What shipped. Release by release.

Pages your agents open for people. A dashboard where every page is an op an agent can call. Money ceilings that stop a turn before it costs. An upgrade that is an agent’s job.

The ten that matter

1. 01 v1.25 Widgets — pages your agents open for a person An agent, a pipeline or a hook opens a page: a Mini App in any chat the bot talks in, a page on the dashboard, a view in Claude, ChatGPT or VS Code. What the person does arrives as events; the agent pushes state back while it is open. A HITL card can open a page that answers it, and widget_preview shows the agent the page as the person will see it first. telegram a Mini App in a DM, group, topic, channel or business chat — or on the menu button dashboard in the chat panel, or on a request card mcp apps Claude, ChatGPT, VS Code the file <name>.widget.html — one self-contained page, sandboxed /docs/widgets/ →
2. 02 v1.20 – v1.25 Every page is an op Every dashboard page reads and writes through the same engine ops an agent’s tools call. Whatever a page shows, an agent can answer — grants, skills, memory files, models, connections, logs, run traces. /docs/dashboard/ →
3. 03 v1.22 – v1.25 The dashboard, rebuilt phone-first An Inbox to triage, with Boss mode in its header. Runs as now, then days. A session page is the whole conversation. One Cmd+K, and /go/<ref> links that survive a page move. The chat dock is the chief’s DM. /docs/dashboard/ →
4. 04 v1.21 Money budgets USD ceilings per brain, scenario or role. A spent budget refuses the next turn it covers before any model is called, and a card asks a person to raise it. Only metered API spend counts — never a subscription. /docs/budgets/ →
5. 05 v1.21 Feature contexts A feature’s rules and data live in its own *.context.md and attach only to a turn linked to it — a run that reaches it, a reply to its message, a chat bound to it, a role that always carries it, or your code. A turn with no link pays nothing. /docs/contexts/ →
6. 06 v1.17 – v1.21 Chat access The role belongs to the chat. Each person gets a tier — owner, member, stranger, blocked — and only an owner approves. Tell the agent, use the dashboard, or edit access.yaml; access_explain traces why. /docs/chat-access/ →
7. 07 v1.17 – v1.25 Least privilege, and one stop A role’s grant is its own file: silence grants nothing. Delegation is capped at the delegator; widening goes through a person. system_stop halts the whole operation in one command; an agent’s call waits for a person’s yes. /docs/contract/ →
8. 08 v1.22 Upgrades are an agent’s job omnislash upgrade moves the pin and writes upgrade/TASK.md; a coding agent finishes it. Every breaking release ships the prompt for that agent. Nothing upgrades a brain from CI. /docs/upgrading/ →
9. 09 v1.23 – v1.24 A voice per role A role speaks in its own voice and hears through its own transcription chain. Voice replies follow the text answer, switched in /settings. A persona speaks through its own bot. /docs/voice/ →
10. 10 v1.21 – v1.25 A harder loop Every cycle is graded improvement · regression · tradeoff · tie · inconclusive. No cycle’s journal outlives a deploy. A window waiting on a person is listed under ‘Needs you’; one that stops moving, as ‘Stalled’. /docs/self-evolving/ →

