v0.2.0

This release makes RoboCo resilient to LLM provider rate limits, gives the dashboard real per-agent token costs, and pushes system state to the panel in real time — plus the fix that finally lets the agents gate their own work.

Highlights

Provider rate-limit handling. When a provider returns a 429, RoboCo no longer drops work or spins. A Redis-backed tracker flags the provider, the spawn gate queues (never drops) further work for it, and affected agents park themselves via i_am_blocked(reason="rate_limited"). A background probe-and-resume loop clears the limit and revives the parked agents the moment it lifts — escalating to the CEO only if probes keep failing. The control panel shows it all live via a rate-limit banner.

Token usage & cost analytics. Every agent session's token usage is now captured from the Claude Code transcript and persisted into per-session rows and daily rollups, with provider-aware pricing (Anthropic models priced correctly; local/Ollama runs intentionally $0). The usage dashboard finally shows real numbers instead of zeros.

/ws/system operator stream. A new system-wide WebSocket stream and event bridge forward system events (starting with the rate-limit lifecycle) from the event bus to the panel in real time — the reusable foundation for future live operator views.

Fixed

• Agents can now run the full quality gate. Workspaces install the project's dev extra, so a spawned agent's environment carries ruff/mypy/xenon/pytest —
  closing the gap that previously let lint/type/complexity issues merge unchecked.
• Token-usage capture (the dashboard recorded zeros because nothing populated the per-session counters).
• The panel rate-limit endpoint shape and the doubled /ws/ws/system WebSocket path.
• Control-panel logo and all /public assets returning 500 (the image copied them without chowning to the non-root runtime user).
• Provider-aware model pricing (Opus corrected to $5/$25 per 1M tokens; non-Anthropic models no longer warn or mis-price).

What's Changed

• Fix: CEO reject conduit and main pm routing by @rennf93 in #91
• Fix: agent idle deadlock and lifecycle hardening by @rennf93 in #96
• [499f9eb1] Token Usage & Cost Analytics — Full-Stack Instrumentation, Persistence, and Visualization by @rennf93 in #90
• Chore: reduce analytics complexity by @rennf93 in #100
• [aaac85d2] Rate limit guardrails for Anthropic and Ollama providers by @rennf93 in #104
• Fix: rate limit real probe by @rennf93 in #110
• [4865ff8b] Add WebSocket support to the usage dashboard by @rennf93 in #115

Full Changelog: v0.1.0...v0.2.0

v0.1.0

RoboCo v0.1.0 — first public release

RoboCo is a virtual software company: 20 AI agents and one human — you, the CEO.
You hand work in at the top; the organization carries it through planning, building, review, and documentation, then brings it back for your sign-off. You act at the two ends; the agents fill in everything between.

What holds it together is structure, not a clever prompt — everything is a task, no task moves without acceptance criteria, and every task walks the same role-gated lifecycle: built, QA'd, documented, PM-reviewed, approved.

⚠️ This is v0 — early-stage homelab software. Expect rough edges, breaking changes, and an unstable schema. Run it on a trusted private network; do not expose it to the public internet as-is.

What's inside

• A real organization — CEO → Intake → Board (Product Owner · Head of Marketing · Auditor) → Main PM → three delivery cells (Backend · Frontend · UX/UI): 20 agents, formal communication, and a silent Auditor watching the whole time.
• The Task Assistant — chat with an agent that reads your actual codebase, interviews you, and drafts a well-formed task (objective, per-cell breakdown, acceptance criteria) ready to launch. RoboCo built this feature itself.
• A gated task lifecycle — backlog → pending → in progress → QA → documentation → PM review → CEO approval → done, every transition enforced by role.
• Real git, real PRs — each agent works in its own clone; cells open PRs into the Main PM's integration branch; only the CEO ever merges to master.
• Two human moments — approve the spec to start, approve the final PR to ship. The company runs itself in between.
• One cockpit — a Next.js control panel (tasks, Kanban, agents, journals, notifications, live metrics, the Task Assistant) behind a single nginx entry point.
• Grounded agents — a piragi + pgvector knowledge base, local inference via Ollama, cloud via Claude.

See it work

The walkthrough in docs/how-to.md follows RoboCo building one of its own features — the Task Assistant — across all three cells, failing and re-running its own QA, and opening the real pull request. RoboCo builds RoboCo.

Run it

Clone and docker compose up (builds locally), or pull the published images:

• GHCR: ghcr.io/rennf93/roboco-* — orchestrator, panel, agent-base, and the eight agent images, at 0.1.0 / latest
• Docker Hub: renzof93/roboco-*, at 0.1.0 / latest

Setup, configuration, and the security model are in the README.

License

AGPL-3.0. Contributions welcome under the CLA.

---

Full Changelog: https://github.com/rennf93/roboco/commits/v0.1.0