v0.5.0

Antares v0.5.0

Built from main commit 3fc4498.

Dashboard: compact navigation

• The sidebar goes from 25 pages to 8 hubs with tabs: Chat, Sessions, Agent, Capabilities, Automation, Security, Studio, and System. Old URLs redirect to their new place, keeping query strings and anchors.
• Optional modules (Automation, Security, Studio) can be hidden. New installs pick a use-case preset during setup: General, Coding, Security, Creator, or Full. Change it any time in System › Settings › Modules. A hidden module still works and stays reachable by link and search.
• Command palette (⌘K / Ctrl+K) finds pages (including hidden modules and old page names), recent sessions, and quick actions.
• Per-feature settings moved next to their pages, behind a gear button on each hub tab. Settings keeps global options, and its search points to settings that moved. Language and theme live in Settings › Appearance.
• The sidebar can collapse to an icon rail.
• Studio › Creator is reorganised around the project: project list, per-project tabs, a Research → Plan → Produce → Publish pipeline with Run all, compact shot rows and reference cards, and a New video dialog. The Models tab is gone; image and video models are set from the page's settings gear.
• Studio › Social shows mailbox, browser, and autopilot as one compact strip, with the IMAP form in a dialog.

Chat

• The transcript stays pinned to the bottom while a reply streams and after it finishes. It no longer jumps to the last prompt at the end of a tool-using turn, scrolling up is respected until you scroll back down, and sessions open at their newest line.
• Each turn is one message in history too, instead of one bubble per model call.

Skills

• Automatic discovery of SKILL.md skills from conventional user and project directories (.agents/skills, .claude/skills, .codex/skills, and others), refreshed live and scoped to the session's project. Discovered skills are read-only in the dashboard (#42).
• Turning a skill on or off is stored in configuration (skills.disabled) instead of rewriting the skill file (#44, delivered through #42).

Terminal UI

• /undo rolls back the last turn; /revert picks any earlier turn, or takes a message id. Both restore changed files and trim the conversation (#43).

Media

• New image_gen.reference_mode: edits (default, OpenAI /images/edits) or generations (reference images sent to /images/generations), for OpenAI-compatible gateways without an edits route.
• An image endpoint that answers with an HTML page or no image now says what came back, instead of "returned nothing usable".

Fixes

• Saving from Settings no longer crashes the page.

Upgrade notes

• Back up your Antares state directory before upgrading, especially the database and configuration.
• On first start, skill enablement is migrated once: enabled: false headers in configured skill directories move into skills.disabled in config.yaml.
• Existing installs keep every module visible. Only new installs choose a preset.
• If your image gateway has no /images/edits route, set image_gen.reference_mode: generations.

Known issues

• The "restart to apply" notice does not appear after saving a restart-required setting; the value is saved.
• With server.auth_token set, a browser that has no token can reload the dashboard in a loop.
• Studio pages are not yet translated in non-English languages.

Verification

go vet ./... and go test ./... pass, frontend tests (166) and typecheck pass, and the dashboard smoke walk loads every route and legacy redirect at desktop and mobile sizes. No paid model generations or social publications were performed during release verification.

v0.4.0

Antares v0.4.0

Built from main commit 4237ac1.

Content Creator

• New Content Creator specialist and dashboard for research, ideation, visual references, shot planning, clip generation, assembly, and publishing workflows.
• Configurable OpenAI-compatible image generation/editing and asynchronous video endpoints. Continuation shots reuse the previous clip's final frame; project state and provider job IDs persist between runs.
• FFmpeg/FFprobe integration assembles clips into H.264/AAC MP4 files and verifies resolution and duration.
• Uploads reuse the existing Social Media browser and accounts. Draft mode is the default; publication requires explicit authorization. Uncertain uploads require reconciliation before another attempt.
• Existing cron jobs can select the Content Creator role, project, stage, and publish mode.

Chat and agent reliability

• Terminal failures are retained as JSON in chat history, with Copy error and Retry controls.
• Retry resends the relevant prompt and attachments without deleting conversation history.
• Subagents cannot block on ask_user. Scheduled actions requiring interactive approval report a blocker instead of waiting indefinitely.
• Streaming, tool-result ordering, subprocess delegation, and file/terminal handling improvements.

