engram was left running. "Needs your decision" presented; user: "1" — accept T22 with the documented gaps and continue; S25 remains the real fix before any public release.
  • L53 (2026-10-08, T18b evidence): Commit 994abce. Guided connectors: Discord via @pasympa/[email protected] (pinned; MIT, 18 releases, last 2026-09-03, readOnlyHint on its tools; the alternatives have not published since 2025), Slack official https://mcp.slack.com/mcp (no DCR; localhost redirect only with PKCE, so the client secret is optional), and Gmail official remote (Developer Preview) with the user's own app; each has an in-app guide, secret prompts, and a fail-closed curated list. Security finding: the engine resolves env/headers/oauth.clientSecret with resolveConfigValue (! runs a shell command, $NAME expands), so every typed or imported literal is escaped ($→$$, !→$!). S19: a propose_connector tool in the guard sends drafts through ctx.ui.setStatus (a namespaced key read only by the daemon); approval adds the server read-only with all tools hidden; at most 5 pending drafts, kept in memory only. S20: a read-only scan of 7 client formats (sse rejected, dedupe, names only in payloads, source hashes unchanged). Also coalesced burst restarts ("Agent is not ready"). RED: 27 failures plus a missing module → GREEN 362 tests ×2, typecheck, lint, e2e 18/18; T18b probe passed (the engine offers propose_connector; drafts and imports are accepted by mcp list; imported tools are hidden from the model; secrets are absent from scan payloads; the private Engram is stopped). Parent re-ran 362 tests + typecheck and reviewed the screenshots (import list with key names only; Slack guide). Risks: Slack and Gmail not tested end-to-end (no credentials; Gmail needs the Developer Preview), masking is heuristic, env refs in imported args are copied literally, and the S25 residual now includes a custom command server injected while the daemon is down. Risk tier high → independent verifier (code review + repo probes) running.
  • L54 (2026-10-08, T18b verification): No path found where a value runs a command or expands a variable, or where a connector is added without approval. Escaping covers exactly what the engine resolves (env, headers, oauth.clientSecret; command/args/cwd/url are not resolved), verified in the real mcp.json (dummy-token-$!x → dummy-token-$$$!x). Drafts pass (state changes only through a window's connector_draft_reply; approved servers are hidden and read-only; max 5). Import passes (read-only, 7 formats, sse rejected; the isolated HOME means ~-relative credentials are not found). Guided connectors pass (pinned Discord; Slack/Gmail oauth fields valid; the guide's redirect equals the engine's redirect_uri). T22 protections are intact. Runs: 362 tests ×2, typecheck, lint, e2e 18/18, both probes. BLOCKER B1: summarize() (connectors.ts:1434-1450) leaks credentials inside imported args to the connector_imports and list() payloads (--header "Authorization: Bearer …", URL query keys, userinfo in connection strings, tokens in URL paths); local authenticated windows only, not the model or logs. Advisories: ${input:} in draft args or URL puts secrets in argv or a model-chosen URL; Gentle Shell !command imports run commands without a visible note; the S25 residual includes a custom command server injected while the daemon is down (medium; not blocking before S25); Gmail's gmail.compose scope can send; unpinned transitive Discord deps; minor UI and dedupe issues; the self-approval risk via the daemon token (S25). The parent cleaned the scratch dirs. One correction batch is running (B1 + advisories A–E).
  • L55 (2026-10-08, T18b correction + recheck): Commit 63c8e9c. B1: summarize() masks structurally (URL userinfo dropped, query values masked, long opaque path segments masked, --header/-H and auth-like header values masked, values after secret-like flags masked, prefix masks kept). Advisories: A, draftSecretProblem in the guard (the model gets "Secrets can only go into environment variables.") plus parseDraft as a second layer; B, notes for values that run a command or read an env var (never showing the command text); C, accurate Gmail/Slack/Discord copy; D, UI drafts capped at 5 and cleared on ready; E, an S25 bullet for custom/imported servers edited while the daemon is down. RED 9 → GREEN; 370 tests ×2, typecheck, lint, e2e 18/18. Parent limited recheck of B1: the five verifier cases (mcp-remote --header Authorization: Bearer, ?api_key=, postgresql://user:pass@, a token in the URL path, --api-key) produced summaries without the secret while keeping command, host, path, and --port 8080; the full suite passed again. Known over-masking (flags containing key or pass). T18 complete.
  • L56 (2026-10-08, user decision on Linux delivery): "Compilar desde el código (Recomendado)". The app is not self-contained today (build.rs records the node path and the daemon script path of the build machine's repo), so packages built here would not run on the tester's machine. S13 delivery changes: instead of .deb/AppImage artifacts, a per-distro setup script (Debian/Ubuntu via apt, Omarchy via pacman) that installs dependencies (Node 24, pnpm, Rust, WebKitGTK and the Tauri system deps), builds, and launches the app on the tester's machine, plus the testing checklist. A Linux compile check runs in a Debian container here. Self-contained packages (also needed to give the Mac app to others) are deferred.
  • L57 (2026-10-08, Docker check): The server image builds at HEAD, and a container smoke run reached idle, setup 3/3, no token in the logs (only the key file path), and the workspace .engram/config.json with gentle-dot. T14 was launched (from source, with setup scripts, the tester checklist, and a Debian container build check).
  • L58 (2026-10-08, T14 evidence + parent fix): Commit 3c002e3. A pure platform.rs decides the Wayland strategy: GDK_BACKEND=x11 set before GTK on GNOME Wayland when XWayland exists, and no self-positioning on native Wayland (Hyprland), where window rules are used. Single-instance with gentle-dot --toggle; Linux tray glyphs; an opaque panel without vibrancy; .cargo overrides limited to macOS. macos-private-api must stay cross-platform (tauri-build allowlist). Setup scripts: Debian/Ubuntu via apt and Arch/Omarchy via pacman; Node 24 from the official tarball with a sha256 check in a private dir (no sudo), pnpm via corepack, rustup; sudo only for packages, after a prompt; plus an uninstall script. Hyprland snippets in Lua (Omarchy v4.0.4, cited) and hyprlang (0.53–0.54 and the older forms). docs/linux-testing.md is the tester checklist. Container evidence (debian:bookworm arm64): the real setup script ran end to end and was idempotent, cargo 53/53, a 16 MB release binary; an X11 smoke run under Xvfb showed the Dot at 72×72 on the right edge and --toggle showing and hiding the panel, with one instance. It found and fixed two GTK bugs (a non-resizable 72×72 window drawn at 200×200; a never-shown panel reporting 0×0). Arch package names were checked (libappindicator-gtk3 → libayatana-appindicator). Not verified: a real GNOME/Hyprland session, an Arch end-to-end run, x86_64. Writer-reported engine issue, "Pi requires the pi executable in PATH" on clean machines, was fixed by the parent test-first (RED → GREEN): the daemon prepends the bundled node_modules/.bin (pi 1.0.4) to the engine PATH. Real-engine check with no pi on PATH: 0 errors, the 4 companion packages installed, setup 3/3. 376 tests, typecheck, lint.
  • L59 (2026-10-08, user): "lo de controlar la computadora por ahora solo para mac". S24 is scoped to macOS; the read-only map was launched.
  • L60 (2026-10-08): T21 map returned (no web access in the explorer; key facts re-checked by the parent): Pi forwards MCP image blocks to vision models and resizes above 2000×2000 (pi-mcp README:32, pi-coding-agent docs/models.md:70-89); Anthropic recommends downscaling screenshots (claude.com best-practices article); ad-hoc signed builds lose TCC grants on every rebuild because the grant is keyed to the cdhash (several public fixes, e.g. github.com/ericwbailey/what-cant-i-press 750a2ff); ScreenCaptureKit SCContentFilter/SCScreenshotManager can exclude windows (developer.apple.com). Decision: option (a), a native Rust helper in the Tauri app serving MCP over loopback, because only there do the grant, panic, blocklist, and risky-action checks live outside the agent (the engine guard and the WS are reachable by the agent, L49). Parent defaults, open to the user's change: 30-minute sessions, panic ⌥⇧Esc, 10 actions per second, long edge 1280 px. S24 refined into S24.1–S24.7; split into T21a (Rust) and T21b (TypeScript), launched in parallel with disjoint surfaces; contract: Tauri commands computer_endpoint → {url, token} | null, computer_permissions → {accessibility, screenRecording}, computer_request_permission {kind: "accessibility"|"screenRecording"}, computer_stop, computer_status → {active, endsAt?}; event computer://state {active, endsAt?, reason?: "granted"|"stopped"|"panic"|"timeout"}; client messages computer_register {url, token} and computer_unregister.
