Repository navigation
feat: each bot's own browser, commands, memory, conversations, and restarts that keep the work going - #22
Open
mrinaldi2 wants to merge 48 commits into
Open
feat: each bot's own browser, commands, memory, conversations, and restarts that keep the work going#22mrinaldi2 wants to merge 48 commits into
mrinaldi2 wants to merge 48 commits into
Conversation
Bot creation and the supervision tick could both enter start_bot for a new bot before either registered its session. With Codex, the second runtime resumed the thread the first had just created and failed with "thread already has an active writer"; the failure path then cleared the healthy first session, crashing the bot until the next retry. start_bot now claims the bot under the supervisor lock before launching anything, and releases the claim when it returns, so a concurrent caller backs off instead of starting a second runtime. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Pair two of the owner's daemons over Tailscale and stand a bot from one in as a linked bot in a project on the other. A linked bot is an ordinary bot row with peer_id set, so addressing, task budgets and history work unchanged; the delivery worker forwards deliveries for it over the /peer link instead of writing to a local session. The receiving daemon stores each forwarded message as if the sender were local: a task is mirrored with the sender's deadline and hop count, a result or cancel closes the mirrored task, and the sender gets a linked bot of its own on first contact so replies route back. Files listed in a result travel with it (from the bot's own directory or the project artifacts only, size-capped) and land in the recipient project's artifacts directory. Pairing is a one-time invite holding a link token bound to the first daemon that presents it. Peers can only list bots, link one, and deliver to bots linked to them; revoking drops the link. `gravityd peer ...` drives pairing and linking until the desktop UI lands. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Refs ahilles107/gravity#13. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Phase 1 of the chat pane (ahilles107/gravity#13). The daemon now reads a bot's runtime transcript and serves it as turns: what woke the bot (the owner from the composer or the terminal, another bot, a routine, a ruling, a background job), its text, its tool steps with diff counts and images, and Gravity's own bus calls as sent messages, results and decisions. A per-bot reader folds the transcript in incrementally and a watcher pushes chat_turns for loaded chats as they grow. Step detail, images and files are fetched on demand; read_file only reaches the bot's own directory and its project's artifacts. New read-grant requests list_chat, get_chat_step, get_chat_image, list_artifacts and read_file, behind a `chat` capability. Also fixes two bugs the research turned up: - transcript_dir kept `_`, while Claude Code files every non-alphanumeric character as `-`, so bots under such paths had no activity preview. - Routine runs never completed from the transcript on Unix: the Stop hook did not forward Claude Code's payload (so no transcript_path), and the run-id scan did not expect Claude Code's peer-message wrapper. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Phase 2 of the chat pane (ahilles107/gravity#13). Opening a bot now shows its conversation as turns read from the daemon's chat model: who woke it, its answer as markdown with highlighted code, tool steps folded into groups that expand to their command, diff and output, step images with a viewer, sent messages, results with their files, and decisions raised. The composer sends through send_user_message and shows a message's delivery state until the bot reads it. The chat follows chat_turns pushes and pages back through older turns. The Terminal and Routines tabs stay (⌘1-3, ⌘L focuses the composer), the side panel gains a Files tab listing the project's artifacts with previews, and a linked bot from a peer shows its bus conversation. A daemon without the chat capability keeps the terminal-first view. Bot markdown keeps the Control Center's policy: links and remote images are inert; images render only from bytes the daemon served. Also fixes list_messages paging, which sent before_id to a daemon reading before_num. Visual baselines for the new stories still need to come from CI. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Phase 3 of the chat pane (ahilles107/gravity#13). Bots get a blocking PermissionRequest hook (not PreToolUse, which fires for every call the bot is already allowed to make). The hook posts the prompt to /hook/permission and waits; the app shows a card above the bot's tabs with Allow once, Allow for this session and Deny (with a reason the bot sees), and the answer goes back as Claude Code's decision. Nothing defaults to yes: - An unanswered prompt is denied when its window closes (permission_timeout_seconds, default 600, below the hook's timeout). - "Allow for this session" sends Claude Code's suggested rules with the destination forced to session; mode changes and persistent rules are dropped, so no settings file is ever written. - The daemon only holds a prompt while an app that shows cards (hello features: permission_cards) is connected with the control grant. Otherwise, or if the hook fails, Claude Code's own terminal prompt appears as before. - Answering is `control`, the same grant as typing into the terminal. New list_permissions / answer_permission requests and permission_request / permission_resolved pushes. Also splits chat/builder.rs (505 lines, over the limit since phase 1) and moves the Unix hook settings out of paths.rs into paths/unix_hooks.rs. The live Claude Code check the issue requires is still to be run. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Next to Info and Files, a bot's Tasks tab shows its work in four sections: Now (open tasks it was given), Waiting on others (open tasks it delegated, with the peer machine for linked bots), Upcoming (its enabled routines' next runs) and Done (finished, cancelled or expired tasks, opening to the result that closed them). It refetches after bus traffic and routine runs. A new read-grant `list_tasks` request serves a bot's tasks from both ends, newest first, with the request, the result and when it closed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Phase 4 of the chat pane (ahilles107/gravity#13). The Codex worker now records every command, file change (with its diff as Claude Code-style structured patch hunks) and MCP call in codex-observations.jsonl as tool_use / tool_result pairs, and marks each turn's end with a turn_duration record, so the chat reads a Codex bot's turns, steps, diffs and bus calls the same way as a Claude bot's. Command, file-change and permission approvals become permission cards: the worker raises a structured SessionEvent::Permission, the supervisor relays it on the internal bus, the approval broker shows the card, and the answer goes back through RuntimeSession::answer_permission as the App Server's accept / acceptForSession / decline (or a turn- or session-scoped permission grant). Whichever answers first, the card or the native CLI, decides; serverRequest/resolved withdraws the card. Questions and elicitations stay in the native CLI. The broker is split into a shared decide step with Claude hook and runtime front ends. The supervisor's session event loop moves to its own file. Also fixes the scheduler's routine-completion listener, which ended for good on the first internal event that was not BotDone, or on a lagged burst. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Phase 5 of the chat pane (ahilles107/gravity#13). A linked bot's chat used to show only its bus conversation. Now list_chat, get_chat_step, get_chat_image and read_file for a linked bot are forwarded to the daemon it runs on, which serves them for bots linked to that peer only (read_file stays confined to the bot's own directory and its project's artifacts there). Turns come back with the local linked bot's id. While a peer has a chat open, the daemon running the bot streams its chat_turns as event frames (no req_id, no answer) and the other daemon pushes them to its clients under the linked bot's id. Daemons advertise `peer_chat`; the app shows the full chat for linked bots on such a daemon, with a note saying which machine the bot runs on, and keeps the bus-conversation view otherwise. Permission prompts of a linked bot are still answered on its own machine. The two-daemon test setup moves to tests/common/peers.rs. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Phase 6 of the chat pane (ahilles107/gravity#13), the parts the issue left out of v1: - Decisions in the thread: a decision a bot raised shows its options in the chat. Picking one, or writing an answer, publishes it in one step (publish_decisions with an inline ruling), as "Answer and publish" does in the Control Center; it needs the approve grant. The card links to the Control Center for drafts, holds and the discussion. - Attachments: attach or paste files in the composer. They upload in chunks through a new control-grant write_artifact request into the project's artifacts/uploads/ (16 MB cap, name confined to its last component), and the message lists their paths for the bot. - Slash commands: a message starting with / is typed into the bot's terminal, as if typed there, with a hint under the composer. - Search: ⌘F filters the loaded turns by their text, steps and trigger. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Search now works like find in page: every match in the loaded turns is highlighted in place (CSS custom highlights), the current one apart, and ▲/▼ or Enter / Shift+Enter step between them, scrolling each into view. Folded step groups open while searching so their text is found. Turns are no longer filtered out. - Tasks: requests, results and routine prompts clamped to two lines get a "Show more" / "Show less" when the clamp hides something or a result is waiting, instead of only an ellipsis. - Image thumbnails keep their image's aspect ratio (fixed height, width from the image) instead of a square crop. