Andromity v0.2.22

[0.2.22] — 2026-10-20

Added

• Customizable VS Code session panel. Wide chat windows show context usage, uncommitted changes, MCP server status and environment controls. Drag a section's handle to reorder it, or use Up/Down, Home and End while the handle is focused. Sections start expanded; their order, collapsed state, visibility and the panel's hide preference are saved across chat tabs and reloads.
• Air-gap mode. Turn it on with [default] airgap = true in config.toml or ANDROMITY_AIRGAP=1. Andromity then makes no network connections except to this machine and private hosts in airgap_allowed_hosts / ANDROMITY_AIRGAP_ALLOW.
  • A guard installed at startup refuses connections, UDP sends and DNS lookups, including on the Windows asyncio path the old guard missed.
  • Cloud providers, web tools, remote and unapproved local MCP servers, connectors, telemetry and background lookups turn off with clear messages.
  • Commands the agent runs get stricter network blocking that an MCP server's env can no longer loosen.
  • New airgap.status and airgap.self_test RPC methods, and an airgap/blocked notification.
  • See docs/air-gap.md, including what the guard does not cover.
• Edit scheduled jobs in place. New cron.update and cron.duplicate RPC methods keep a job's id and run history; changes apply from the next run.
• Schedule preview. cron.preview validates a schedule and returns a readable description plus the next run times.
• More schedule formats: weekdays at 09:00, mondays and fridays at 5pm, every 1h30m, every 2 hours, cron day and month names (MON-FRI, JAN), and @hourly/@daily/@weekly/@monthly.
• New job settings, all optional:
  • catch_up: run once after downtime (the default) or skip.
  • max_retries: default 2.
  • notify: failures (default), always or never.
• Missed runs. A daily or cron-schedule job that Andromity was closed for runs once when it starts again, if the miss was within 24 hours. Otherwise the history records it as Missed.
• New run statuses:
  • Needs attention: a tool was blocked, or the agent ended with a question nobody could answer.
  • Skipped: for example, the workspace is untrusted.
  • Missed: Andromity was not running at the scheduled time.
  • Runs also record their trigger (schedule, manual, catch-up or retry) and the tools that were blocked.
• Run history keeps the last 100 runs per job.

Fixed

• A job's run history was wiped from the database every time any job was saved. It survived only through the .andromity/cron_runs JSON copies.
• Duplicate runs.
  • The VS Code engine started a new copy of a running job every 10 seconds.
  • A terminal and a VS Code window open on the same folder each ran the same jobs.
  • Jobs are now claimed in the database before they start, so each scheduled run happens once. Run now on a running job says it is already running.
• Provider failures were recorded as successes. Missing key, rate limit and quota failures now fail the run with a plain-language reason. Only temporary errors (rate limits, provider outages) are retried, with growing waits.
• Auto-pause. A job pauses itself after 5 failed runs in a row and shows why.
• Unattended runs.
  • The agent is told it is unattended and which commands it may run.
  • A blocked command now says "Blocked by this scheduled job's allowed commands" and lists them, instead of "rejected by user".
• Wrong folder. Scheduled jobs no longer run in the editor's install folder when VS Code has no folder open. Cron requests without a real project folder are refused, and jobs whose folder disappeared are paused.
• Stale session view. Opening a scheduled run's session while it ran could leave a stale copy showing only the prompt. It now shows the live run, and missing run sessions are never recreated empty.
• Live updates for open scheduled runs.
  • An open scheduled run streams like a normal chat: agent/started, text, thinking and tool events, then exactly one finish signal. That is agent/done for success or needs-attention, agent/error for a failure or timeout, or agent/cancelled when stopped.
  • Chat turns and scheduled runs now share one event-forwarding helper.
  • Stop in the chat cancels the run.
  • These notifications carry scheduled: true.
• Wrong permission mode. A scheduled run's session records the job's permission mode and profile instead of your default. session.get adds a scheduled_run field with the job, its mode and whether the run is still going.
• Messages during a run. Sending a message into a scheduled run that is still going is refused with "This scheduled run is still in progress. You can continue the chat when it finishes." Before, it started a second agent on the same session. Once the run ends, the chat continues normally.
• Untrusted workspace. A job no longer records a failed run every 10 seconds. It records one skipped run and pauses until you trust the workspace and resume it.
• Cron expressions fired only if a check happened to land inside the matching minute. They now run on time, and their next-run time is correct.
• Schedule parsing.
  • Out-of-range fields such as 99 * * * * were accepted and never ran.
  • Day-of-month and day-of-week now follow standard cron (either can match).
  • every 1h30m was read as one hour.
• Daily jobs (daily at HH:MM) created after that time no longer run immediately.
• A chat no longer shows as running forever when Andromity stopped mid-reply (VS Code closed, the engine crashed or was restarted). Each engine now records itself as the owner of a running chat, and any window can tell when that engine is gone; such chats show as interrupted. Chats left running by earlier versions count as interrupted after 30 minutes without activity. A run in another open window is never marked interrupted.