  • L61 (2026-10-08): T21a scope additions approved by the parent: build.rs COMMANDS + generated permissions/**; rustup component add clippy; minimumSystemVersion 11.0 → 12.3 (ScreenCaptureKit is strongly linked; screenshots need macOS 14, checked at runtime, computer_permissions adds screenshotsSupported). T21b done (5b59c0e). Accepted risk, deferred to S25: any authenticated WS client (the agent included, L49) can send computer_register and repoint the built-in entry; an Origin check would not help because curl can forge it, and the helper still enforces the native grant, so control is never gained this way. The bearer key sits in mcp.json (file tools blocked, bash best-effort) and only reaches the helper.
  • L62 (2026-10-08): T21a done (176f547): ScreenCaptureKit capture excluding our pid's windows, JPEG ≤1280 px, NSAlert dialogs via tauri-plugin-dialog; clippy clean; macos.rs not yet run on a real Mac. Seam check found that the engine's MCP client cuts a call after 30 s (pi-mcp client.js:3) while the helper blocks on native dialogs; Pi reads a per-server timeout in seconds (pi-coding-agent mcp/runtime.js:160), so the built-in entry now renders timeout: 300 (RED observed, GREEN 400 passed; 3ebfe5c). Independent verifier launched on 85bd25d..3ebfe5c.
  • L63 (2026-10-08): independent verifier on 85bd25d..3ebfe5c: all checks green (400 TS, 145 Rust, clippy, typecheck, lint); S24.2, S24.3, S24.5–S24.7 PASS; S24.4 FAIL with B1 (pointer owner resolved to the WindowServer cursor window, blocklist bypass) and B2 (same-point drag skips risky confirmation); critical R1 (daemon spawned without TCC responsibility disclaim, so the agent's bash likely inherits the app's grants, contradicting S24.1). One correction batch launched: B1, B2, R1 (posix_spawn with responsibility_spawnattrs_setdisclaim), and advisories A1 (space on send-like button), A2 (control characters), A4 (system dialog blocklist + focused-element pid), A5 (fresh screenshot per session), A6 (rename a saved computer connector instead of dropping it). Recheck limited to those items.
  • L64 (2026-10-08): correction batch 3e29886 (B1, B2, R1 disclaim via posix_spawn, A1, A2, A4, A5, A6). Recheck: B1, B2, A1, A2, A4, A5, A6 FIXED; R1 correct in code (the child is its own responsible process), real-Mac confirmation pending. New advisories closed inline by the parent: N1 (type refuses a space or newline after a tab, since the focus may have moved onto a send button) and N2 (Unicode line/paragraph separators and C1 controls refused); RED 2 failing → GREEN, cargo test 169/169, clippy clean.
  • L65 (2026-10-08, user, real-Mac run): screenshot of the panel toast "A change to your connectors was blocked." right after the app started. Reproduced: daemon.log showed "the desktop app registered computer control" then, 34 ms later, "put back …/connectors.json". Root cause: enforce() treated a missing file as untouched only with no saved connector and no computer helper, for both files, so on an install with no saved connector the registration made the missing record count as tampering. Fix d099a89: each file is judged by what it would hold (RED reproduced the exact log line; GREEN 402 passed; the key-leak test no longer assumes the record exists). App restarted: registration logs no revert. Note: the registration restarts the engine during startup and logs one "agent failed to start: Agent is not ready; retrying", which recovers to idle.
  • L66 (2026-10-08, user): real-Mac run: "anda muy bien pero lento, bastante lento". Measured from the engine session: the model takes 5–9 s per step and the helper 0.8–1 s per action (300 ms settle + capture + JPEG in an unoptimized debug build); mentioning five people took ten round trips (~70 s); one type with an emoji took 18 s. Proposed batch, faster helper, lighter screenshots; user: "hagamos todas si". Recorded as S24.8/T24.
  • L67 (2026-10-08, user): "y mete un modo yolo". Asked what it skips and how it turns on; answers: "Solo las confirmaciones (recomendado)", "Interruptor, se apaga solo (recomendado)". Recorded as S24.9, added to T24.
  • L68 (2026-10-08, user): "ah y pon un ocultar la rosa, que la puedas abrir bajo demanda y un full screen para el chat cuando quieras". Checked: the shortcut and tray "Open" already open the panel without the rose (shell.rs:280,397). Recorded as S26/T25, queued after T24 because both edit shell.rs and packages/ui.
  • L69 (2026-10-08, user): "y se le puede poner su propio mouse? en vez de utilizar el mio?". Answered: macOS has one cursor per session; recommended Accessibility actions first, keyboard to an app, real mouse with cursor restore as the last resort, and a visible ghost pointer; a separate cursor needs a VM.
  • L70 (2026-10-08, user): "que sea como un glow rosa". Recorded as S27/T26 with the ghost pointer as a pink glow, queued after T25.
  • L71 (2026-10-08, user): "no no no lo pongas lo del segundo mouse no te preocupes". S27/T26 dropped, the pink glow included; nothing was implemented.
  • L72 (2026-10-08, user): "lo del glow rosa si eso si para saber que esta haciendo el agente y solo mientras este interactuando". The glow alone is back as S28/T27, queued after T25.
  • L73 (2026-10-08): T24 done (952830e): batch through one per-step perform path with pacing instead of refusal and 100 ms between steps; SETTLE 150 ms; JPEG 65 (synthetic UI 254,592 → 211,264 bytes); dev profile opt-level 3 for image/zune and 1 for the app crate (JPEG encode 200 → 40 ms, accepted by the parent: the generic encoder monomorphizes in our crate); yolo in memory, native confirm, 1 h expiry, tray item and UI switch. The 18 s type is not the helper's doing (two chunks + 150 ms settle, pinned by a test); likely a native confirmation or the model. Open: Stop and panic end the session but leave yolo on. Independent verifier on 8e99e7a..952830e and the T25 writer launched in parallel.
  • L74 (2026-10-08, user): "si y si" to: the panic shortcut also turns yolo off (recommended), and removing the verifier scratch dirs. Panic now ends yolo with or without a session; Stop leaves it on (RED → GREEN, cargo test 195/195; the old test pinning the previous rule was updated). Removed $TMPDIR/tmp.WlsFPGwLSP (13 GB), tmp.I3NLBN7JIb, tmp.49ULuteyHt.
  • L75 (2026-10-08): independent verifier on 8e99e7a..952830e (detached worktree copy): cargo test 193, clippy clean, pnpm test 407, typecheck, lint, 9/9 probes; S24.8a–c and S24.9 PASS, S24.8d partial (the 18 s needs a real-Mac look); no blockers. Advisories kept for later: A1 100 ms between batch steps narrows the post-tab focus re-read window (check with slow/Electron apps); A2 the yolo and grant dialogs default to the affirmative button; A3 a Stop right after yolo expiry can report a stale yolo for up to 1 s; A4 a batch with many confirmations can outlast the engine's 300 s timeout; A5 wait with a risky intent asks nothing (baseline). Scratch removed.
  • L76 (2026-10-08): T25 done (8b274ed): rose_hidden in desktop.json, tray item toggling "Hide the rose"/"Show the rose", header buttons, panel placement without the Dot (last spot on the current display, else right edge), full screen to the work area with restore, ⌘⇧F, Esc precedence (handled → restore full screen → hide). Risk tier medium (window behavior covered by tests; no high-risk item), so no separate verifier. Gaps: the tray icon does not show in-control (the panel banner and the yolo tray check do); real-Mac checks pending (set_size on the non-resizable panel, mixed-scale displays). T27 launched with the explicit constraint that pointer-owner resolution must skip the overlay window.
  • L77 (2026-10-08): T27 done (f01dfdf): marks after each successful post, a click-through glow window (level 102, all Spaces, orderFrontRegardless/orderOut), owner resolution skips only the overlay's window number, the AX hit test re-runs inside the owning app when it lands on the overlay, a glow capability with listen/unlisten only, reduced motion static. Verifier launched on a2fbc65..f01dfdf; app bundle rebuilt for a real-Mac round of T24, T25, T27.