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Tasks list trims each request to 400 characters and each result to 600, so "Show more" could only reveal a cut-off preview. The list now says when it trimmed (request_truncated, result_truncated), a new read-grant get_task request returns one task with its full request and result, and opening a row loads it when the preview was cut. Opened requests and results render as the markdown bots write. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A mic button in the composer turns speech into text with the operating system's own recognizer, so audio never goes to Gravity or a third party. On macOS, AVAudioEngine feeds the microphone to Apple's Speech framework (objc2-speech / objc2-avf-audio), set to recognize on the device whenever the Mac supports it. Partial transcripts stream into the draft after what was already typed, and the final one stays; pressing the button again stops. The first use asks for Speech Recognition and Microphone access: the app now carries NSSpeechRecognitionUsageDescription and NSMicrophoneUsageDescription (src-tauri/Info.plist), and signed releases get the hardened-runtime audio-input entitlement (Entitlements.plist). The mic is hidden where dictation is unavailable: a plain browser, and Windows until its recognizer (Windows.Media.SpeechRecognition) lands. The composer's attachments move into useAttachments and its buttons into ComposerParts. Notices gain the seven objc2 framework crates, which ship no license file and use the reviewed objc2 LICENSE override like their siblings. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The full-size image viewer reused the command palette's overlay panel, which is 580px wide, at most 60vh tall and clips its content, so wide screenshots were cropped into a box. Images now open in a full-window lightbox: fitted to the window at their own aspect ratio, or at actual size (click the image) with the stage scrolling. ←/→ and the side buttons page through a step group's images; Esc or the backdrop closes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The composer's mic was macOS-only. On Windows, SAPI's in-process recognizer with its dictation grammar now feeds the same partial and final events, so audio stays on the PC. WinRT dictation was not used because it requires Windows' online speech service. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ation fec6522 changed apps/desktop/src-tauri/Cargo.toml and Cargo.lock without regenerating the notice inventory, so `scripts/notices.py` refused them. It added no crates: the regenerated records and bundles are identical, so only the two input hashes change. The records are kept as committed rather than regenerated here, since a macOS run swaps the marketing build tools' Windows platform packages for macOS ones. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Eleven stories added with the chat pane had no baseline, so the visual suite failed on each: chat composer, code blocks, lightbox, permission cards, tasks panel and turn view. Generated in the pinned Playwright container with --update-snapshots=missing, so the existing 72 baselines are untouched and still match byte for byte. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
After revoke_peer, pairing the same two daemons again failed with "a peer named '<name>' already exists": the revoked row kept its name under the UNIQUE constraint. It also kept the other daemon's id under its own UNIQUE constraint, so even a new name would have left the link unable to bind. Revoking now tombstones the name the way archiving a bot does (`name#<id8>`) and releases the daemon id. Views show the owner's name, not the tombstone. Migration 15 does the same for peers revoked before it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Linking a project here with one on a peer makes the two one team: every bot that runs on either side is exposed to the peer and stands in on the other side, and stays mirrored while linked. This is the daemon half of the protocol GravitiOS is built against (gravitiOS docs/linked-projects.md). - hello_ok carries `daemon_id`; capabilities add `linked_projects`. - A `project_link` table (migration 16) records each link on both sides; one link per (project, peer), and per (peer, remote project). - Projects gain `links` (peer_id, peer_name, online, remote_project_id, remote_project_name, linked_at), and `project_updated` is pushed when they change. - Control plane: `list_peer_projects` (read), `link_project` and `unlink_project` (control), and `create_bot` with `peer_id`, which creates the bot on the peer and replies with its stand-in here. Errors keep their code across the link: `conflict`, `not_linked`, `not_found`, `unavailable`. - Peer frames: `list_projects`, `link_project`, `unlink_project`, `create_bot`, `update_bot`, `delete_bot`, and the `project_roster` and `project_links` events. A peer acts only on projects linked through it, except `list_projects` and `link_project`; accepted links are