Changed

• A new interval job waits one interval before its first run. Use Run now to start it right away.
• A run stopped by shutdown is recorded as interrupted and does not count as a failure.

Andromity v0.2.21

[0.2.21] — 2026-10-18

Added

• 15 more app connectors, for 65 in total: Context7, Tavily, Firecrawl, CircleCI, Heroku, Pulumi, PlanetScale, ClickHouse, Clerk, WorkOS, LaunchDarkly, Better Stack, Ahrefs, Make and Calendly. Each is the vendor's official remote MCP server and connects with one browser sign-in.
• AI disclosure on posts. When the agent posts text other people read on your behalf (issue and pull request comments, descriptions, messages, emails, wiki pages) through an app connector or a CLI such as gh, it ends the text with "Written by Andromity, an AI agent, on behalf of the user who posted this." Drafts it hands back to you are not marked.
• Assisted-by trailer (off by default). [default] include_assisted_by = true, or Assisted-by Trailer in the terminal settings, adds Assisted-by: Andromity/<version> (<provider>/<model>) to commits the agent makes. A model switch mid-session updates the trailer.
• Each saved answer records the provider and model that produced it. Session exports name the model on every answer and say they were generated by Andromity, with a generator field in JSON and <meta name="generator"> in HTML.

Changed

• The GitHub Action's PR review footer says the review is AI-generated and names the model.

Andromity v0.2.20

[0.2.20] — 2026-10-15

Added

• App connectors. A catalog of 50 official remote MCP servers (Notion, Linear, Jira & Confluence, Gmail, Google Calendar, Google Drive, Stripe, Sentry, Supabase, Neon and more) that connect with one browser sign-in. In the TUI: /settings → MCP → Connect an app. Connectors are saved to ~/.andromity/mcp.json; nothing is saved if sign-in fails or is cancelled. Gmail, Calendar and Drive take your own Google OAuth client ID and secret. GitHub, Airtable, Render and PagerDuty connect with an access token: Andromity checks it with the service, then stores it in the encrypted token store, not in mcp.json.
• MCP server entries accept an optional token_prefix (default Bearer ) for servers that expect another Authorization scheme, such as PagerDuty's Token token=.
• New RPC methods connectors.list and connectors.connect.
• The terminal's Connect an app note links the setup guides at https://andromity.agenticmarket.dev/integrations.

Fixed

• MCP OAuth discovery rejected servers whose metadata differed only by a trailing slash in the resource or issuer URL, such as Vercel and Google. Google sign-ins now ask for offline access, so their tokens refresh instead of expiring after an hour.
• OpenRouter listed Andromity under two different app names. All Andromity requests now identify as "Andromity" from https://andromity.agenticmarket.dev, so OpenRouter counts them as one app.