  • L78 (2026-10-08): independent verifier on a2fbc65..f01dfdf: cargo test 217, clippy, pnpm test 426, typecheck, lint, 11/11 probes; owner resolution, element re-hit-test, marks, overlay/capabilities, and UI PASS; no blockers. A1 (an empty type marked a glow without input) fixed inline: RED → GREEN, cargo test 218, clippy clean. Advisories kept: A2 reduced motion read once at mount; A3 an unresponsive app under the overlay yields an empty AX chain (same as baseline for hung apps); A4 out-of-lock state reports can reorder grant/end (baseline); A5 ?surface=glow on the web is a blank page. App rebuilt and restarted for the real-Mac round (permissions must be granted again after a rebuild).
  • L79 (2026-10-09, user): "anda increible!! ahora lo unico que te falta... es hacerle un instalador a la app" (after a real-Mac run where Gentle Dot controlled Ghostty and typed into this session). The yolo mark became an amber rim (28049af, user: "y cambialo por mas integrado"). Checked: no codesigning identity ("0 valid identities found"); the app embeds build-machine Node, daemon script, and PATH; Engram is the user's own binary.
  • L80 (2026-10-09, user): installer for "Solo para mí y gente de confianza", platforms "macOS y Linux". Recorded as S29/T28; read-only map launched.
  • L81 (2026-10-09): T28 map (explorer without web; parent re-checked): ship a full Node 24 distribution (first run does npm install --prefix <agentDir>/npm; Discord uses npx); SEA and bun compile do not fit (the daemon re-executes process.execPath with gentle-shell.mjs and the pi CLI; gentle-pi loads extensions/skills from disk via jiti); per-platform dependency installs (ffi-rs N-API, fff-bin, the gentle-ai Go binary from gentle-pi's postinstall with mode 0700); the pnpm .bin/pi shim is not relocatable; Engram ships as a fallback binary (MIT, goreleaser darwin/linux amd64/arm64). Verified on the web: macOS 15 removed the Control-click Gatekeeper override (developer.apple.com/news/?id=saqachfa); a persistent self-signed identity keeps TCC grants across updates while ad-hoc does not (github.com/duongductrong/Snapzy docs/SELF_SIGNED_CERT.md). Plan as T28a–T28g with the S29.5 layout as the contract; T28a and T28b launched in parallel. Risks: the engine writing into a read-only package tree, first-run network dependence, Engram version skew on ~/.engram, amd64 under emulation.
  • L82 (2026-10-09): T28a done (84a38a4). T28b done: build:runtime (esbuild ESM bundle of daemon + protocol, engine external, lockfile-exact package.json, relocatable bin/pi), runtime-mode UI/engine/pi paths, Engram resolution passed as ENGRAM_BIN, --self-check; smoke: bin/pi --version → 1.0.4 from a hand-built runtime. Advisory: the login-shell lookup for Engram runs synchronously with a 5 s timeout at daemon start. T28c launched; its smoke must use a private Engram on a temp port so it never touches the user's 7437 server or ~/.engram.
  • L83 (2026-10-09): T28c done: pnpm package:runtime --target darwin-arm64 stages 346.4 MB (408 MB on disk: daemon/node_modules ~267 MB, Node ~128 MB, Engram 20 MB) in 11 s with cache; Node 24.14.1 and Engram 3.2.1 (latest release; the user's is a dev pseudo-version) checksum-verified; refuses non-host targets. Clean-HOME smoke (scripts/package/smoke-runtime.mjs): idle in 23.1 s, private Engram on a temp port from <rt>/bin/engram, first-run npm install of the companions, 0 runtime files changed. Read-only smoke: idle, 0 changes, no EACCES/EPERM (risk R1 cleared). Fix: the engine gets a private TMPDIR when none is set (jiti cached in shared /tmp/jiti). Notes for later: transitive deps resolved by npm at staging; .nvmrc says only "24"; the engine log says "Gentle AI sends anonymous usage metrics" (privacy/white-label to review); size trims possible (esbuild ~21 MB, @shikijs/cli 17.6 MB). Scratch removed.
  • L84 (2026-10-09): T28d done (546aa92): pnpm package:mac → Gentle Dot_0.1.0_aarch64.dmg 111.2 MB (app 368 MiB); runtime injected with tauri build --config (a resource in tauri.macos.conf.json would be copied on every cargo build and fail without a staged runtime); nested Mach-O signed inside-out except gentle-ai, whose sha256 gentle-pi checks; codesign --verify --deep --strict valid; mounted-DMG copy smoke idle in 23.1 s and read-only smoke ok, 0 runtime changes; ad-hoc DR changed between builds (57c5… → 43e0…). Biome now ignores build/ and .cache/. Quarantine/Gatekeeper on a downloaded DMG still needs a real second Mac. User consented to T28e ("Sí, crealo"): identity "Gentle Dot Local Signing" created in the login keychain; the trust step was not needed (codesign signs with the untrusted identity and the DR reads certificate leaf = H"4611e0a8…"), so the script skips it. Two signed builds running to compare the DR.
  • L85 (2026-10-09): T28e proven: two release builds signed with "Gentle Dot Local Signing" both have designated => identifier "dev.gentleman.gentle-dot" and certificate leaf = H"4611e0a8b9c71afb97cafdbf79dccf4816d80e6b", codesign verify ok, DMG 116,354,516 bytes. The user saw many keychain prompts on the first signed run (each of 8 nested signatures asks); fixed by "Permitir siempre" once. User (screenshot, the assistant saying its memory search is limited to Gentle Dot): "porque esto? si tiene access a la memoria global?". Cause: identity.md also scoped searches to gentle-dot, beyond S23's "siempre escriba". Fixed: save to gentle-dot, search with all_projects: true (RED 2 failing → GREEN, pnpm test 478).
  • L86 (2026-10-09, user): "te pido una cosa mas?... posibilidad de audio a chat, cosa que puedas "hablar" con el agente, y que esto use la sub o api que ya tenes conectada"; choices "Conversación por voz", "Botón de micrófono". App restarted so the engine took the memory-recall identity (verified in ~/.gentle-dot/identity.md). Recorded as S30/T29; read-only map launched (which connected accounts can transcribe/speak; fallbacks; mic in WKWebView).
  • L87 (2026-10-09): T28f done: pnpm package:linux --arch arm64|amd64 → gentle-dot_0.1.0_arm64.deb 116.1 MB (38 s cached), gentle-dot_0.1.0_amd64.deb 119.1 MB and gentle-dot-0.1.0-1-x86_64.pkg.tar.zst 107.5 MB (emulated, 2 min 20 s); pnpm package:linux:check passed in clean debian:bookworm and ubuntu:24.04 (both arches) and archlinux (x86_64): daemon smoke over the root-owned tree, read-only smoke, and an Xvfb launch of /usr/bin/gentle-dot to /health idle (18–23 s native, 40–58 s emulated), 0 runtime files changed, GLIBC_2.34 max. pacman needed DisableSandbox under emulation. No Arch aarch64 (no official image). Docker images (gentle-dot-linux-builder:arm64/amd64, gentle-dot-arch-builder) and cache volumes kept for rebuilds. T28g done: docs/install.md Linux section and the signing identity note; root scripts package:mac, package:linux, package:linux:check. Pending real-hardware checks: a downloaded DMG on a second Mac (quarantine, Open Anyway), real x86_64 and GNOME/Hyprland desktops.
  • L88 (2026-10-09): T29 map: one connected account, OpenAI "Sign in with ChatGPT" OAuth (auth.json provider ids only, no secrets read); pi-ai 1.0.4 has no audio input or transcription; Info.plist lacks NSMicrophoneUsageDescription; WS messages cap at 1 MiB. Verified on the web: Sign in with ChatGPT does not support the transcription API or audio input (OpenAI SIWC preview limitations); Codex CLI removed its voice transcription (openai/codex PR #16114). User choice: "Voz de macOS, y API key si la conectas (recomendado)". S30 refined into S30.1–S30.4; contract: Tauri commands voice_status → {available, reason?}, voice_start, voice_stop → {text}, voice_cancel; events voice://partial {text}, voice://level {level 0..1}, voice://error {message}; WS voice_transcribe {mime, data(base64)} → voice_transcript {text} or error voice_unavailable; voice_speak {text} → voice_speech {mime, data}; capabilities voice: {transcribe, speak} in the ready/accounts payload. T29a and T29b launched in parallel.
