Repository navigation
Releases
0.11.0
Highlights
A new agent: IT Support. It sets up and fixes the machine the agents run on, whether Windows, Linux or macOS: installing programs and SDKs, changing system settings and environment variables, and starting the services an environment needs, such as Docker.
- It plans before it acts. The agent first looks at the environment without changing anything, then asks you to approve a plan in plain language that explains why each step is needed. Nothing changes before you approve it.
- The CEO only suggests it. The CEO never delegates to IT Support on its own: it asks you first, and the server rejects a delegation that does not follow your message or answer.
- It shows up in the squad. The web app gives it its own mark and color, and
/profileson Discord lists it so you can choose its agent, model and effort.
What's Changed
- feat: add the IT support agent by @douglasjunior in #29
- build: migrate to Rsbuild and Rstest by @douglasjunior in #28
- ci: block AI agent authors in commits by @douglasjunior in #27
Full Changelog: 0.10.0...0.11.0
0.10.0
Highlights
Large conversations open over remote access. A conversation with a long history no longer drops a WebRTC connection when it opens.
The whole history used to travel as one event, which took longer to cross a TURN relay than the web app's heartbeat timeout. The connection was dropped and opened again, in a loop.
- The history comes in pages. Opening a conversation loads its latest entries, about 200. The earlier ones load as you scroll up, or from "Load earlier messages", and the feed keeps your place.
- Pages hold whole turns. A page never splits a delegation from its steps, the squad panel still shows every agent at work, and an open question is always loaded.
- A long message keeps the connection alive. Every part of a message still arriving over WebRTC counts as an answer from the server.
Agent settings live in the squad panel. The settings button sits in the header of "Your squad", next to the squad status.
The header fits phones. From 320 px wide, the header hides the title and shows the connection status as an icon, so no button overlaps another.
Upgrading
The protocol version goes to 3. The hosted web app of remote access speaks only the latest version, so update the server to keep reaching it remotely.
What's Changed
- fix: page conversation histories by @douglasjunior in #26
Full Changelog: 0.9.0...0.10.0
0.9.0
Highlights
QA prepares its own test environment. The QA agent sets up everything a test needs by itself: it brings up local services and containers, applies migrations, runs seeds, starts the backend locally, points the application and its clients at it, and rebuilds the app. To reach a case that data alone cannot, it changes local databases, configuration or code (a fixed clock, a mocked response, a record on a given plan or date).
- Nothing reaches the repository. QA never commits, stages, pushes or opens a pull request with those changes. It reverts the tracked files before finishing, so
git statusshows what it showed at the start, and its final answer lists the local data or ignored settings it kept. - The developer comes back only for a change to commit. The CEO routes environment setup to QA as part of the test. A FAILED verdict sends the task back to development only when the report calls for a change to commit; when only the environment blocked the tests (an access, a credential, a decision of the user), QA is resumed with what it needs.
- Mobile. The
mobile-testingskill covers how a device reaches the host:adb reverseor10.0.2.2on an Android emulator,localhoston an iOS simulator.
What's Changed
- feat(qa): prepare the test environment itself by @douglasjunior in #25
Full Changelog: 0.8.0...0.9.0
0.8.0
Highlights
Remote access from anywhere. Reach your server from any browser in the world, with no port to open on your router and no VPN. Set REMOTE_ENABLED=true and REMOTE_URL, and the server prints a link at startup. The browser and the server talk over a WebRTC data channel, encrypted end to end, directly when the networks allow it and through a TURN relay when they do not. The connection badge in the header shows which of the two carries it.
- The link pins your server. The instance id derives from the server's certificate, and the page refuses any server that does not present it, so not even the signaling server can stand in for yours.
- The token stays between the browser and your server. The browser proves it holds the access token with a code bound to its own certificate before the server spends anything on it; the signaling server relays it without learning the token. Wrong proofs back off per address.
- The signaling server stores nothing. Hosts and sessions live in memory only, and it never sees your conversations.
- Self-hosting.
packages/p2p-server(signaling server and hosted web app) anddeploy/coturn(TURN relay) come with Dockerfiles, CapRover definitions and a deployment guide: docs/remote-access.md.
Public signaling server. https://arc-p2p-server.arc.douglasjunior.work is available for anyone to use at their own risk, as is and without warranty. Using it means trusting its operator with the page that reads your access token; read the disclaimer or host your own.
Upgrade notes
- Remote access is off by default. Turning it on requires
REMOTE_URL(HTTPS, or HTTP only forlocalhost) and an access token of at least 32 characters; leaveWEB_ACCESS_TOKENempty to use the generated one. - The certificate and key that identify the server live in
DATA_DIR/remote; keep that folder to keep the same link across restarts.
New settings in .env.example: REMOTE_ENABLED and REMOTE_URL. The signaling server and the relay read P2P_* and TURN_* variables, described in the deployment guide.