Andromity 0.2.19

Fixed

• A background command (shell_bg) that exits within its first two seconds is no longer reported to the agent as running. The tool result now includes the exit code and output, with a reminder of the active shell's syntax. On Windows that shell is PowerShell, so a bash loop such as for i in $(seq 1 180); do …; done exits at once with a parse error. Previously the agent was told the process had started and kept saying it was running. The agent is no longer woken a second time for an exit it already saw.
• Background processes are registered before their output reader starts, so a process that ends instantly is no longer missed.
• Background process notices are no longer lost:
  • During a turn: if a process ends or you stop it while the agent is working, the notice is added to the running turn. Before, it was dropped.
  • Several at once: when several processes end within a moment (for example, you stop them all), the agent gets one message listing every one. Before, only the first reached it.
  • Settings: these turns use the session's own mode capped at TRUST, plus its profile and model, as before.
• Compacting by hand no longer stores a rough estimate as if the provider had reported it. As with automatic compaction, the size is unknown until the next request. Both kinds of compaction now send an estimate of the compacted context, flagged with context_estimated; it covers the system prompt, tool definitions and messages.
• A config.toml with a syntax error could lose your API keys. Andromity started with defaults, and the next setting it saved overwrote the file. It now keeps a copy as config.toml.unreadable-<time> before anything can be saved, says where in its warning, and reports it through the new config.status RPC. The warning goes to stderr instead of stdout.
• MCP:
  • TRUST mode checked the whole tool name, server included, for words like post or exec, so every tool of a server named postgres asked for approval. Only the tool's own name counts now.
  • TUI "Use PAT" rewrote the server into an npx mcp-remote … --header command whose header was never sent, so the token didn't work. The token is now stored like an OAuth or VS Code token and sent on connect.
  • mcp-remote configs keep their --header "Name:Value" arguments when Andromity connects to the URL directly.
• The Max Parallel Subagents setting wrote [subagents] max_parallel, which nothing read. config.get now reports [subagents] max_concurrent, the limit the subagent orchestrator uses (default 5).
• TUI text that didn't match the app:
  • Trust prompt: it said reading files is always allowed in an untrusted folder; reading is blocked until you trust it. Its dismiss button now reads Not now instead of Read-only.
  • Error hints: they pointed to Ctrl+M, Ctrl+P -> Settings and an andromity auth login command that doesn't exist. They now say Ctrl+L to switch model, Ctrl+E for Settings, and signing in from the VS Code extension.
  • Help and command palette: they listed Ctrl+J for switching profile; the key is Ctrl+U.

Security

• Skills that come with a project (.agents/skills/, .andromity/skills/, skills/) are left out of the system prompt until the folder is trusted. Before, an untrusted repository's skill names and descriptions reached the model even though the agent could not open them. Settings still lists them.

Added

• session.context RPC: an estimate of what fills a session's context window, by category. The categories are the system prompt, tool definitions, MCP and deferred tools, skills, the compacted summary, your messages, replies, tool calls and tool results. It also returns the provider-reported size of the last request, the model's window and the auto-compact threshold. The VS Code context popover uses it.
• Anonymous feature-usage events for automatic compaction (compact_auto) and TUI slash commands (tui_<command>).
• config.status RPC: the path of config.toml, and the parse error and kept copy if it couldn't be read.
• config.set_ollama_host RPC: sets the Ollama address, with the same URL checks as custom providers. The TUI's Ollama Base URL field uses the same checks and now reports an invalid URL instead of ignoring it.

Changed

• The shell_exec tool description now points long-running commands to shell_bg instead of suggesting & or Start-Process.
• The auto-compaction threshold (80% of the context window) is a named constant shared with the context breakdown.
• The telemetry notice in telemetry.py lists feature-usage events: feature names only, with the anonymous install ID and session ID.

Telemetry Dashboard

• A new Feature Usage page lists every tracked feature grouped by area, including ones nobody has used. Each feature shows users, uses and adoption, with search and used/unused filters. The overview's feature list is no longer cut at 25 entries, and its stale labels are fixed. These dashboard changes deploy independently of the core release.

Andromity 0.2.18

Fixed

• Ollama model discovery now returns only models installed on the running server. Offline servers and empty installations no longer fall back to catalog entries or stale cached models.
• Plan approval uses the active session's permission mode instead of the global default. Builder instructions now match automatic plan approval in Full and YOLO.
• Switching permission modes rechecks pending approvals against workspace trust and profile restrictions. Switching to Trust preserves checks for sensitive paths, execution-control files, commands, and external actions.

Changed

• VS Code chat has clearer permission and working-style menus, model activity indicators, tool targets, and Markdown formatting. See the extension changelog for details.

Telemetry Dashboard

• Session timelines show recorded error categories, such as “Model not found” and “Provider unavailable”, instead of a generic error badge. Older records without a category show “Error type not recorded”.
• A later successful turn no longer hides an earlier recorded session error. These dashboard changes are deployed independently of the core release and use existing anonymous telemetry without collecting error message text.

Andromity 0.2.17

Added

• Commits the agent creates end with Co-authored-by: Andromity <333054755+andromity-bot@users.noreply.github.com>. Turn this off with include_co_author = false under [default] in config.toml or the Co-author Commits switch in TUI settings.