  • L89 (2026-10-09, user): "lo del certificado es algo inseguro para mi? porque sino la cuenta apple te pido ser develor por 99 al año, tambien, no seria mejor sino tener un modelo de ia local chiquito? Parakeet es muy muy bueno"; then "pero que sea opcional para descargarlo si solo quiere". Verified: Parakeet TDT 0.6B v3 covers 25 European languages incl. Spanish (Hugging Face, NVIDIA NIM matrix); sherpa-onnx has an official Rust Parakeet example; FluidAudio offers a CoreML build (~110× real time on Apple Silicon). Recorded as S30.5/T29c after T29a; T29a told to keep a pluggable backend over native PCM capture.
  • L90 (2026-10-09): incident: commit 70c3f69 wrote back a 50 KB-truncated read of this document (log L12–L87 lost from HEAD, reported by both T29 writers). Restored from 70e5498 and the S30/L88/L89 changes reapplied with a full-file edit; lesson: never rewrite this document from a paged read.
  • L91 (2026-10-09, user): "y una cosa mas :) la posibilidad de pasar archviso por chat, eso estaria genial"; choice "Solo yo al agente". Checked: the engine's RPC prompt accepts images?: ImageContent[] (pi-coding-agent rpc-types.d.ts:18-29) and the WebSocket caps messages at 1 MiB (daemon.ts:197), so uploads go over HTTP. Recorded as S31/T30, queued after T29b.
  • L92 (2026-10-09): T29a done (93ee637): SFSpeechRecognizer (SpeechAnalyzer is Swift-only), native 16 kHz mono capture behind a Recognizer trait, Info.plist usage strings, voice capability for the panel only. T29b done (fd7c25f): protocol voice messages and capability, daemon VoiceService using only an OpenAI API key (checkAuth("openai").type === "api_key"; the ChatGPT sign-in is refused), mic button, source choice, spoken replies, mute. The lint failure T29b reported was T29a's capabilities formatting, fixed. User: "una cosa mas, no veo para conectar NaN, fijate que si lo tenemos en gentle shell, hasta tenemos nuestro propio plugin". Cause: the Accounts screen's ModelRuntime runs without gentle-pi's extensions, and NaN is registered by extensions/nan-provider.ts. Fix 627f2ad: the daemon registers gentle-pi's NaN provider as a native provider, loaded through jiti 2.7.0 (gentle-pi ships TypeScript inside node_modules); RED → GREEN, pnpm test 521; jiti added to the runtime's dependencies. App restarted.
  • L93 (2026-10-09, user): "una cosa mas, da la posibilidad de cambiar los modelos en el chat mismo". Checked: the engine RPC has get_available_models, set_model, set_thinking_level (rpc-types.d.ts:45-56) and profile apply already switches live (profiles.ts:670-698), so this is mostly UI. Recorded as S32/T31, queued after T30.
  • L94 (2026-10-09): T30 done (8a79b48): POST /upload (Bearer only, WS Origin allowlist, CORS preflight), streaming limits with cleanup, private modes, one-time uploads, <attachments> block in the prompt hidden from history, images only when the model accepts them. Pending parent change after T29c: .disable_drag_drop_handler() on the panel's WebviewWindowBuilder (shell.rs) so HTML5 drops reach the panel. Verifier launched on 4e2efc5..8a79b48; T31 writer launched. User: "con deepseek v4 no anda de NaN". daemon.log: signed in to nan (api_key) at 10:57, then three assistant error: 524 Cloudflare "A timeout occurred" responses (11:02, 11:04, 11:15) with model nan/deepseek-v4-flash; a 524 means NaN's origin did not answer within Cloudflare's limit. Not yet known whether NaN was slow for everyone or for this request (large prompt/tools).
  • L95 (2026-10-09, user): "si lo pruebo desde pi funciona, en gentle dot esta MUY lento" (screenshot: computer control active in yolo mode, several "Something went wrong while answering"). Measured: the current session holds 91 screenshots, 11.1 MB of base64 images, and the engine resends the whole history with every call, which plain Pi sessions do not carry; the 524s are NaN's origin not answering within Cloudflare's limit on those heavy requests. Pi's extension context event can replace the messages sent to the model without touching the session (pi-coding-agent extensions/types.d.ts:659-662, 1047-1049). Recorded as S24.10/T32, inline after T31.
  • L96 (2026-10-09): T29c Rust done: sherpa-onnx 1.13.8 Parakeet backend preferred when installed (Apple Speech otherwise on macOS; parec/arecord capture on Linux), ModelManager with pinned asset (k2-fsa/sherpa-onnx release asr-models, 487,170,055 bytes, SHA-256 5793d0fd…16bf, matches GitHub's digest), atomic install, cancel, remove; commands voice_model_status|download|cancel|remove, event voice://model. Real model: English and Spanish clips transcribed correctly, load ~0.6 s, decode ~0.3 s; install 24 s; release binary 14.3 → 42.2 MB; ~0.7 GB RAM while loaded (unloads after 5 min idle). Risks: sherpa-onnx's build.rs downloads unverified static libs (pin with SHERPA_ONNX_ARCHIVE_DIR in the packaging scripts); Linux voice not compiled here; a model that fails to load returns empty text silently. Pending UI: download button with progress, a "Transcribing…" state while voice_stop runs.
  • L97 (2026-10-09, user): "no seria mejor solo enviar la ultima y de las anteriores mandar un resumen?" after "no te acueras que no podemos tener mas de una conversacion? esto es justo lo que te decia con lo de autocompactar y el costo". Parent's earlier advice to open a new conversation was wrong (one continuous chat, T19); the session (11.8 MB) had not rotated (20 MB or 10 compactions). S24.10 updated: keep only the last screenshot, summarize older ones from existing text, no extra model calls. T32 is first priority after T31.
  • L98 (2026-10-09): T31 done (d0b63ce): models_list/model_set, picker chip above the composer, choice saved to settings.json (Pi's RPC set_model does not persist). User: "o una bitacora" — answered that the one-line entries left in place of older screenshots, in order, are that log (tool calls must keep their results, so they cannot be removed). T32 done (d9f0fa8): screenshot-context.ts pruneScreenshots keeps only the last computer screenshot as an image and replaces each earlier one with [earlier screenshot omitted: <tool> in <app>], keeping the action's own text; hooked on the guard's context event; the runtime build now copies the guard's sibling extension. App restarted; the engine runs with the guard (-e …/approval-guard.ts), no extension errors.
  • L99 (2026-10-09, user): "no veo el chip ni el selector de voz ni nada" — the UI is bundled into the app and had not been rebuilt. Added .disable_drag_drop_handler() to the panel (2e9b087), rebuilt the debug app, and signed it with "Gentle Dot Local Signing" (designated => … certificate leaf = H"4611e0a8…", verify ok), so the grants given now survive the next rebuilds. T30 verifier on 4e2efc5..8a79b48: no blockers; advisories kept: A1 a user can type a fake <attachments> block that renders as chips (no privilege gained); A2 two raw sends can reuse one upload (UI clears chips); A3 unsent uploads expire only on the next upload and are orphaned by a daemon restart (docs say "removed after an hour"); A4 the 120-unit name limit can exceed ext4's 255 bytes with CJK, and can split a surrogate; A5 a pre-existing symlinked uploads dir is followed and chmodded (needs the agent's own bash); A6 nul.tar.gz not renamed (Windows only).
  • L100 (2026-10-09, user): "donde elijo lo de voz? lo de descargar el mdoelo" — the download UI did not exist yet. Added a Voice entry in the desktop Connectors screen (4848d24): current engine, "Download local voice model (487 MB)" only on request, progress, "Checking and unpacking…", Cancel, failure reason, "Remove local model (670 MB)"; RED (missing module) → GREEN 5/5, pnpm test 589. Rebuilt, signed with the stable identity, restarted.
