Skip to main content

Release Notes

The platform's changes, newest first. For technical detail see the repository's commit history and docs/adr/ records.

v1.5 — 18 October 2026: department view, learned guidance, multi-company authorisation​

  • Assistant view (ADR-0022 §7): the new assistant_user role; chat with their own assistant, send files, status events and "my context". Company-wide lists, developer credentials and installation-wide settings are closed to this role. → Assistant
  • Learned guidance (ADR-0022 §2): the harness change-request type; four allowed kinds, versioned storage, rollback, per-chat version pinning. Guidance enters the prompt framed as lower-authority suggestions.
  • Multi-company authorisation: policies, skills, departments, memberships, flows, sessions, invites, token minting, Git, audit, A2A, task streams, journal, dashboard and Telegram are now tied to the role in the owning company. Every endpoint requires sign-in (protected by a test).
  • Evaluation (ADR-0023): gateway failures (timeout, connection, unparseable reply, cut stream) are written to the audit; deterministic judges, success intervals and the Evaluation page.
  • A2A: the target agent's result does not reach the chat or the requester until the responsible human approves it.
  • Security: chat output HTML is sanitised synchronously and fails closed; Git sync uses a separate working copy per connection; the legacy admin-identity endpoints follow the same rules.

Limits: the proposer/probe/judge half of the learning loop, an outputs list and sources panel, the herdr plugin and the local Assistant folder do not exist yet. → Known limits

v1.4 — 18 October 2026: Authorisation hardening, evaluation, monorepo​

  • Change requests: company and role checks, scope (a department head only for their own department), the same person cannot approve both stages, one winner on concurrent approval.
  • Invites and role grants: who may grant which role is enforced in the API; an existing founder/executive membership cannot be lowered.
  • Evaluation (ADR-0023, stage 1): run diagnosis from audit events and the /eval API. → Evaluation
  • Audit chain: workstation events are stamped origin: client by the server; policy decisions carry the verified run.
  • Multi-company authorisation: policies, skills, departments, memberships, flows, sessions, invites and token minting are tied to the role in the owning company; a chat session belongs to its owner and the agent's managers; a draft agent cannot be used; an invite never resets the password of an account that belongs to another company.
  • Gateway: timeouts, connection errors, unparseable replies, cut streams and empty output are written to the audit event.
  • ADRs: 0003, 0006, 0014–0023 accepted; some decisions of 0014 and 0019 are narrowed/superseded by ADR-0022.
  • Repository: the documentation (apps/docs) and landing page (apps/landing) now live in this repository; the version number matches the code; the licence is Commons Clause + MIT in LICENSE.

Limits: judges, a viewer page and replay do not exist yet. → Known limits

v1.3 — 30 September 2026: Department Workspace (fab)​

The fab client in which every department works from the terminal, and the server side behind it.

  • fab client: sign-in, sidebar, local agent pane, fab workspace, fab sync.
  • Workspace API: create, resume, suspend, file manifest, single-use attach ticket.
  • One agent per person, token-budgeted prompt assembly, company context and job description, intent classification with Jev (first turn only, fail-open, off by default).
  • Work review and fit ratings: "My review", contesting, rubric and its linter, sampled rating, calibration tool, training-need signal. Off per company by default.
  • 5 new database migrations; hourly/daily scheduled jobs.

Limits: there is no real workspace runtime (only a FakeRuntime for tests); rating has not been measured with live Jev and has not passed legal review. → fab

v1.2 — 9 September 2026: Developer Workstation​

Software developers working on their own computers with opencode, inside a sandbox connected to the server. The web interface did not change; this layer works alongside it, optionally.

  • Central LLM gateway: OpenAI-compatible, bring your own key; model allow-list, rate and token quota, usage metering.
  • Executable policy engine: allow / ask / deny; denies on failure; off, dry_run, enforce modes.
  • Tamper-evident audit chain: hash-chained record, external anchoring (S3 Object Lock / local log).
  • Identity: short-lived persona tokens, refresh, revocation, sign-in with OIDC.
  • 3pa CLI, opencode plugin, signed managed configuration, sandbox and egress filter, prompt injection defense, LLM audit severity scoring and Telegram alerts.

→ Developer Workstation

v1.1.2 — 15 July 2026​

  • License note updated. Note: on 18 July 2026 the LICENSE file was turned back to plain MIT; the "Commons Clause" wording is no longer valid.

v1.1.1 — 14 July 2026​

  • AI provider settings were split into two sections: local models and cloud API services.
  • Delete is now visible on a provider card with an invalid key; deleting prevents the key from being re-added from .env.
  • Hard-coded Turkish text on the Dashboard moved into the translation system; the English title of the Organization page was fixed.

v1.1 — 14 July 2026: Production readiness​

  • One-command install (install.sh) and the .env wizard (setup-env.sh); production Docker configuration and an HTTPS (Let's Encrypt) Nginx template.
  • New version notice: the Dashboard checks the latest release on GitHub (1-hour cache).
  • Agent SLA table: for each agent sessions, tokens, flow success and task completion rate.
  • System requirements and production installation documentation.

v1.0 — 14 July 2026​

  • Local models: Ollama and LM Studio became first-class providers; the server address is entered and models are fetched automatically.
  • Change request flow: agent edits are not saved directly but enter the approval flow (submitted → dept_head_approved → admin_approved → committed), with a YAML commit to GitHub after approval.
  • Archiving: archiving instead of deleting an agent (inactive); data is kept.
  • The built-in tool list comes dynamically from the server.
  • Security: a critical warning at startup with the default JWT_SECRET.

v0.9 — 12 July 2026​

  • Token counting fixes (all tool steps are accumulated; Qwen fallback calculation).
  • Image generation in flows (Qwen Image / Wanx / Flux); task routing now also considers the company skill library.
  • Integration tests for GitHub, Telegram, PostgreSQL and scheduled flows.

v0.8 — 9 July 2026​

  • AI Setup: web research + chat + preview + one-click creation.
  • Skills and Policies pages (Markdown editor, change request flow).
  • In the sidebar, the structure pages were gathered under Structure.

v0.6 — 9 July 2026​

  • Dashboard rewritten: company and personal telemetry, memory summaries, company identity (vision, mission, values, goals).
  • Telegram integration for invitation and password notifications; the provider screen was simplified; sidebar collapse.

v0.5 — 9 July 2026​

  • Security: company membership check on all endpoints; password change required at first sign-in; typed input schemas.
  • Infrastructure: Alembic migrations, JSON logs, production Docker/Nginx configuration (login limited to 5 requests per minute).
  • Backup tab (S3-compatible) and social media skills (Instagram, WhatsApp).

v0.4 — 2 July 2026​

  • Qwen (DashScope) provider; a first-setup form instead of a fixed admin password; interface languages limited to Turkish and English.

v0.3 — 23 June 2026​

  • Audit log: a record of create, update and delete actions; Settings → Audit Log.

v0.2 — 19 June 2026​

  • Persistent memory across agent sessions; tool calls in flows; Turkish and English interface.

v0.1 — 1 June 2026​

First release.

  • Company, department, human and agent management; the org chart; five platform roles and the invitation system; authentication with JWT.
  • Provider key management (Anthropic, OpenAI, Google, Mistral); an agent runtime with tool calls, skills, session history, MCP support.
  • GitHub / GitLab / Gitea synchronization; invitation by email; delegation between agents (A2A).
  • SvelteKit 5 interface, multi-company use, light/dark theme.