Skip to main content

Changelog

Major capability milestones, newest first. This is a curated highlight reel — the full per-slice history lives in roadmap.json.

2026-07-21 — adaptive executive support​

  • Automatic task adaptation — a deterministic private router detects observable activation, correction, load, reorientation, and task-complexity cues before each model call; no diagnosis or durable personal claim is inferred.
  • Harness self-redirect — action turns that accumulate six read-only calls or repeat the same failing call receive one bounded change-of-approach directive before the existing hard stop.
  • Transcript privacy — automatic support directives exist only in the prepared provider call and are never written into the session transcript; /support remains an explicit override.
  • Quiet MCP startup — configured connectors are dormant by default and no longer spawn or add tool schemas to every session; explicit panels, commands, runtime mounts, and an opt-in auto-mount setting preserve the extension path.

2026-07-20 — v0.9.5 Sight and workflow graphs​

  • Explicit screen context — /look, /look window, and /look screen attach a selected macOS area, window, or all displays to the next task turn with scoped capture receipts.
  • Desktop image context — the composer exposes the same capture modes and accepts clipboard PNG, TIFF, and JPEG images as visible, removable attachments.
  • Privacy-bound capture — capture is operator-initiated, native temporary files are removed after ingestion, unsent context expires after five minutes, and permission denial opens one concrete Screen Recording recovery path.
  • Graph engineering v1 — durable shared state, typed handoffs, review/rework cycles, adaptive topology policy, and replayable operator receipts coordinate existing Vanta loops.
  • Honest proof boundary — the CLI visual-question path and source/signed-package Desktop smokes are executed; one successful visual question from the signed app remains gated on macOS Screen Recording permission.

2026-07-19 — v0.9.4 desktop release​

  • Notarized public desktop build — the Apple Silicon DMG and ZIP are published with checksum and Apple submission 374f7536-59ba-4657-a437-b6d151d81445.
  • Independent Gatekeeper proof — a clean hosted Mac downloaded the public DMG, verified its checksum and staple, quarantined the extracted app, and passed signature and Gatekeeper assessment.
  • Release regression boundary — 36 Ghost light/dark captures, three fresh-profile startup samples, first-use latency, memory, CPU, and package-size budgets run in CI against the tagged source.
  • Cold-start repair — critical Work data enables the composer without waiting for optional Connect, Outputs, Canvas, messaging, or release-proof data.

2026-07-17 — desktop session safety​

  • Recoverable session operations — archive and Trash now expose pending, success, error, and Undo states; permanent deletion is restricted to confirmed actions from Trash.
  • Bulk session control — Select chats supports Shift-click ranges, All visible, bulk archive/restore, bulk Trash Undo, and explicit bulk permanent deletion.
  • Keyboard-safe menus — session actions use menu semantics, arrow-key traversal, Escape focus return, and outside-click dismissal, verified in Electron against an isolated session store.
  • Project context picker — the composer now attaches safe project files through Changed, Mentioned, Recent, and Search groups; ignored/private paths stay hidden and selected files remain removable before send.
  • Desktop release proof — one command now builds and signs Vanta.app, then runs the complete cold-start, approval, recovery, attachment, session Undo, Outputs/Connect, and three-viewport matrix against both source and packaged Electron.
  • Reliable access mode switching — stale background refreshes can no longer overwrite a newly saved project mode; the picker updates immediately and rolls back if persistence fails.
  • Outcome-oriented Connect — models, capabilities, and messaging adapters now show Ready, Needs setup, or Unavailable with safe local test actions; project file/catalog failures no longer misdirect operators into model setup.
  • First-class MCP connectors — CLI, TUI, and Desktop now share one project connector registry for source, trust, OAuth, enablement, health, tools/resources, redacted errors, and lifecycle receipts; kernel Block remains immovable.

2026-07-12 — Hermes delta and documentation release​

  • Executed product acceptance — real model read/write/readback, corpus recall, delegation, packaged desktop chat and destructive-command blocking, clean-install launchd scheduling, cold-start capability reporting, and bounded cited research now have recorded behavioral evidence. The refreshed public artifact remains the final local acceptance boundary.
  • Session-scoped model switching — /model and both model pickers now change only the current conversation by default, persist that route with the session, and reserve .env mutation for explicit Set as default/--global actions.
  • Per-call route usage ledger — completed agent-loop calls now persist the provider/model/base route that actually served them, including fallback depth, model switches, cache/reasoning tokens, and zero-cost local or subscription-included calls. /usage breakdown and /dashboard read the route ledger without double-counting legacy spend rows.
  • Real-headroom compaction guard — automatic compaction now uses the next provider input-token count to detect a fixed system/tool floor. Two ineffective passes suppress a third automatic rewrite, while /compact <focus> remains available.
  • Interrupted-tool effect safety — every tool attempt now records none, confirmed, or unknown; session restore preserves dangling mutations and requires state inspection instead of blindly replaying them. A redacted metadata receipt is kept in .vanta/tool-effects.jsonl.
  • Prompt presets and routed agents — /prompt can switch the current operating role, while delegate {agent_type} spawns a worker from the same project/home markdown definition with a bounded prompt, narrowing-only tools, and an optional model default. The base Vanta safety prompt and kernel remain enforced.
  • Current-main audit — compared Hermes Agent 4281151a with Vanta production paths, kept only six material local gaps, and ordered effect disposition plus real-headroom compaction first.
  • Hermes workflow wave shipped — persistent profiles and distributions, profile-routed Kanban, delegation receipts, corpus memory, typed context references, workflow blueprints, vault-backed secrets, credential pools, bounded plugin model calls, native deliverable attachments, and the public skill registry are now represented by shipped roadmap slices.
  • Honest release boundary — the roadmap reports 1,177 shipped, 2 remaining local cards, and 10 parked external acceptance gates. Parked provider, credential, and physical-device proofs are not reported as live.
  • Docs synchronized — comparison, roadmap, generated command/tool reference, and Cloudflare deployment instructions now derive from the current repository state.