Release by release

1. v1.25.1 Oct 31 nothing to upgrade
   • A brain image installs from the same package store the engine image came from.
   • The doctor’s browser sign-in hint names the BotFather switch and fields to set.
2. v1.25.0 Oct 30 25 upgrade entries
   • Widgets: interactive pages in Telegram (any chat, or the menu button), on the dashboard and in MCP Apps clients, two-way.
   • A HITL card can open a page that answers it.
   • widget_preview renders a page in its real host and shows the agent the shots before a person sees it.
   • Sign in with Telegram in a browser, outside Telegram; agents list and revoke web sessions.
   • Emergency stop: system_stop / system_resume halt and resume the whole operation; an agent’s call waits for a person, the dashboard’s click does not.
   • New reads for agents: container logs, the event tail, the running configuration (secrets by presence only).
   • A scanned PDF’s pages reach the calling model, never the person’s chat.
   • The loop lists a window waiting on a person under ‘Needs you’, and one that stops moving as ‘Stalled’.
   • A feature ships its own skills.
   • The last legacy dashboard pages move to the new tree: Permissions, Chats, Knowledge, Connections, System.
3. v1.24.0 Oct 24 27 upgrade entries
   • Agents write grants (grant_create, grant_extend, grant_revoke); a widening still takes a person.
   • Memory, skills, models and providers, services and connections, repos and endpoints, artifacts and the tool catalogue become ops, each an agent tool.
   • A role speaks in its own voice (speech:) and hears in its own chain (transcription:); whisper steps take a prompt.
   • Voice replies follow the text answer, switched in /settings.
   • Every self-evolving cycle carries a grade.
   • Engine notices fold per subject and are gated by a notify level; Grafana alerts reach people through the engine.
   • A Claude turn costs its own calls, priced from models.json or list prices; a lapsed subscription login stays subscription.
   • A run reaches the contexts of the entity that ran, not of its whole feature.
   • Dashboard: a features page with lego figures, a session page that is a conversation, Runs as now over the days, Home in nine sections.
   • A pipeline handler’s plain throw fails the run at once; a retry is something you ask for.
4. v1.23.0 Oct 19 nothing to upgrade
   • A persona writes to you through its own bot, so your reply reaches the persona, not the chief.
   • A card a persona’s heartbeat raises stays in that DM.
   • The pipeline test harness runs handlers inside a pipeline invocation.
5. v1.22.1 Oct 19 nothing to upgrade
   • The engine image builds again.
6. v1.22.0 Oct 19 42 upgrade entries
   • The dashboard, rebuilt phone-first: one Cmd+K, /go/<ref> links, the chat dock as the chief’s DM.
   • An Inbox with Boss mode in its header; agents can start Boss mode, a person approves.
   • Upgrades become an agent’s job: omnislash upgrade writes upgrade/TASK.md; the auto-upgrade workflow is retired.
   • System messages ride the chief’s bot.
   • Workflows and pipelines start through one scenario_start and plan through one scenario_plan.
   • Every execution records the brain commit it started on.
   • New reads inside one run: run_trace, run_compare, run_history, run_timeline.
   • Heartbeats, hooks and the event catalogue become ops.
   • Every entity file write goes through one checked writer.
7. v1.21.1 Oct 11 nothing to upgrade
   • 1.21.0 shipped a migration out of order and refused to start; this is the stepping stone.
8. v1.21.0 Oct 11 13 upgrade entries
   • Money budgets: USD ceilings that refuse the next covered turn before any model is called.
   • Feature contexts: a feature’s rules and data attach only to turns linked to it.
   • The daily brief: an engine tick, one paid turn a day into the chief’s DM, on every brain.
   • Memory keyword recall is stemmed.
   • Chat access lives in constitution/access.yaml; a chat rule caps what a member’s turn may reach.
   • The self-evolving cycle is a Virtual Object — no journal outlives a deploy; the next cycle shows ahead as a dry run.
   • A run’s cost includes its sub-agents and turns that finish after a cancel.
   • Moving a role onto Codex takes a person.
   • A pipeline’s approval card expires after 168 hours by default.
9. v1.20.0 Oct 9 9 upgrade entries
   • The engine-api kernel: dashboard pages and agent tools call one layer of ops.
   • An agent’s model follows its role file.
   • Run control is one run_* surface: cancel, pause, resume, retry and recover any run under run:execute.
   • Read a loop round’s whole proposal.
10. v1.19.0 Oct 8 8 upgrade entries
    • An approved grant widens the agent — no capability ceiling without a way to ask.
    • Delegation is capped at the delegator: no spawning around a grant card.
    • The agent that holds the operation is the chief.
    • Loop targets are entities: a scenario, a heartbeat or a role.
    • One document reader: document_read takes PDF, Word, Excel, PowerPoint, OpenDocument, RTF, EPUB and HTML.
11. v1.18.0 Oct 7
    • “Owner” means one thing everywhere: a person with the owner tier.
    • Fix: an older OAuth login saves under the default MCP connection.
12. v1.17.0 Oct 7
    • Chat access: roles per chat and a tier per person, for every bot.
    • A role’s grant is its own file and silence grants nothing; its skills: list is an allowlist; Claude builtins follow the grant.
    • The context diet: progressive disclosure of rules, API prompt caching, wide tool sets deferred behind tool_search, a role-prompt budget.
    • A durable inbound sink, a GitHub App as a shared receiver, resource watches, missed deliveries recovered.
    • A sub-agent’s question asks its parent first, then a person with context.
    • Every doctor warning carries its remedy; Doctor rows end in “Fix with agent”.
    • The overview becomes a helicopter view, with a KPI band and an evolution feed.
    • One timezone and one brake over every clock.
13. v1.16.1 Oct 3
    • Deploys pull the private engine image.
14. v1.16.0 Oct 3
    • Agents create and run the self-evolving loop through validated tools; the human gates hold.
    • The loop’s proposer gets its own settings block and per-role prompts.
    • Approving a grant takes a person.
    • One vocabulary for the loop: self-evolving, loop, cycle, policy.
    • generate_image runs the brain’s image model and returns a file.