Changelog

Updates no longer report a problem while they're still starting up, and messages passed to your agent by other agents are no longer turned away.

Your agent runs on a newer engine with everything shipped since June, and updates no longer report a problem while they're still starting up.

Maintenance and reliability improvements.

• Download a backup of your agent and restore it later.
• The agent's own dashboard now works properly on phones and tablets.

Maintenance and reliability improvements.

Nova's system prompt now plainly declares what it can do (shell/filesystem, persistent memory, task list, scheduled runs) instead of instructing or guardrailing. The previous 'if you don't have the tool say so' / 'never crontab' phrasing confused weaker models (gemini-flash, openrouter auto) into falsely refusing to schedule or falling back to bash, even though the tools were present. Verified injection on live meshes: all capability flags were correctly set — this was purely a prompt-clarity fix.

Nova's system-tools hint no longer tells it to 'do real work yourself' via the shell. As a coordinator, Nova follows the user's request and delegates to member agents when it judges fit; the shell/filesystem are simply available if a request needs them. bash-is-not-a-scheduler guardrail retained.

Nova now understands its persistent capabilities.

• Prompt guardrail: bash is an ephemeral shell, not a scheduler/DB. Nova no longer fakes scheduling/persistence with crontab/at/systemd/disk; if a tool is absent it says so.
• Dynamic capabilities hint names the memory/tasks/cron tools actually enabled for the mesh, so Nova calls schedule_run / remember / create_task instead of the shell.

Pairs with control-plane fix (agent-aggregator #116) that stopped the nightly credential-refresh from blanking NOVA_MEMORY/TASKS/CRON.

Nova can now fully manage its own schedules and give them an expiry.

Added

• pause_schedule / resume_schedule brain tools — Nova could previously only create, list, and cancel (delete) schedules; it now pauses/resumes them without losing the job.
• Schedule end time — schedule_run takes an optional ISO until; backed by a new nullable end_at column (idempotent ALTER for existing volumes). The ticker auto-disables a job once it passes end_at.
• list_schedules now surfaces on/off state and end time.

Management API

• POST /setup/api/cron accepts endAt; PATCH /setup/api/cron/:id accepts endAt (and clears with null) alongside enabled.

Builds on v0.20.0 (per-run cron history). Full suite green (95 pass).

Scheduled runs (cron, docs/15) now record a full per-run history so a fire that failed is observable instead of vanishing into a Telegram message.

• cron_run history — each fire records status (running/ok/error), timings, error, summary, the full reply, orchestration trace, generative-UI parts, and reasoning.
• listCrons enrichment — every schedule carries lastStatus/lastError from its latest run.
• New endpoint — GET /setup/api/cron/:id/runs?limit=&before= (CONTROL_TOKEN), cursor-paginated newest-first.
• Deleting a schedule cascades to its run history.

Consumed by the dashboard's Schedules surface (list + detail view + run-history timeline rendered like a chat reply).

Nova gains three opt-in, durable capabilities (docs/15), each gated by an env flag and persisted on the /data volume — so what Nova learns or schedules in one conversation is available in every later one and every channel. Toggle them per-Nova from the dashboard (Routing & limits); curate them in the "Memory & automation" panel.

Global memory

Nova remembers durable facts (agent roles, your preferences, setup notes) across every conversation — a compact digest is injected into its reasoning each turn, so it already knows your setup in a fresh chat. Tools: remember / update / forget / recall.

Task list

A durable to-do list Nova can create, list, and complete — persists across sessions.

Scheduled runs (cron)

Nova (or you) can schedule a prompt on a 5-field cron expression; an in-process ticker fires it through the normal orchestration pipeline (it can consult agents on a timer). When Telegram is connected, results are also pushed to your chat — not just the hosted chat.

Management

• /setup/api/{memory,tasks,cron} control-plane API for the dashboard to view/curate everything.
• Capability flags surfaced on /meta, /nova/card, and /setup/api/status.

Off by default; a mesh opts in explicitly. Fully backward compatible.

Design

• The /chat sign-in gate is reskinned to the Build Your Future identity: the live animated Conway-cloud sky background, dark glass card, BYF cloud logo + wordmark, and a "by Swarm Network" byline.
• Fleshed out the gate to match the other agents' login — a Nova heading + subtitle and labeled username/password fields — and the credential hint now points to your Build Your Future dashboard.

No orchestrator / mesh / API changes.

Setup, settings, logs and loading screens now match the rest of Build Your Future.

Setup is now two steps instead of three, and setup, settings and logs match the rest of Build Your Future.

Your agent now looks like the rest of Build Your Future.

Signing in to your agent lands you on the same cloud-sky identity as your dashboard, and the sign-in hint tells you where to find your credentials — in your agent's settings — instead of naming things only we understand.

Reasoning effort & visible reasoning

• Consumes the composer's reasoning-effort pref. Reads the dashboard's {nova:"prefs", reasoning} directive (replacing the old model pref) and applies it as the AI SDK top-level reasoning param on Nova's brain + synthesis calls — sent to the gateway as reasoning_effort (minimal/low/medium/high/xhigh).
• Streams the model's reasoning back to chat. The brain captures reasoningText; it accrues live on task.reasoning, is mirrored as a {nova:"reasoning"} artifact part, and is exposed on both transports — A2A task frames/GET (dashboard) and hosted-chat SSE (new reasoning event + done). A collapsible "Thinking…/Reasoning" panel renders it in the hosted web SPA, and reasoning persists on the stored transcript so it survives reload.

Ships the same brain (Vercel AI SDK ai@7 / @ai-sdk/openai-compatible@3).

Scan a code and your Telegram bot is created and connected for you, with nothing to copy or paste. Disconnecting and reconnecting are one step too.

Scan a code and your Telegram bot is created and connected for you, with nothing to copy or paste. Disconnecting and reconnecting are one step too.

Connect Telegram by scanning a code.

Scan, and your bot is created and connected for you — no BotFather, no tokens to copy, nothing to paste. Disconnecting is one step too, and reconnecting later just works.

fix(telegram): render Markdown replies as Telegram HTML instead of raw text (bold/italic, code, bullets, headings, links via parse_mode HTML, with plain-text fallback).