What's Changed
- feat: reach the server from anywhere over WebRTC by @douglasjunior in #24
Full Changelog: 0.7.0...0.8.0
0.7.0
Highlights
- The server now serves the web interface. The Docker image includes the web app and serves it on the same port as the server (
WEB_STATIC_DIR), so the page and the WebSocket share one origin and no nginx is needed. - Open
http://localhost:3000and read the access token withdocker logs <container>. WEB_ENABLED=falseleaves the web interface out.WEB_ALLOWED_ORIGINSis unchanged.
Upgrade notes
- Images derived from this one must run
node packages/server/dist/main. The oldnode dist/maindoes not exist.
What's Changed
- feat(web): serve the web app from the server by @douglasjunior in #23
0.6.0
Highlights
Web interface. A browser chat alongside Discord, served by the same server over a WebSocket and protected by an access token (generated into DATA_DIR/web/access-token when WEB_ACCESS_TOKEN is empty). It shows the conversation feed with each agent's steps, the squad panel with the coding agent, model and effort of every role, per-conversation settings, file attachments and voice messages, agent usage, and a skills page.
Task-driven board. The work of each conversation lives on a board of features and tasks that the agents read and write themselves, routed by the CEO through their statuses. The web interface shows it as a list or a kanban: create tasks, edit their sections, move them by drag and drop. Moving a task an agent is working on to done, blocked or cancelled stops that agent, after a confirmation. A task that loops between agents freezes after ARTIFACT_MAX_DELEGATIONS delegations (default 25) until the user takes it out.
Skills on demand. Agents read their know-how from skills listed in their instructions instead of carrying it in every prompt. The CEO creates and changes custom skills when asked, and they can be viewed and deleted from the web.
Agent login. Log the agent CLIs in from the web interface or Discord (/agent-auth), without a terminal on the host (AGENT_AUTH_ENABLED, on by default).
Team. A new QA agent tests the work in practice. Only the CEO, the Product Owner and the Tech Lead ask the user questions; an unanswered question falls back to its recommended option after QUESTION_TIMEOUT_MS (default 5 minutes). The CEO runs one agent per role and gets each result as soon as that agent finishes, and Git Ops performs destructive operations only on explicit instruction.
New settings in .env.example: WEB_ENABLED, WEB_ACCESS_TOKEN, WEB_ALLOWED_ORIGINS, AGENT_AUTH_ENABLED, QUESTION_TIMEOUT_MS, ARTIFACT_MAX_DELEGATIONS and LOG_RETENTION_DAYS.
What's Changed
- feat(git-ops): gate destructive operations on explicit instruction by @douglasjunior in #16
- feat(ceo): run one agent per role, pipeline work and hand results as agents finish by @douglasjunior in #17
- docs(readme): add portfolio footer by @douglasjunior in #18
- feat: web chat interface, protocol package and delegation improvements by @douglasjunior in #19
- feat(auth): agent CLI login from web and Discord by @douglasjunior in #20
- test(auth): fix flaky login code status check by @douglasjunior in #21
- Feat/task driven board by @douglasjunior in #22
Full Changelog: 0.5.0...0.6.0
0.5.0
What's Changed
- feat(shell): add /shell command to run host commands by @douglasjunior in #15
The new /shell <command> Discord command runs a shell command directly on the host, in the workspace directory, without going through the agents. The reply is visible only to the person who ran it and shows the output and exit code; /cancel stops a running command. The feature is disabled by default: set SHELL_COMMAND_ENABLED=true to enable it (optional SHELL_COMMAND_TIMEOUT_MS and SHELL_COMMAND_MAX_OUTPUT_CHARS). When enabled, anyone who can reach the bot can execute arbitrary commands, so use it only on private servers.
Full Changelog: 0.4.1...0.5.0
0.4.1
What's Changed
- fix(docker): build whisper.cpp without native CPU flags, fix(process-runner): treat signal kills as failures, fix(docker): drop NPM_CONFIG_PREFIX to silence nvm by @douglasjunior in #14
whisper-cli is no longer built for the CPU of the machine that builds the image, which crashed with Illegal instruction on hosts lacking those instructions. A process killed by a signal is reported as a failure instead of exit 0, so a crashed transcription shows an error instead of "No speech was detected". The image no longer sets NPM_CONFIG_PREFIX, which made nvm warn on every new shell.
Full Changelog: 0.4.0...0.4.1
0.4.0
What's Changed
- feat(speech-to-text): transcribe audio attachments by @douglasjunior in #13
Audio attachments and Discord voice messages are transcribed on CPU with whisper.cpp and the text is sent to the agent as the prompt. The provider is selected with STT_PROVIDER (whisper-cpp or none). The Docker image ships ffmpeg and whisper-cli and downloads the model on startup; see the README for the STT_* variables.
Full Changelog: 0.3.0...0.4.0
0.3.0
What's Changed
- feat(usage): show progress bars in /usage by @douglasjunior in #12
Full Changelog: 0.2.0...0.3.0