Security

• VS Code: SAFE-mode approvals for write_file, edit_file and edit_file_multi show the lines being removed and added, instead of only the file path or raw parameters.
• Auto-woken sessions now run with their own permission mode and profile, capped at TRUST, and never in an untrusted folder.
• Session tools that can wake or steer another session (send, ask, broadcast, watch, shared state writes, handoffs) require workspace trust again and need approval in SAFE mode. Sessions only reach peers in the same project, and only the addressed session (or the user) can answer a question.
• Writes to .git/, .andromity/, .vscode/, .husky/, package.json, .envrc and .mcp.json need approval in SAFE and TRUST, since they can trigger later command execution.
• Cron jobs that arrive in a repository's .andromity/crons.json are imported disabled, and FULL/YOLO jobs are lowered to SAFE.
• VS Code: the sign-in token only goes to AgenticMarket hosts set in user settings, sign-in links must come from this editor, and model output can no longer render raw HTML or load remote images in chat.

Fixed

• Inter-session messages, questions and answers now appear in the TUI, and VS Code keeps them for sessions that are not on screen.
• Collaborator badges and per-session renames in VS Code, plus shared-state and handoff notifications.
• Sessions whose last turn was cancelled or failed can be woken again.

Changed

• Error telemetry records the error category (for example error_auth, error_rate_limit, error_ollama_offline), never the message text.

Andromity 0.2.16

Fixed

• Preserve LiteLLM proxy runtime modules (spend_tracking) in bundled daemon binaries to resolve Linux runtime crashes (ModuleNotFoundError: No module named 'litellm.proxy').
• Prune Next.js frontend chunks and stray test certificates from litellm/proxy to resolve VS Code Marketplace automated validation and security scanner rejections.
• Eliminate duplicate binary outputs (serverb) and enforce strict packaging exclusions for certificates and keys.
• Validate AgenticMarket account sessions before saving sign-in credentials and clear rejected sessions from the extension and daemon.

Added

• Automated provider runtime checks (--check-runtime) integrated into both the build script and GitHub Actions CI matrix across Windows, Linux, and macOS runners.
• Dedicated regression test suites in Python and TypeScript (packagingGuards.test.ts) guarding .vscodeignore rules and bundled binary integrity.

Andromity 0.2.15

Fixed

• Close Git review repository handles after RPC requests so Windows can release workspace directories.
• Refresh the VS Code chat, keep BYOK setup visible, clarify cloud trial status, and preserve manual scrolling while reliably jumping to the latest turn.
• Harden Docker defaults and release workflows with locked dependencies, restricted credentials, pinned Actions, and release validation.
• Reduce telemetry dashboard D1 reads with normalized shared snapshots, compact historical activity, grouped feature/session statistics, and session/day error checks. Ordinary ingestion no longer invalidates statistics snapshots.
• Count telemetry activity across complete windows, apply database cohort filters, and avoid multiplying session totals when multiple turn endings exist.
• Separate gateway first-token timing from full response duration and record failures within streams.
• Generate initial session titles from the user message before appending IDE context, preventing active-document metadata from appearing in titles.
• Reconnect enabled MCP servers when the IDE daemon starts, even when a manager already exists. Failed startup can retry, and changing projects closes previous sessions.
• Support remote MCP HTTP transport, protected resource discovery, and OAuth token endpoint client authentication required by providers such as Supabase.
• Refresh the TUI file sidebar after file creation, deletion, and moves. Reconcile affected visible directories without cascading into unrelated expanded folders; refresh collapsed folders when reopened.

Added

• Anonymous task outcomes and active turn timing, durable telemetry retries with idempotent ingestion, and MCP startup/cancellation signals. Dashboard statistics and administrative actions require separate server-side credentials and authenticated access.
• IDE RPC support for MCP configuration management and removal of selected installed skills, with workspace trust checks.
• MCP tool names and descriptions for collapsible IDE inspection while keeping schemas lazy.

VS Code Marketplace publishing is manual. Platform VSIX builds are produced by GitHub Actions. Python package version: 0.2.15; PyPI publishing is separate.