logged. - Mirroring: a watcher on the daemon's own pushes sends a linked project's whole roster on any bot change; the receiver creates, updates and archives stand-ins to match. Events on a link are applied in order, and each link-up resends every roster and drops links the other side no longer holds. - A link whose rosters share a bot name is refused, naming the bots. Unlinking from either side, revoking the peer or archiving the project archives the stand-ins on both sides (cancelling their open tasks) and stops the peer reaching the project's bots. - Stand-ins no longer count towards `max_bots_per_project`. - MCP `create_bot` takes `machine`, allowed only when the project is linked through that peer; `update_bot` and `delete_bot` on a bot created there are forwarded, and the peer holds them to the creator rule. The system prompt names the machines a project is linked with. - The owner's `update_bot`, `set_bot_runtime` and `delete_bot` on a stand-in in a linked project are forwarded to its machine, since changing only the stand-in would be undone by the next roster. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
protocol.md lists the new requests, `create_bot`'s `peer_id`, the project's `links`, `daemon_id` in hello_ok and the `linked_projects` capability. peer-bots.md gains a "Linked projects" section (what a link means, how linking, mirroring and creating on the other machine work, and the peer frames), notes that revoking frees the peer's name, and adds the schema. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ions
Bots drove the owner's own Chrome through the Claude in Chrome extension:
the owner's global `claudeInChromeDefaultEnabled` reached every bot
session, so tabs opened in their Chrome with nothing saying which bot
asked or why.
Each bot now gets a browser of its own:
- Its session runs the Playwright MCP server (pinned, with the vision
tools for coordinate clicks, plus tabs) against a headless Chrome whose
profile lives in the bot's directory, so logins persist and stay the
bot's. Claude and Codex bots both get it. The daemon finds Node itself,
since launchd's PATH misses version managers, and picks Chrome, else
Edge, else Playwright's Chromium. `[browser]` in gravityd.toml can turn
it off or pin Node, the channel, headless and the package.
- Bots start with `--no-chrome`. A per-bot `user_chrome` switch
(`set_bot_user_chrome`, migration 17) gives one back the owner's Chrome
for tasks that need their logged-in sessions; it restarts the session.
The system prompt tells the bot which browsers it has.
- `watch_browser` streams a bot's browser to the app: `browser_tabs` when
its tabs change and `browser_frame` screencast frames of the tab the bot
used last, or one the owner picked. Chrome is found through the
DevToolsActivePort file in the profile; no new dependencies.
- `list_browser_activity` lists the bot's browser steps, newest first, with
their turn and what started it, and says whose browser each step used.
Browser steps are titled by what they did ("Opened example.com",
"Clicked Sign in").
Conversations between bots: `list_agent_conversations` returns the pairs
of bots in a project that have talked, most recent first, and
`list_agent_conversation` one pair's messages oldest first, paged, with
each message's sender, recipient, kind and task state. Both read the two
bots' DMs together; no schema change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Clients show the Browser tab, the Chrome switch and the Conversations section only on daemons that advertise `bot_browser` and `agent_conversations`. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…' conversations - Browser tab: the bot's own browser, live. Its tabs (following the one the bot used last, or one you pick), the page as the daemon streams it, and every browser action under the request that started its turn, with actions in your own Chrome marked. It streams only while shown. - "Can use your Chrome" in the bot's Info panel: off by default; allowing it restarts the bot with access to your Chrome. - Memory tab: the bot's FACTS.md as markdown, reread after its turns. - Conversations: a row at the top of each project in the sidebar opens the pairs of bots that have talked, most recent first, and each pair's messages as a chat. Every bot has its own bubble and avatar, on a side that stays put (the one higher in the sidebar on the left), with the message kind and a task's state. It refreshes as they talk and pages back on request. Each appears only on daemons advertising `bot_browser` or `agent_conversations`. Visual baselines for the new stories were made in the pinned Playwright container. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