  • L101 (2026-10-09, user): "el modelo no funciona muy bien (el de voz) tanto en ingle como en español no toma lo que digo, o mejor dicho lo toma pero es cualquier cosa"; "ahhh porque estaba usando el de mac era". Checked: the local model was not installed (the parent's app restart killed the first download, leaving .parakeet-…download), so Apple Speech was used; suspicion: our 16 kHz box-filter conversion degrades the audio (Parakeet decoded test WAVs correctly). The user downloaded the model again ("Using: local model (Parakeet)"), and asked about the "Built in" chip (it marks the Voice entry as part of the app; proposed showing the engine instead). User: "otra cosa, hasta que no hago click en el model arriba del input no toma mi modelo actual, aparece como que noengo ninguno". Cause: ready carries the model only when known (bridge.ts:155) and the computer registration restarts the engine at launch, so nothing told windows the model later. Fix: broadcast the models payload whenever the engine becomes ready (RED reproduced → GREEN, pnpm test 590).
  • L102 (2026-10-09, user): "bien pero, mientras estoy hablando no me aparece lo que se va enviar como cuando uso el bui† in de mac" — Parakeet transcribes correctly with our capture, so the earlier garbage came from Apple Speech with our converted 16 kHz audio (open: feed Apple the device's native buffers). Live text with the local model (e027b19): the recording so far (last 30 s at most) is decoded again every 0.8 s of new audio, skipping when the model is still loading or a decode runs (never blocking the audio thread), serialized with the final decode, and never shown after stop/cancel; RED (compile) → GREEN, cargo test 296. The Voice chip now names the engine ("Offline", "macOS", "Off"; 1396457). Rebuilt, signed, restarted (also applies the model-chip fix 7ab0c0a).
  • L103 (2026-10-09, user): "estaria bueno que si pongo el boton de grabar de nuevo en vez de enviarlo, que se quede escrito en el input para confirmarlo, como se comporta ahora estaria bueno que sea una opcion que uno elijiese". S30.4 updated: review in the input is the default; "Send when I stop" (switch in the recording strip, stored in localStorage) keeps send-on-stop; the mic button reads "Stop and review" or "Stop and send". e0010eb: RED 3 failing → GREEN, pnpm test 593 (the earlier tests now cover the option). Rebuilt, signed, restarted.
  • L104 (2026-10-09, user): "sube todo esto al repo y fijate el readme de gentle ai como es para copiar el estilo y hacer el propio para Gentle Dot"; choices "Gentleman-Programming/gentle-dot, público" and "main + feat/gentle-dot, y abro el PR (recomendado)"; then "usa la rosa con glow como banner". TruffleHog over the whole history: no verified secrets (unverified hits are masking-test fixtures). README rewritten in Gentle-AI's layout with an honest Security status; banner and divider rendered from the Dot's own rose strokes and glow (docs/brand/tools/build-banner.mjs); MIT LICENSE as in Gentle-AI. Created the public repo github.com/Gentleman-Programming/gentle-dot, main at the first commit 5ec596c, pushed feat/gentle-dot (153 commits, 286 files, +57,416), and opened PR #1 (single-pr). Merge is the user's decision.
  • L131 (2026-11-02): Review of #170 (egdev6): S24.10 is already "Prune old screenshots" (T32), and docs/design.md reused it for Windows; the Windows computer-control spec is S24.11. The same review asked the screenshot sentence to match #169: exclusion per screenshot, restored right after, not for the whole session.
  • L130 (2026-10-31): Review of #159 (egdev6) asked to record the keyboard guards in S24.4: the screenshot binding (#89), the recipient recheck per chunk, Return and key, and the 500 ms restore after a confirmation, all observable on macOS and Linux. It also found that a stop in the middle of type said "nothing was sent" next to "Typing stopped after 20 of 45 characters"; the guard now says "the rest was not sent" once characters landed. S24.4 rewritten; docs/design.md S24.4 updated.
  • L129 (2026-10-10): T23h committed (35b942b): server mode GENTLE_DOT_VPS=1 (image only; refuses without root); daemon root with SETUID, SETGID, CHOWN, FOWNER, DAC_OVERRIDE, KILL and no-new-privileges; engine dot (uid 1001), stdio connector servers dotmcp (uid 999) via setpriv with <data>/connector-home (they cannot see the engine's workspace; documented); data dir root 0711, token/state/connectors/secrets/PIN root 0600; secrets AES-256-GCM per record (id as AAD) under GENTLE_DOT_SECRETS_KEY, scrubbed from every child, old file secrets migrated once; web PIN pin_set/pin_command/pin_approve (6-12 digits, scrypt + salt, constant-time, 5 failures -> 15 min persisted lockout; declining needs no PIN); desktop with the app channel unchanged. RED across protocol (4), UI (5), env scrub, spawn uid, new modules, vps-gateway 10/10, migration, CLI -> GREEN; pnpm test 778. Writer deviation: one git stash to observe a RED (checked: stash empty, HEAD intact; told not to repeat). The first independent verifier crashed twice (runtime errors) after a lead the parent confirmed by code: B1 the root daemon ran Pi's ModelRuntime on dot-owned auth.json/models.json (resolveConfigValue runs !cmd values; voice refresh calls it on every provider list) and wrote engine files dot could replace with links. Correction 162473c (class fix): ModelRuntime in cli.ts --auth-helper spawned as dot (no capabilities, scrubbed env, Node IPC, plain data back); every other daemon access to engine folders (profiles/settings, mcp.json + shim, .pi, .engram, history/rotation, uploads + prompt images, legacy mcp-auth.json, isolation) inside a synchronous engineAccess bracket (setgroups/setegid/seteuid to dot and back, exit 70 if it cannot restore); root files through writePrivateFile (O_EXCL temp, owner check, refused inside a bracket); per-launch hand-over removed; static guard against async fs (seteuid applies to every thread). RED: 2 non-root + 5 root tests in the container against a mktemp copy of 35b942b; GREEN: pnpm test 790 (13 root-only skipped on macOS), root tests 18/18 in the image, 3 mutations caught. Fresh full verifier (7ff8021..162473c): no blockers; S25.8 PASS, desktop S25.1/2/5/6 unchanged; as dot every daemon file EACCES, environ denied, links and hardlinks refused, planted !cmd ran only as uid 1001, PIN lockout persisted, PIN never in logs. Advisories fixed in 4d1035e: the user switch moved inside its try (root tests 18/18 in the image), entrypoint .engram hand-over only of root-owned files (--from=0:0), capability reasons in compose.yaml/deploy-vps.md corrected, stale token-folder comment. Advisories kept: connector !cmd values (user- or root-import-supplied only) resolve as root; root existsSync on engine-reported paths (boolean only); upgrade-only chmod in prepareVpsLayout follows a link left by an older image; docker exec shells inherit the key (documented); dot can ptrace/kill the helper (spoofed data or DoS, no root execution); a future library doing async fs as root is not caught by the guard. Not run: startup hand-over link probe (the safety policy blocks recursive chown commands), a daemon restart for lockout, a live connector pin_approve. S25 stage 2 complete (T23a-T23h). Next step: PR for T23f-T23h (about 4,700 changed lines; delivery decision pending), then the user's real-Mac round.
  • L128 (2026-10-10): T23g independent verifier: S25.1 (Linux channel) and S25.7 store PASS, Docker Linux check rerun green (379 tests, clippy, GNOME Keyring cases); two blockers. B1: the new docs said a keyring installed later needs no restart and that the assistant names it, but a keyring missing (or a prompt dismissed) at start leaves unchecked for the whole run (T23e fail-closed, by design), so the assistant gets INTEGRITY_UNCHECKED until a restart. B2: the daemon's mapping matched every the secret store is unavailable: answer, changing macOS's locked-keychain wording and dropping the proxy's log line. Correction 920455b: only the Linux answer (/^the secret store is unavailable: .*Secret Service/) is passed on; macOS keeps its wording (regression test); the proxy logs connector <id> cannot get its secrets: …; install.md, linux-testing.md, and design.md say to unlock/install the keyring and restart Gentle Dot. RED 2 → GREEN 43/43. Recheck: B1 and B2 closed (probes against a git archive copy). Advisories kept: getrlimit not on the POSIX async-signal-safe list (safe in glibc/musl); fcntl fallback never exercised; the 120 s bound is per prompt and the crate never dismisses a timed-out prompt; no-prompter maps to Denied; provider error text reaches the model unbounded (same-user only). T23h design accepted (writer's notice): GENTLE_DOT_VPS=1 only in the container; daemon root, engine dot, stdio connector servers dotmcp via setpriv; /app root-owned read-only; root 0600 token, connectors.json, secrets, PIN; AES-256-GCM per record with GENTLE_DOT_SECRETS_KEY scrubbed from env; PIN over the WebSocket, scrypt + salt, 5 failures → 15 min persisted lockout; desktop unchanged. Parent added: no-new-privileges, cap_drop: ALL, only justified cap_add, and a smoke check that the engine and dotmcp have no effective or ambient capabilities.