Andromity 0.2.12

🌿 Git & Multi-Repository Targeting

• Multi-Root Git Commit Generation: Resolved target repository directly from SCM button context, preventing cross-repo diff contamination in multi-root workspaces.
• Dedicated Staged vs. Unstaged Diffs: Prioritizes staged changes cleanly without polluting with unstaged modifications when generating commit messages.
• Commit Input Error Guard: Blocked AI error strings and backend exceptions from ever being injected into the Git commit input box.

💻 Terminal TUI & Clean CLI Formatting

• Core Error Decoupling: Decoupled error classification from HTML presentation in core/provider.py.
• Zero-HTML Terminal Rendering: Terminal TUI and standalone CLI output clean Markdown blockquotes and native Textual .session-error-card with zero raw HTML, <svg>, or <button> tags.
• Webview HTML Isolation: Preserves rich interactive HTML error cards exclusively for webviews.

🛡️ UX Hardening & Gateway Quota Transparency

• UTC Quota Countdown Timer: Dynamic live countdown badge (Quota resets in Xh Ym (00:00 UTC)) indicating exact daily reset times.
• Raw Exception Suppression: Omitted raw litellm.RateLimitError and internal stack traces on quota limits to maintain a clean, trustworthy developer experience.
• Zero Hardcoded Numbers: Removed artificial turn counts (30, 5) across all UI cards; limits and status derive directly from gateway allocations.

🔑 Authentication & VS Code Forks

• Universal IDE Fork Support: Deep-link login automatically detects editor URI schemes (cursor://, vscodium://, windsurf://, trae://, vscode://, code-oss://).
• Open-Redirect Hardening: Validated callback URLs via strict pattern matching to eliminate open-redirect risks while ensuring tokens route safely to the originating editor.
• Cross-View Auth & Limit Sync: Unified authentication status and live usage quotas across both sidebar and editor-tab views with bi-directional broadcast.
• Interactive Action Target Fix: Wired up [Sign In] and [Open Settings] click handlers in the webview error card.

📐 UI Dimensions & Modal Bounds

• Prompt-Aligned Interactive Cards: Constrained #interactive-slot, permission approval cards, and clarifying question carousels to match the prompt composer bounds across wide editor layouts.
• Typo Fix: Corrected CSS background property in markdown file pills.

---

Andromity v0.2.11

Andromity v0.2.11

Autonomous AI coding agent with dual TUI and VS Code experience. This release brings the new Live Waterfall Execution Profiler auto-docking, multi-agent session collaboration, and the official GitHub Actions CI/CD coding agent.

---

🌊 Live Waterfall Execution Profiler

• Auto-Open & Auto-Dock: Added configurable andromity.waterfall.autoOpen setting that automatically docks and opens the Live Execution Waterfall panel during multi-step reasoning and subagent runs.
• Span Synchronization: Eliminated timeline jitter during concurrent tool calls, ensuring live durations and status codes stream smoothly without dropped spans.

🤝 Multi-Agent & Session Collaboration

• Collaborative Session Inbox: Added dedicated Session Inbox in the webview to inspect and answer incoming messages between concurrent agents.
• Session Watch & Auto-Awake: Background sessions can enter watching mode and automatically awaken when another agent posts an answer or file handoff.
• Inline Permission Cards: Clarifying questions render inline using the permission-card aesthetic with custom radio options for fluid pair-programming.

🤖 Official GitHub Action (action.yml)

• Automated PR Reviews (mode: review): Autonomous code reviews on Pull Requests with 1-click GitHub suggestion blocks.
• Interactive Issue Tasks (mode: task): Solve issues and apply fixes directly from PR comments via @andromity <instruction>.

🎨 Session UI & Chat Experience

• Stepped Reasoning Popover: Minimal stepped slider (Off, Low, Medium, High) for reasoning effort.
• Cursor-Parity File Pills: Automatic stripping of raw ambient context blocks with clean, interactive file pills and diagnostics badges.
• Intelligent Auto-Scroll: Preserves reading position when reviewing history while maintaining automatic scroll-to-bottom for live incoming tokens.
• Post-Response Meta Footer: Sleek turn duration badge (3.2s), humanized relative timestamps (now, 2m ago), and one-click copy actions.

🛡️ Telemetry & Stability

• Repaired tool call tracking for shell_exec, shell_bg, find_files, and edit_file_multi.
• Fixed UUID session joining in telemetry worker and API key scrubber.
• 100% test coverage passing across all 513 Python backend and 81 extension tests.