bot-browser.md explains why bots no longer use the owner's Chrome, what each bot's browser can do, how the app watches it, and the [browser] settings. protocol.md lists the new requests, pushes and capabilities. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Chrome keeps a DevTools HTTP connection open whatever the request asks, so reading the tab list to the end of the stream timed out and the Browser tab saw no tabs. Found by driving a real bot browser through a throwaway daemon; the test's fake DevTools now keeps its connections open the same way, and fails without this fix. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The app now watches a bot's browser from the moment the bot is selected, whatever tab is open, so the Browser tab is live when opened; its label shows a red dot while the bot's browser is open. Since the phone will watch too, the daemon no longer gives each watching connection its own screencast. Every viewer of a bot reads one shared stream: one tab-list poll per browser, one screencast per tab. Streams hold only the latest value, so a viewer that joins gets the current tabs and screen at once and a slow one (a phone on a weak link) skips to the newest frame. A stream stops when its last viewer leaves. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Bots come back from a daemon restart resuming their conversation, but a turn that was running when the session stopped did not resume: the bot sat idle and its open tasks waited forever. At startup, before sessions start again, the daemon now reads each bot's transcript for a turn that never ended, and its open tasks. Each bot with either gets one note: what started the interrupted turn, its open tasks with their task ids and requests, and to pick up where it left off (checking its files first, since running commands stopped), or keep waiting if it was waiting on someone. A note still waiting to be delivered is not sent again, so quick successive restarts tell a bot once. `resume_after_restart = false` in gravityd.toml turns it off. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…r machine A linked bot's terminal was only reachable on its own machine. Now the daemon a client is connected to mirrors it: the bot's machine feeds its terminal output over the peer link (`term_attach`, then `term_frames` events) into the stand-in's own terminal buffer, so the desktop and a phone attach to the stand-in exactly as to a local bot, with replay and resume, and the link carries one feed however many of them watch. Input and resizes go back as `term_input` and `term_resize` events. The feed runs while anyone watches, picks up after the last frame it sent when the link comes back, and stops (`term_detach`) when the last viewer leaves. A peer only feeds or takes input for bots linked to it; typing still needs the `control` grant on the client's own daemon. The desktop shows a linked bot's Terminal tab on daemons advertising `peer_terminal`. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A linked bot's Browser tab now works from the other machine, and from a phone connected to it. The bot's machine watches its browser for the peer through the same shared stream its own clients read (`browser_watch`) and forwards the pushes as `browser_feed` events: tab lists at once, frames newest only and at most a few a second, so a slow link never queues them. The other daemon keeps one feed per stand-in and tab choice, shared by all its viewers, and renames the bot to the stand-in. When the link drops, viewers are told the machine is offline, and the feeds resume when it is back. The browser log is read from the bot's machine. Daemons advertise `peer_browser`; the desktop shows a linked bot's Browser tab on them. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A new side tab beside Info, Tasks and Files shows every command a bot runs: what is running now and what finished, foreground and background, each with its description, command line, how long it took, its exit code and its output. A background command still running shows the live end of its output file. The daemon reads commands from the bot's transcript, alongside the chat: each `Bash` call from start to result, and a background one until the runtime's task notification or the bot's `TaskStop`. A command left running when its session ended reads as stopped. Codex sessions record commands the same way. Checked against a real bot transcript: 771 commands, every one resolved. `list_bot_commands` serves it (read grant, `bot_commands` capability); for a linked bot it is read from its machine, so the phone sees the same list. The panel refetches as the bot's turns change and polls while a background command runs. The refetch-on-turns effect the Memory, Commands and browser activity panels share is now one hook. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Picking up unfinished work only happened when the daemon started. A bot restarted on its own (a crash restart, a runtime or Chrome change) came back resuming its conversation but was not told about the turn it was cut off in, which just stopped there. The check now runs before every session start: the supervisor calls a hook while the transcript still ends where the last session stopped, and a bot cut off mid-turn, or holding open tasks, gets one note once its session is up. That covers the daemon starting, crashes, runtime and Chrome changes, and the two new requests: - `restart_bot`: the session restarts and picks its conversation back up. - `clear_bot_session`: the bot restarts with a fresh conversation (Claude Code without `--continue`; Codex without its saved thread). Its workspace, FACTS.md and tasks stay, and its note says the conversation was cleared and points at what it keeps. Both need the `control` grant, work on a linked bot by asking its machine, and are advertised as `restart_bot`, so the phone can use them. The desktop adds Restart and Clear chat to the bot's header, each confirmed first. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each row in the Commands tab gets a Copy button for its command, and opening a row shows the full command and its output, each with its own Copy. The copy button is now one component, shared with code blocks. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each artifact now carries `created_by`: the bot (or the owner) a file came from, and how, shown under the file's name with the bot's avatar. - Written or edited: the first bot whose file tools touched the path, read from its transcript alongside its chat and commands. - Made by a command: otherwise, the first bot whose command named the path (a build, a copy, an image render). - Uploaded: files in `uploads/` are the owner's. - Sent: files a bot on another machine handed back with a result, traced through the message that carried them to that bot and its machine. A file nothing can account for is left unattributed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…d bots On a real project a quarter of the files had no maker: bots usually copy into the artifacts directory through a shell variable (`cp run.csv $A/31620468-vtrig.csv`, `cp *.png $A/b8b97477-frames/`), so their commands never held the file's full path. A command now also counts when it names the file's path inside the artifacts directory, or its folder, when that name is distinctive enough (12 characters or more) not to be any file. Archived bots' transcripts are read too, since they made files before they were archived. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A script a bot writes into the artifacts directory, once run, produces files no command names. Bots name a task's files after the task (`<task id>-summary.json`), so a file still unaccounted for goes to the maker of the other files named after the same task, as made by a command. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On native Windows `npx` is `npx.cmd`, a batch script that neither Claude Code nor Codex can start directly, so every bot's own browser would have failed to come up. The session now runs it as `cmd /c npx.cmd …`, the form Claude Code documents for `npx` servers on native Windows. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Owner
|
would be great if you could rebase it. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Seeing and steering what each bot is doing, from the desktop app and (through the same requests) a phone.
A browser for each bot
visioncoordinate tools, and so on.[browser]ingravityd.tomlcan turn this off or pin Node, the channel, headless mode and the package.--no-chrome. A per-bot switch, "Can use your Chrome" (set_bot_user_chrome, migration 17), gives it back for a task that needs the owner's logged-in sessions. The system prompt tells the bot which browsers it has.watch_browser, thenbrowser_tabsandbrowser_framepushes).DevToolsActivePortfile; no new dependencies.list_browser_activitylists each browser action under the request that started its turn, and marks actions in the owner's Chrome.Commands, memory, files
list_bot_commands). Everything the bot runs, foreground and background, from its transcript:FACTS.md, rendered as Markdown and reread as the bot works.created_byonlist_artifacts). Tried in this order:cp x $A/<task>-run.csv);Conversations between bots
list_agent_conversationsandlist_agent_conversationgive a project's bot pairs and each pair's thread, read from the two bots' DMs together, with no schema change.Restarts that keep the work going
resume_after_restart = falseturns it off.restart_bot: restarts the session, resuming its conversation.clear_bot_session: starts a fresh conversation, keeping the workspace,FACTS.mdand tasks.Across peer daemons (with #21)
peer_terminal). It's mirrored into the stand-in's buffer, so clients attach as to a local bot.peer_browser).Capabilities:
bot_browser,agent_conversations,peer_terminal,peer_browser,bot_commands,restart_bot. Clients show each feature only on daemons that advertise it.Why
How tested
agent_browser.rs: the Chrome switch and spawn flags, the per-bot browser config, the live view (two viewers sharing one screencast against a fake DevTools endpoint that keeps connections open as Chrome does), and the activity log.peer_browser.rsandpeer_terminal.rs: across two in-process daemons.agent_conversations.rs,bot_commands.rsandfile_authors.rs.resume.rs: restart, clear, a crash restart, the off switch, and restarting a linked bot on its machine.scripts/verify.shpasses: 397 Rust tests, 698 desktop tests, 89 visual stories.dffd258.🤖 Generated with Claude Code