  • L127 (2026-10-10): T23f done (e9869d4): subagents on by default (GENTLE_DOT_SUBAGENTS=off opts out); children load the approval guard through a daemon-written, protected shim in <agentHome>/extensions/ that is inert outside assistant children and imports the guard from the daemon's install path; no propose_connector in children; inherited child markers scrubbed. RED 10 → GREEN 21/21; mutations (inert shim, &&→||) caught; opt-in offline real-engine probe (GENTLE_DOT_SUBAGENT_PROBE=1) passes in ~4 s. Risk tier high (item 2). Independent verifier: one blocker B1, an unwritable extensions folder made enforce() throw (daemon failed to start with subagents on or off; a planted .pi/mcp.json stayed; error logged every second). Correction b6e235d: any error placing the shim is reported once, subagents start off, the other checks still run; RED 2 tests → GREEN 9/9; recheck: B1 closed (scenarios A, B, C). Advisories kept: A1 shim written even with subagents off (harmless now), A2 planted <shim>.<pid>.tmp symlink (same as mcp.json at baseline), A3 children in other worktrees (baseline), A4 up-to-1 s unguarded window and a running engine keeps subagents on until its next start (documented), A5 probe uses a launcher fixture, not full gentle-shell. T23g done (837a42d): Linux crate 379 tests + clippy in the container; Secret Service put/get/delete against GNOME Keyring; no bus/no provider → Unavailable with the keyring reason; locked keyring → Denied fast; 120 s prompt bound; the app marks other fds close-on-exec in pre_exec (close_range, fcntl fallback); Node marks fd 3 CLOEXEC at startup so no daemon child holds the channel (strace; package smoke on debian and ubuntu: other holders none); same-user keyring limit documented (install.md, design.md, README, linux-testing step 17). Parent added the daemon side of the writer's finding: an app error starting the secret store is unavailable: becomes SecretsUnavailableError, so the assistant names the missing keyring instead of "Could not reach X right now" (RED 1 → GREEN). Checks at 837a42d: cargo test 385 passed 4 ignored, pnpm test 734 passed 6 skipped, typecheck, tracked lint clean. Not run: the GTK approval dialog on a real Linux desktop (S25.3 on Linux, tester step 17), KWallet, amd64. Next step: T23g verifier verdict, then T23h.
  • L126 (2026-10-10, user): "Sigue" (continue with the rest of S25 before the user's Mac checks). T23f (subagents behind the proxy) and T23g (Linux Secret Service, channel inheritance on Linux, documented limit) launched in parallel with disjoint surfaces; T23h (VPS) follows T23f because both touch the daemon's spawn and isolation.
  • L125 (2026-10-10, user): "Si a tus preguntas". PR #41 merged (ec44bdb; closed #8, #21, #22, #25, #28 through its commits). Before acting on the open questions every claim was checked: #16 duplicates the race fixed on main (models-gateway.test.ts:126) → closed with thanks; #14 superseded by the #13 fix on main (claimEngineHome, lifecycle tests) → closed with thanks, explaining why its timeout and npmrc changes were left out. The triage's claim about #27 was wrong: its new test fails on main without the fix and passes with it (full suite 720 passed), so instead of asking for a test it was merged (b12c2cb) with a note about a tiny duplicate window. #10 and #13 had been closed by their commits without our note; comments naming the closing tests were added. Remaining open: PRs #30 (changes requested), #33 (KSNI, draft), #34-#40 (Niri, asked to stack), #20 (--toggle-rose, owner decision); 8 issues (feature requests and proposals).
  • L124 (2026-10-10): T36 correction batch (6932a1b): the app emits dot://assistant-ready from on_ready after every successful ensure_running (Spawn, Attach, ReplaceOrphan) and the panel and Dot ask for their connection again, keeping a connection whose address and key did not change (B1, A3); the pid file is written through a create_new 0600 temp file renamed over it, read with O_NOFOLLOW|O_NONBLOCK only as a regular file of at most 64 bytes, and pids 0, 1, the app's own, and its parent's are never acted on (B2). Recheck: B1, B2, A3 FIXED, no regressions (cargo test 384, clippy, pnpm test 718; R1/R2/R4 probes rerun: symlinks not followed on read or write, new inode with no leftover temp, own/parent/0/1/FIFO/oversized refused, start time checked before every kill; a SIGTERM-ignoring orphan replaced in 10.08 s, then ready fired once). New advisory A9: a failed create_new removes whatever is at its pid-and-nanoseconds temp name (negligible). PR #41 merged next.
  • L123 (2026-10-10): Independent verifier on 30d532c..8705ab0 (S35): pnpm test 706 ×2, e2e 17, cargo test 377, clippy, typecheck, lint; T35, T37, T38 PASS; T36 FAIL with two blockers: B1 the panel calls connection_info once, so after the refusal's exit ("Restart assistant") it keeps showing the refusal until the app is reopened (the realistic #22 upgrade path, since an old orphan has no pid file); B2 the pid file follows symlinks on read and write, is rewritten in place (not atomic), and accepts the app's own pid (the app would SIGTERM itself). PASS within T36: pid plus start time checked before every signal, a mismatched process never signalled, SIGKILL only after 10 s, malformed or unreadable pid files ignored, pid 1 excluded; lifecycle (no respawn when the channel closes during restart, backoff, or start; no channel keeps running); frames over 1 MiB never sent; the engine home marker (#13) exact, wx, never on a custom home. Advisories: A1 a microsecond window between the start-time check and kill (Linux could use pidfd); A2 GENTLE_DOT_AGENT_HOME equal to <data>/agent is marked; A3 the worst-case orphan replacement can exceed connection_info's 20 s; A4 T38 releases note.object() inside the handler; A5 a SIGKILLed orphan leaves its engine and Engram; A7 "Open in browser" still opens a refused daemon (baseline). One correction batch launched: B1, B2, A3. #41 stays unmerged until the recheck passes.
  • L122 (2026-10-10, user): "Mergea todo y explica lo del 30 y niei"; then "30 pedir cambios y niri pedir apilar", "Y comprobar también cada uno", "Go". Merged community PRs #18 (c99abd9), #23 (a3e9152), #26 (73e8958) after testing them merged together on main (680 passed); main merged into feat/gentle-dot without conflicts (715 passed); #41 waits for the S35 verifier because it now carries T35-T38. Not merged: #27 (its test passes without the fix, so it does not prove #7), #16 and #14 (to close with thanks; asked). Before posting, every claim in the drafts was checked against the PR diffs and metadata: two were corrected (the #30 extension does not read auth.json itself, and the impact of an unprotected extension was overstated; the Niri "flicker" is the author's documented reopen-at-anchor-then-move). Posted: a changes-requested review on #30 (protect the extension like the guard's file, drop the ODD edits that collide with our T35/L118, list models from the gateway) and a comment on tracker #34 (stack the chain, note real-Niri verification, track the reopen behavior).
