📋 Work Board
Real-time state straight from GitHub + the on-disk queue — nothing cached or hand-edited. Open issues are pending work; PR lanes are in review; merged PRs are landed.
🚦 PR Lanes
Each agent prefix gets one open PR lane (CLAUDE.md monoworkstream rule). CI rollup is live from GitHub checks — green/red here is the real gate.
🤖 Auto-Pull Fleet
When on, the fleet pulls the top-priority open issue and works it into a draft PR unattended — one at a time, cloud-paused, one PR per lane, nothing auto-merges. This toggle is the kill switch (no restart needed); it persists across restarts.
📥 Manual Work
Picking issues by hand, researching them, copying coding-agent handoffs, and the "Recently Landed" shipped log now live on their own page — so this dashboard stays focused on the autonomous loop above.
📥 Open Work page →👥 Agent Slots
The roster from .claude/agent-slots.json, cross-referenced with the assigned queue — who's actively on an issue vs idle.
🔬 Research Team · Repo Memory
The fleet's grounding. Every doc + live module is learned into the one Convergence Memory as a grounded, evidence-cited record — so delegated agents and chat retrieve real repo facts instead of guessing. This is how "actually remember you" becomes true. Counts are read straight from the memory log; nothing is fabricated.
🧠 Local Model
The offline model wired into the coder/agent path — runs fully on this computer, no cloud. OLLAMA_MODEL is the pinned target; it only runs if it's actually being served on the local API.
⚡ Training Control
Fire a self-improvement run on the free GPU providers. Start fans out to every automatable provider with quota; or dispatch one provider at a time.
🔑 AI Provider Keys
Paste an API key to connect a model. Saved to this computer's Windows user account only — never uploaded.
🖥️ GPU Account Keys
Connect free GPU accounts so the Start button can dispatch training automatically. Set Hugging Face first — it's the checkpoint parking spot.
🔗 Provider Chains
Each task type tries providers in order — the first connected one wins. Set keys above to light these up.
🤖 AI Provider Reliance
Share of real AI work (Claude Code engineering turns + chat model calls). Reliance = how much each provider carries; reliability = its success rate. Local is bucketed under one Ollama row.
🏆 Chat Model Reliability
Granular view of the local/cloud models that serve chat — the small sliver above. Ranked by success rate.
⚖️ Trust Calibration
How much each agent's confidence has matched verified reality.