2026-06-24 — model freedom & creative ideation​

  • Add any model provider with no code — declare any OpenAI-compatible backend in ~/.vanta/providers.json ({baseURL, keyEnv, model}); VANTA_PROVIDER=<id> just works. The secret stays in .env (the file holds only the env-var name); keyless local endpoints supported; a user entry can even override a built-in. See Providers.
  • Routers reach every model — pick tokenrouter/openrouter in vanta setup and free-type any model the token serves, instead of a pinned list.
  • Creative ideation, routed — the ideation-methods skill grew to 22 named methods from artists, scientists, and designers (Eno, Jarry, Cage, de Bono, Pólya, Alexander, Meadows, Vonnegut, Tharp…), with a feasibility↔creativity dial and a deterministic, tested router — so "give me a fresh angle" gets a deliberate method, not generic brainstorming. Fires automatically on ideation-intent prompts.

2026-06-22 — public launch (v0.2.0)​

  • Open source — the repo is public at github.com/jpoindexter/Vanta under the MIT license.
  • Zero-toolchain install — only git is required. The Rust safety kernel ships as a prebuilt binary and a portable Node 22 is fetched automatically on first run (both checksum-verified); no Rust or system Node to install.
  • Messaging — 20 adapters wired (Telegram · WhatsApp · Signal · Discord · Slack live today; the rest configurable).

Voice, desktop & on-device tuning​

  • Voice input — hold-to-talk speech-to-text, transcribed locally on your machine.
  • Native desktop control — Vanta can see the screen and click/type/scroll to drive native apps, in addition to the browser.
  • Terminal capture — read a live terminal pane into context.
  • Slack channel autocomplete — type # in the composer to pick a channel.
  • Parallel swarm — fan independent tasks across a live multi-pane terminal session, one worker per pane.
  • Personal tuning — train a small on-device adapter from the choices you've accepted and rejected, so Vanta drifts toward your preferences.
  • Self-serve setup — the setup wizard turns these on and walks you through the one-time OS permissions.

Docs & polish​

  • Full documentation site (this site) — diagrams, complete tool + command reference generated from source, use cases, guides, FAQ.
  • Modularity made the standard: ports/adapters seams + the code-size fitness function.

Operator runtime​

  • Operator systems — world model, Money OS, opportunity radar, background teams, life-wide search, self-repair compartments, reach doctor, verification organ.
  • Executive-function gates — research/inhibit/set-shift/closure/stall guardrails to keep long runs on-task; the task stack.
  • Self-improvement — the factory (vanta improve), the evolve loop, the eval harness, an independent critic.

Interface​

  • Terminal UI rebuild on real Ink — inline render + committed scrollback, markdown, diffs, approvals menu, cockpit, ⌘-style overlays; mission-control shell (VANTA_TUI=v2); desktop renderer.
  • Slash-command surface grew to 149 commands; per-turn tool scoping over 142 registered tools.

Selfhood & knowledge​

  • Brain — one cohesive memory unit (markdown regions + structured, decaying entries) with recall guardrails.
  • Knowledge graph + references for cross-session facts and ingested sources.
  • Corpus compiler — ingest note/transcript folders or guarded URLs, recall with keyword/semantic/entity fusion and source freshness receipts, refresh stale material, and preview/apply linked Obsidian exports.
  • MCP both directions — mount external servers as tools; serve Vanta as an MCP server.

v1 — feels like an operator​

  • Hook to any model (OpenAI/Ollama/Anthropic/Gemini/OpenRouter) via a first-run setup wizard.
  • Continuity — age-gated resume, fork, handoff, session memory, standing loops.
  • Self-improvement loop, bundled skill library.
  • Gateway — run as a service, messaging (20 adapters wired; Telegram · WhatsApp · Signal · Discord · Slack live), webhooks.

v0 — all the parts​

  • Rust safety kernel — risk classifier (allow/ask/block), approval queue, goal ledger, tamper-evident event log, HTTP sidecar.
  • Core agent loop + the first tools (files, web, browser, vision, code, comms).
  • Skills & per-goal memory; cron, subagents, operator modes.

Dates and exact slice history are tracked in roadmap.json; this page summarizes the waves of capability as they landed.