What shipped — Jun 21–28, 2026
A weekly, verifiable changelog from a studio running on its own agent fleet. 242 PRs merged across 32 repos this week. Most of that is internal or automated; here's what actually changed, and every public claim links to a real PR or release.
Update, July 16, 2026: warden has since been renamed redstamp (github.com/askalf/redstamp): same firewall, same code, same guarantees; this post keeps the name it shipped with, and every repository link redirects.
Security
Canon: signing keys move into the OS keychain. The canon signing key now lives in the OS keychain rather than a flat file. Tamper detection was already there, now the key itself is properly protected too. Canon's CI also gained macOS-runner support this cycle.
Keeper: secret-store socket tightened. Keeper's Unix socket is now created owner-only (0600) at startup, so only the owning process can connect, closing a local secret-exposure gap at the secret-store boundary. Platform-level warden integration also landed this cycle.
Dashboard: internal auth hardened. The operations hub's internal proxy gained an auth layer this cycle: rate-limit exemption paths now require internal auth headers rather than being trusted by IP.
Ships
deepdive v0.26.1. A source-trust badge now appears in synthesis output, so you can see whether a source was verified, inferred, or scraped from a content farm, and the multi-source fan-out now hard-biases toward primary sources. Faster synthesis that's less likely to hallucinate through aggregated noise. @askalf/deepdive on npm. Full case study →
dario v4.8.101. Output-format preservation improved, dario now holds the response shape the caller requested across more model configurations, plus the ongoing SDK-drift management that tracks upstream API changes and absorbs them before they break callers. 17k installs/month and climbing. How the release pipeline heals itself →
Update · August 2026: we quote lifetime installs now — 85K+ across npm, ghcr, and release assets — because monthly counts decay and lifetime only grows.
Platform internals: how we run the studio
Not a public product, this is the tooling we use to operate the fleet. Sharing it because the “how” is half the story.
Infra-health monitoring. Our internal cockpit now polls infra-health endpoints across all services. If something is degraded before a job runs, the fleet knows before the job does.
Human-in-the-loop triage. The cockpit surfaces flagged agent actions with approve / reject / escalate controls, so a human resolves them directly instead of hunting through ticket IDs, a first proper pass at human-in-the-loop review. The Discord remote's slash commands were also folded into the cockpit directly.
Cleanup & docs
Pre-launch repos scrubbed from public listings. Three not-yet-ready pre-launch tools were removed from public-facing docs, landing pages, and READMEs across all repos. They're real internal tools, just not ready to document publicly. Keeping the public surface accurate is part of the work.
SDK-drift tracking, made explicit. The dario README now spells out the drift-management approach: we track upstream SDK changes continuously and absorb them before they break callers. A busy upstream week is what that looks like in practice; a quiet one means there was genuinely nothing to absorb.
Install / upgrade
npm i -g @askalf/dario@latest # standalone dario
git pull && docker compose up -d # askalf platformaskalf is self-hosted. When something breaks, we say so here.
Early-access waitlist → askalf.org. Public claims above are verifiable on GitHub and npm (as of 2026-06-28).
We build and run AI systems in public: the security stack, the release pipeline, the research agent, on our own fleet. If you're putting an agent's output in front of someone who will act on it, that's the kind of problem we go deep on.
Start a conversation →