Models and configuration

• models.dev metadata enrichment with a shared provider allowlist, last-known-good handling for empty refreshes, deterministic model lookup, and correct vision capability flags (#40).
• Custom provider HTTP headers across dashboard and onboarding, with credential isolation between newly named providers (#41).
• Fresh non-container installs default to loopback. ANTARES_HOST overrides apply at runtime rather than contaminating first-run configuration; container defaults remain supported (#39).
• Generated TypeScript build metadata and local tooling artifacts are ignored (#38).
• Restored Windows builds with a pinned HNSW portability fix; SQL-backed RAG persistence is unchanged.

Upgrade notes

• Back up your Antares state directory before upgrading, especially the database and configuration.
• Content Creator requires configured image/video provider credentials. Custom endpoints must implement the supported OpenAI-compatible image and video protocols.
• Install ffmpeg and ffprobe separately for video processing. They are not bundled in these binaries.
• Connect the intended social account in Social Media before enabling publishing. No paid model generations or real social account publications were performed during release verification.
• The previously removed Cursor Cloud Agents integration remains removed.

Downloads

Choose the binary matching your operating system and CPU architecture. The dashboard is embedded in every binary. checksums.txt contains SHA-256 hashes for all six platform binaries.

Full changelog: v0.3.0...v0.4.0

v0.3.0

What's new in v0.3.0

Features

• Managed background processes. terminal(background=true) now returns a stable process_id immediately, plus a session-scoped process tool with list, poll, log, wait, and kill. Incremental output with absolute offsets and a bounded rolling buffer, explicit lifecycle states, and process-group termination on kill/timeout/Stop/session close. (#9)
• Background server lifecycle. antares serve now detaches by default and returns once the daemon owns the listener and health is ready. New antares stop (graceful SIGTERM → process-group SIGKILL) and antares status (PID, dashboard URL, version, log path). antares serve --foreground and bare headless antares are preserved for systemd/containers. Stale/reused PID files are rejected via kernel start-time, UID, executable, cmdline, and socket ownership. (#10)
• Reasoning-effort picker in the composer. Choose the reasoning effort per turn (Default / Off / Low / Medium / High) straight from the chat composer; the request value overrides the configured default.
• Correct per-model context window. The context gauge and compaction now resolve each model's real budget (e.g. glm-5.2's 1M) instead of a flat 200k default, via a per-model catalogue with a model_meta override.

Fixes

• Context accounting. Cached prompt tokens are no longer double-counted: providers now expose a normalized context size (OpenAI/Gemini use total prompt tokens; Anthropic adds its separately-reported cache read/write), and tool schemas are included in request-size estimates so compaction fires at the right time. The runaway "3.5M/200k (100%)" gauge is fixed — it now plots the current turn's context, not the run's cumulative input. (#7, #5)
• Chat renderer hangs on long sessions. Idle /chat/attach polls no longer re-download and re-hydrate the entire transcript every 1.5s, reconnects resume from a real cursor instead of replaying the whole backlog, and text/reasoning deltas are coalesced on both the client and the server. (#8, #5)
• Terminal hangs on non-POSIX shells. Persistent terminal sessions no longer inherit a $SHELL such as fish that rejects the POSIX completion sentinel. (#7)
• Sub-agent workspaces. Background sub-agents inherit the parent project workspace (and honor isolated worktrees) instead of falling back to the global workspace. (#7)
• Tools page crash on ungrouped/MCP tools (toolsets encoded as [] rather than null). (#11)
• uTLS panic (extension already initialized) under parallel http_request, via httpcloak v1.6.8 (requires Go 1.26).
• Edit / revert the last user message now works right after a turn without a manual reload.

Notes

• Building from source now requires Go 1.26.

Thanks to @nvdorman for #7, #8, #9, #10, and #11.

v0.2.0

Antares v0.2.0 — a big feature drop on top of v0.1.0. Native RAG, project-aware sessions, an agent identity file, dashboard security, and a lot of UI polish.

Single self-contained binary (Go backend + embedded React dashboard). Download the build for your platform below, chmod +x, and run antares serve. First run walks you through a setup wizard.