  • L121 (2026-10-10, user): "2" (community triage first), then "lo de gentle shell no lo toques no te preocupes, solo mira lo de gentle dot" (the gentle-shell reports the user forwarded were dropped: triage cancelled, its scratch removed, the gentle-pi checkout untouched). gentle-dot triage (read-only explorer over a dump of all 19 open issues and 17 PRs, against HEAD 3cc71f8): buckets A 2 (#11 superseded by c6d7c6f, #15 already fixed), B 1 (#28 duplicate of #8), C 8 (#6, #8, #10, #13, #21, #22, #24, #25), D 7 (#4, #9, #17, #19, #29, #31, #32), E 1 (#7: not reproducible from code; ask the reporter). Clusters: C1 immutable bundle (#8/#28: createJiti default fsCache, auth.ts:68-69), C2 daemon lifetime (#22: daemon.ts:269-276 keeps running after its channel closes; daemon.rs attaches without a channel), C3 engine home ownership (#13: gentle-shell treats a non-empty home without .gentle-shell-home as foreign, and the daemon writes auth.json/mcp.json there first), C4 uploads/composer (#6 fixed by PR #23; #21 useDot.send returns void and the composer always clears), C5 activity titles (#24 fixed by PR #26), C6 voice (#10 configuration-change handler fires while running, macos.rs:126-129, probably since T34; #25 useVoice checks status once). Urgent: #13 blocks clean installs, #22 strands users with locked connectors and no exit. PR notes: #23, #26 merge as is; #27 merge after changes; #14 superseded by the C3 fix; #16 close (already on main); #18 merge; #30 (a new provider by rodrigovzq) and the Niri chain #35-#40, KSNI #33, #20 --toggle-rose are the owner's decisions. Recorded as S35 with T35 (inline) and T36-T38 (parallel writers with disjoint surfaces).
  • L120 (2026-10-10): Independent verifier on 02784d6..9925c39: pnpm test 688 ×2, e2e 17, typecheck, lint, 23 own probes; the MAC reproduced by an independent Python implementation; 12 kinds of tampering set aside once with content preserved; ordering, A2 closure, regression PASS; no blockers. Fixed inline (RED reproduced both → GREEN, pnpm test 690 ×2): A1 with the app present but the integrity key unreadable or unstorable, the file was loaded unverified and secrets stayed reachable; now the check fails closed (INTEGRITY_UNCHECKED: file kept as is, no secret read, every window told) until a start that can check it; A3 a signed file deleted while closed started empty in silence; now MISSING_NOTICE. Documented: the HMAC has no counter, so putting back an older signed version is accepted (A2; sending still needs native approval), and deleting the key plus removing mac reads as a first start. Kept: A4 a legacy mcp-auth.json planted while a migration is retrying is still imported.
  • L119 (2026-10-10): T23e done (9925c39) after the writer stalled 4 minutes on a shell command and was resumed in its own session with timeouts on every command. connector-integrity.ts: HMAC-SHA256 over gentle-dot/connectors.json/v1\n plus the canonical record without mac (sorted keys, no spaces), key = 32 random bytes in the store as integrity/connectors; verification after listen and before migration, enforce/watch, and the engine start, with app commands waiting for it; a signed file is kept byte for byte when no key is available so it is never rewritten unsigned. A2 closed: an unreadable legacy sign-in file is reported and kept, and the migration retries. Documented limit: a same-user process that deletes the Keychain key and removes mac makes the next start treat the file as a first start. Independent verifier launched.
  • L118 (2026-10-09, user): "si" to merging PR #12 and continuing with T23e. PR #12 merged (3954682). T23e design: an HMAC-SHA256 key kept in the Keychain through the app channel (secret_get/put, id integrity/connectors); the daemon signs connectors.json on every write (a mac field over a canonical serialization of the rest); at start with the channel, a missing or wrong mac on a file that was signed before means it changed while Gentle Dot was closed: the file is moved aside (connectors.rejected-<timestamp>.json, private), the daemon starts with no connectors, and the user is told once; the first start after T23e signs the current file (migration); without the channel the file cannot be verified and is loaded as today (connectors needing secrets are already locked there).
  • L117 (2026-10-09): Independent verifier on eec6776..c6d7c6f: pnpm test 669 ×3 (no flake), e2e 17, cargo test 369, clippy, typecheck, lint; no secret on disk, in WebSocket messages, logs, or errors after connect, refresh, mode change, restart, disconnect, and remove; secret_* unreachable from the WebSocket; every proxy method fails closed without the app; OAuth PKCE, wrong state, replayed code, pasted address for another flow, cancel, loopback-only callback, single refresh on 401, no loop without a refresh token; migration survives a kill between put and file rewrite; no blockers. Advisory A1 (serious, would have waited for T23e): a {secretRef} planted in connectors.json while the daemon was down, in another connector's slot, made the daemon fetch that item and hand it to an agent-chosen stdio command (reproduced with Notion's sign-in). Fixed inline: a reference is honored only in its own slot (ids follow from the connector and field: connector/<id>/{value|env|header}/<k>, connector/<id>/signin), so refsOf, open, isSignedIn, and signInState ignore any other; signInState also stopped a planted sign-in reference from sending Notion's cached token to another server. RED (both attacks reproduced on the unfixed code) → GREEN, pnpm test 671 ×2. Other advisories kept: A2 an unparsable legacy mcp-auth.json is marked migrated and its secrets stay on disk, and a 0-byte one is deleted; A3 the browser e2e for connector flows were replaced (they failed at the baseline since T23c); A4 an imported server's callbackUrl decides the callback host (non-loopback would listen outside; same as baseline); A5 isSignedIn is true for an uncached reference, so without the app the list says connected while the proxy fails closed; A6 serve_secret copies the secret into a serde_json::Value not zeroed; A7 every imported env and header value goes to the Keychain, non-secret ones included.
  • L116 (2026-10-09): T23d done (8b29bab): secret-source.ts (AppSecretSource over the channel, in-memory cache cleared when the channel closes), connector-oauth.ts (daemon-run pi-mcp OAuth with PKCE and a loopback callback, replacing mcp login), connectors.json holds {secretRef} only, one-time migration with a version: 2 marker (N1 planted files reported and ignored, N2 empty or unreadable files do not block), app-side secret_* handlers with id validation and no echo. Limits: security-framework 3.7 cannot read an item's access list, so the daemon only reads ids it recorded; secrets inside a server's URL or args stay in connectors.json. Parent fixes: two tests raced engine restarts under load (the proxy key rotates on each restart; and the models list pushed at engine start since L101 could be taken for the reply) — both now wait for the right event (8 consecutive full runs green); e2e/serve.ts dropped the removed GENTLE_DOT_MCP_CLI; two e2e tests still drove connector changes in the browser, which T23c made app-only, so they now check the browser shows the list with every change disabled and the reason (c6d7c6f, e2e 17/17). Independent verifier launched.
  • L115 (2026-10-09, user): "si" to merging PR #5 and continuing with T23d. PR #5 merged (ed14f2d). T23d design: the daemon gets connector secrets from the app over the private channel (secret_get/put/delete/list), the app serves them from Keychain items it creates, and the daemon keeps them in memory only; OAuth sign-in and refresh run in the daemon with pi-mcp's OAuth and a channel-backed state store, replacing the mcp login CLI and the sign-in home files; existing secrets in the sign-in home, connectors.json, and the legacy mcp-auth.json migrate once into the Keychain and are removed from the files; without an app channel (attached daemon, pnpm dev) connectors that need secrets fail closed with a reason (the VPS store comes in T23h). Writer launched.
  • L114 (2026-10-09): Independent verifier on 03c5994..8ac11aa: pnpm test 658, cargo test 366, clippy, typecheck, lint; B1 recheck PASS (only the Allow response allows; refuse, abort, stop, cancel, unknown, main-thread call, or failure refuses; the plugin path maps Escape, close, and failure to the refusing label), A2 PASS, channel authentication PASS (10 privileged commands over the WebSocket, including a spoofed clientId, return app_required with connectors.json, mcp.json, the proxy's tools/list, and the engine pid unchanged), channel integrity PASS (no listening Unix socket; fd 3 reaches no child: engine, mcp login, bash, detached, shell, exec, spawnSync, pi-mcp StdioTransport; frames over 1 MiB close the channel and the WebSocket keeps working), approvals PASS except blocker B2: the import re-read the latest scan after the native answer, so a scan sent while the dialog was open swapped the imported server (reproduced: the dialog showed the filesystem server, connectors.json got /bin/sh -c "echo pwned"). Fixed inline (4d1eafb): importPlan freezes a copy of the chosen servers at confirmation and the import uses it; RED reproduced the swap → GREEN, pnpm test 659. Advisories: N1 a legacy mcp-auth.json planted after migration adds sign-ins for connectors the daemon lacks (T23d/T23e close it); N2 a 0-byte current mcp-auth.json blocks migration; N3 an aborted approval leaves its native dialog open up to 120 s, queuing later ones; N4 anyone launching the daemon with their own socketpair on fd 3 becomes "the app" (by design; T23d/T23e); N5 a malformed computer_register over the WebSocket answers bad_message instead of app_required (still refused). User: "actualiza TODO" (todo list rebuilt). App rebuilt, signed, restarted: the daemon holds fd 3 as a Unix socket connected to the app and the computer helper registered over the channel.
