Tailscale, Research Papers, and a Brand-New Blog: b4arena's Busiest Week Yet
The week of March 9 saw b4arena shift from "getting things running" to "getting things right" — hardening infrastructure with Tailscale and SSH lockdowns, publishing a wave of research papers on the Tabula documentation site, and extending the agent platform with new capabilities across the board.
Infrastructure Gets a Security Overhaul
The infra repo dominated this week with 37 commits, most of them focused on a comprehensive security hardening sprint. Tailscale landed on mimas, replacing direct port exposure with an encrypted mesh overlay — observer services like beads-ui and PinchChat now bind exclusively to the Tailscale interface, eliminating accidental LAN exposure. The gateway switched to Tailscale Serve mode with explicit allowed origins, and a centralized SSH authorized-keys playbook replaced ad-hoc key management across hosts.
Alongside the security work, the platform completed its OpenClaw 2026.3.x upgrade, switching the model backend from direct API keys to OpenRouter auth profiles with claude-3.5-haiku. A dedicated Dolt standalone playbook was extracted from the monolithic provisioning flow, making database operations independently deployable. The week also saw IPv4-only container networking fixes for pasta, SOPS recipient additions for rpi5, and a fresh lessons-learned.md capturing upgrade gotchas for posterity. → SSH hardening
Arena's Knowledge Architecture Expands
The arena repo — b4arena's conceptual backbone — saw 13 commits adding substantive research and architectural content. Three new research papers landed: "Agentic Company Archetypes" classifying different organizational patterns for AI-native companies, "Agent Specs and Failure Modes" analyzing how specifications break down in agentic systems, and a "Verification Loop Gap" document identifying the missing feedback layer between agent execution and outcome validation.
On the operational side, Marcel contributed a gateway-auth reference doc documenting authentication modes and device pairing behavior, while the delegation-chain FSD was updated to reflect the platform's current flat routing model. A CI workflow was added to automatically trigger Tabula deploys whenever research documents change — closing the gap between writing and publishing. → verification loop gap
Tabula Becomes a Real Documentation Platform
Tabula went from a bare Docusaurus scaffold to a branded, content-rich documentation platform this week. The site received a complete visual identity — custom logos (light and dark), a social card, and a redesigned homepage that communicates what b4arena actually is. Google Tag Manager and GA4 tracking went live for usage analytics.
More importantly, the content architecture took shape. All seven arena research papers received Docusaurus frontmatter and are now browsable under a dedicated Research Papers section. The documentation hierarchy was organized into conceptual layers (L0 philosophy through L3 runbooks), with mental models and architecture overviews filling the gaps. A blog section was enabled with b4arena-specific authors and tags — the post you're reading right now is the first output of the new Dispatch blog mode. CI pipelines handle cross-repo deploys via repository_dispatch, so research paper updates in the arena repo automatically trigger a Tabula rebuild. → branding and homepage
Ludus Grows New Agents and Sharper Tools
The ludus CLI saw 33 commits spanning agent development and operational improvements. Two new agents joined the roster: Glue, a coordination agent, and Indago, a research-focused agent — both wired into the sandbox infrastructure with proper workspace configuration. The sandbox system itself received significant attention with UID mapping, permissions fixes, and ca-leash workspace configuration updates.
On the tooling side, the observe repo contributed pinchchat management recipes, an ops-dashboard command, and a refactoring pass that removed PinchChat lifecycle management from the CLI (now handled by Quadlet containers). The meta repo extended the arena skill with an issue-triage workflow, a full agent-browser skill for web automation, and updates to the observe and healthcheck workflows to reflect the Tailscale migration. → sandbox enhancements
The AI Behind the Curtain
Claude Code burned through 172 million tokens ($77.75) across roughly 267 prompts this week. Friday was the heaviest day at $31.23 — an intensive infrastructure debugging session that resolved container networking issues and finalized the Tailscale migration. The model mix shifted from Haiku-dominated early in the week to Opus 4.6 for the more complex architectural and research work. The bulk of AI-assisted effort went into the meta repo (242 prompts), driving infrastructure configuration, Tabula branding, and intercom bootstrapping. Token spend is trending upward as the platform's complexity grows, but the ratio of commits-per-dollar remains healthy at roughly 1.7 commits per dollar spent.
By the Numbers
| Metric | Value |
|---|---|
| Commits | 134 |
| Active repos | 7 |
| Claude Code spend | $77.75 |
| Tokens consumed | ~172M |
| Period | week of 2026-03-09 |
This post was generated by Dispatch, the b4arena activity recap system.