---

✨ Highlights

Native RAG (no external service)

• Dropped the enowx RAG backend entirely — retrieval now runs natively inside Antares against its own DB.
• Full pipeline: recall → rerank → compress (dedupe) → top-K.
• Smart reranking — rerank_mode: llm auto-uses Voyage rerank when a Voyage key is present, falls back to LLM reranking otherwise, and can be turned fully off. Custom rerank model/URL/key supported via rag.rerank_*.
• Embeddings: native Voyage or any OpenAI-compatible embeddings endpoint, chosen in the setup wizard (rag.embed_provider).
• Auto-context: relevant snippets are retrieved and injected into the turn automatically (rag.auto_context).
• Measured search latency dropped from ~6.9s to ~0.6s on the same query after the rerank rework.
• RAG playground on the Memory page: pick a collection, tune top-K, and see the exact pipeline + latency + per-result scores.

Project-aware sessions

• Bind a chat to a project folder — the agent is write-confined to that project and picks up its context.
• Project sidebar with tabs: Info · Plan · Git · Changes · Files · Scripts · Tools.
• Per-project env editor and a one-click Analyze / index action that builds a readable RAG collection named after the folder (<folder>-<hash>).
• Tools-usage panel: which tools the session used, background-tool activity, and when each was last used.
• Sessions page now splits into Chat and Project tabs, with a folder chip on project sessions.

SOUL — agent identity

• A SOUL.md identity file the agent reads to shape how it speaks and behaves, editable from the new Soul page and via a set_soul tool.

Dashboard security

• Password login for the dashboard (bcrypt, cookie-based) — separate from the API bearer token.
• Sensitive features gated behind the dashboard password: VPS, proxies, file manager, and project/workspace browsing.
• The password hash is hidden from the config editor.

VPS manager

• Connect, monitor, and manage servers over SSH from the dashboard, with a vps_run tool for the agent.
• Encrypted credential storage, TOFU host-key pinning, host/IP spoilered in the UI, and a scrollable processes modal.
• Dedicated VPS Manager role.

OSINT & proxies

• osint_email_full — emailosint.org lookup with an in-browser Turnstile solve (headless-reliable, no zombie processes).
• Cookie-gated Google profile lookup (osint_google) accepting Cookie-Editor JSON + SAPISIDHASH auth; multi-account selection.
• Email → handle pivots and an osint_pivot tool for branching investigations; fewer username false positives from soft-404s.
• Global proxy store with a management page, batch/file add, and agent-picks-a-proxy routing.

Other

• Update checker with an in-dashboard banner when a newer release is available.
• Edit a past message and revert — per-turn checkpoints roll the workspace back to that point.
• Context-window gauge in chat showing how full the window is.
• Mobile-responsive dashboard throughout (slide-over sidebar, icon-only composer chips, tap-to-open context ring).
• Discord/Telegram gateway: role-gated server bindings, reliable "addressed" detection, per-channel reply style, DM-only pairing.

🐛 Fixes

• Sessions no longer show 0 messages / 0 tokens despite having a chat (session-counter ownership fix).
• Chat remembers the last-used role and shows the persisted model on load.
• Default workspace is no longer created before the setup wizard finishes (#4).
• Dashboard auth endpoints correctly exempted from the bearer gate.

📦 Install

Download the binary for your platform, then:

chmod +x antares_v0.2.0_<os>_<arch>
./antares_v0.2.0_<os>_<arch> serve

Verify your download against checksums.txt (SHA-256).

Full commit log: v0.1.0...v0.2.0

v0.1.0 — early release

First tagged release of Antares — a self-hosted AI agent (Go backend, React dashboard, one binary).

Early release. It runs and is used daily, but the surface is still moving and rough edges are expected. Bug reports, feature ideas, and pull requests are welcome — open an issue or send a PR.

Install

Linux / macOS

curl -fsSL https://raw.githubusercontent.com/enowdev/antares/main/scripts/install.sh | bash

Windows (PowerShell)

irm https://raw.githubusercontent.com/enowdev/antares/main/scripts/install.ps1 | iex

While the repo is private, install the GitHub CLI and run gh auth login first — the installer uses it to fetch these assets. Full guide: docs/installation.md.

The installer downloads the prebuilt binary for your platform, installs it (~/.local/bin on Unix, %LOCALAPPDATA%\Antares\bin on Windows), and adds it to your PATH. Then run antares setup.

Binaries

Linux, macOS, and Windows on amd64 and arm64. checksums.txt has the SHA-256 of each. The single binary embeds the dashboard.