  • L113 (2026-10-09): T23c done (8ac11aa): socketpair end passed as fd 3 (channel_pair with CLOEXEC, posix_spawn adddup2 / Linux pre_exec), AppChannel (line-delimited JSON), childInherits startup probe (on macOS Node 24's libuv spawns children with POSIX_SPAWN_CLOEXEC_DEFAULT, so fd 3 does not reach them); privileged actions over the WebSocket answer app_required ("Open the Gentle Dot app to change connectors."); the panel sends them with the Tauri command connector_command; proxy approvals and widening changes are confirmed in native dialogs from a worker thread; WebSocket ask cards are informational; the computer helper registers over the channel and unregisters when it closes (closes the L61 risk). B1 fixed: NSAlert directly on macOS, allow only on NSAlertSecondButtonReturn, everything else refuses; off macOS OkCancelCustom(allow, refuse) so cancel or failure maps to the refusing label. A2 fixed: legacy sign-ins migrate whenever the old file exists, keeping the daemon's own. Open: engine-own asks for servers the proxy does not cover stay answerable over the WebSocket; Linux fd inheritance unverified; NSAlert behavior needs a real Mac. Independent verifier (T23c + B1/A2 recheck) launched. The running app is an older build: rebuild only after the verifier, since a new daemon with an old app would refuse connector changes.
  • L112 (2026-10-09): Independent verifier on 7cde26d..6016b43 (detached worktree): pnpm test 636, cargo test 348, clippy, typecheck, lint; T23a PASS on auth, policy (case, unicode lookalike, NUL, ZWSP, prefixed names, late-listed tools, same name across connectors), approvals (one ask, full arguments, late approval never forwarded), secrets (no token, env, command, or upstream URL in the engine's mcp.json; failing !cmd leaks nothing), migration, regression. T23b FAIL, blocker B1: alert.rs passes OkCancelCustom(refuse, allow) and tauri-plugin-dialog 2.8.1 maps rfd's Cancel (Escape-like responses, display or channel failure on macOS's CFUserNotification path; Escape or close on GTK) to Custom(<second label>), the allow label, so cancelling or failing allowed computer grants, risky confirmations, yolo, and approvals (the baseline refused). The user's running build predates 1c345b5. Fix assigned to the T23c writer: show NSAlert directly on macOS (Refuse first with Return and Escape, allow only on the Allow response), failures refuse elsewhere. Advisories: A1 a daemon ask survives the window closing and the next window can answer it (T23c's native path replaces it); A2 the migration skips when the sign-in home exists, so a partial failure strands legacy tokens (fix assigned to T23c); A3 modes are not re-imposed after a chmod; A4 cat ../connector-signin/* passes the best-effort bash check (same class as baseline); A5 storedSignIn reads the token from disk each request; A6 the WS preview is rewritten by presentText, and the native dialog clips values at 600 characters or 12 lines with a marker; A7 concurrent requests with the same JSON-RPC id share a slot.
  • L111 (2026-10-09): T23a done (6016b43): mcp-proxy.ts (Bearer per engine launch, Origin check, JSON-RPC over POST, mode filtering of tools/resources/prompts, hidden-tool refusal, one approval with the full preview before sending, upstream via pi-mcp HTTP or daemon-run stdio), engine mcp.json with proxy entries only, sign-in in <data>/connector-signin (0700/0600, protected, put back, one-time verified migration), guard policy.proxied. The user's install has no legacy mcp-auth.json and an empty connectors.json, so the migration is a no-op there. Design change for S25.1 by the parent: instead of a key, the app passes one end of a Unix socketpair as fd 3 when it spawns the daemon; the connection itself authenticates, with no key to steal and no listening socket to impersonate. Independent verifier on 7cde26d..6016b43 and the T23c writer launched in parallel.
  • L110 (2026-10-09): T23b done (1c345b5): secure_store (Keychain via security-framework, items created by the signed app; Secret Service on Linux, not compiled here; in-memory fake; redacted Secret), approvals (one native dialog at a time, full sanitized preview, Decline is the default button, 120 s timeout declines, abortModal for a stale dialog), and advisory A2 from L75 applied: computer-control grant, risky-action, and yolo dialogs default to their refusing button (Return no longer approves). A real Keychain put/get/delete under a throwaway service passed with no prompt and left no item. Open advisories for T23c/T23d: a same-user process can delete items (DoS, fail closed) or plant an item before the first put; get does not verify the item's ACL; item ids are readable. T23a (proxy) still running: sign-in moves to a daemon-owned <data>/connector-signin/ home, protected like today's credential files.
  • L109 (2026-10-09): S25 map (explorer without web; platform facts inferred): any WebSocket client with <data>/token can answer approvals and also switch a connector to read and send, connect, or remove it (bridge.ts:253-273); pi-mcp 1.0.4 is client-only (stdio, streamable HTTP, OAuth with an injectable state store), so the daemon writes its own MCP server side, as the Rust computer helper already does (plain JSON over POST, accepted by the engine); a Keychain item created through the security CLI does not protect from the agent, one created by the signed app does; Linux Secret Service has no per-caller ACL. User choices: "Mac primero, Linux documentado (recomendado)" and "Motor con otro usuario + PIN (recomendado)". S25 refined into S25.1–S25.8 and T23a–T23h; T23a and T23b launched in parallel. Also: T34 done (7a8773c, native buffers for Apple Speech) and Voice language in Settings (b2b2c79) after the writer found the Mac is set to en-US, so Apple Speech always listened in English; app rebuilt, signed, restarted.
  • L108 (2026-10-09, user): "merge" — PR #3 merged (231f413). Then "da a todos" for: Apple Speech with native buffers (S30.6/T34, writer launched), S25 (T23, read-only map launched in parallel), and the real-hardware checks (need a second Mac, an x86_64 PC, and GNOME/Hyprland desktops: handed to the user and testers as a checklist).
  • L107 (2026-10-09): User: "merge el pr" — PR #2 merged (0c3a7a3). T33 done (2d1f5e4): shortcut_get/shortcut_set (panel only), pure shortcut.rs rules, apply = register new → save → release old → tray label, rollback when saving fails, config.json read-modify-write 0600, Settings screen with a capture-phase recorder (event.code, so ⌥ does not change the key). Parent changes: the writer's proposal to refuse Shift-only combinations adopted (S33.2 updated: Shift+K would capture capital K everywhere); a pre-existing crash in VoiceModel on a null status (6 unhandled rejections making pnpm test exit 1) fixed (bd819a8). Known limits: macOS may accept a system-reserved combination (e.g. ⌘Space) that never fires; a combination already global cannot be recorded in the webview. Rebuilt, signed, restarted.
  • L106 (2026-10-09, user): "porque tiene un space el banner?"; then "sacalo y creo que deberiamos de cambiar por otra cosa, que uno pueda configurar su atajo no siempre space"; placement "Nueva pantalla de ajustes (recomendado)". Banner regenerated without the key chip and README names ⌥ Space as the default (2556674, in PR #2). Recorded as S33/T33.
  • L105 (2026-10-09, user): "merge". PR #1 merged with a merge commit (ae9e9e9) to keep the work-unit commits; no CI checks exist on the repo yet. The README and banner are live on main.
  • Next step: the user's check of review-before-send, live text, the model chip at launch, and NaN with computer control; then Apple Speech fed with native buffers; S25 before a stable release; T30 verifier verdict; the T29c UI hook (download button, "Transcribing…"); drag-drop handler; pin sherpa-onnx libs; rebuild for a real-Mac round; the T29c UI hook; pin sherpa-onnx libs in packaging; drag-drop handler change; T29c, T31, T30 verdict; then the T29c UI hook, the drag-drop change, and a rebuild for a real-Mac round; grant both permissions, take a screenshot (the Dot must be absent), click and type in a harmless app, a risky "Send" confirmation, ⌥⇧Esc panic, a blocked app (System Settings), and screencapture -x /tmp/t.png from the agent's bash must fail (R1); the community tester runs docs/linux-testing.md when the user shares it; T23/S25 before any public release.