5 more viral repos, distilled

A multi-tenant agent harness that isolates each user's data and credentials by scope, gates write operations behind approval, and supports pluggable skills across Browse, Linear, GitHub, and Admin. Sessions persist through disconnects, and every call can be run with full egress auditing.

The wedge is real: controlled, auditable agent workflows for startups that don't trust consumer chatbots with internal data. It's undercut by SECURITY.md itself, which admits the command policy is bypassable via obfuscation, and by a memory layer that enforces "no secrets" as an instruction rather than a code-level constraint.

A voice coding agent that keeps one conversation alive across desktop, web, and CLI frontends, surviving network partitions without losing context. Persistent per-owner sessions plus a FIFO async queue let the agent keep reasoning while voice input stays interactive.

The architecture is sound and well-tested, but the personas split hard on how permission consent should work: CTO wants auto-allow within policy limits to avoid fatigue, CISO wants explicit per-turn consent with zero inference. Persistent sessions make that tension worse, since resuming a conversation invites reusing permissions granted turns earlier.

A role-pinned delivery workflow that separates architect from implementer via TOML-configured agent roles, template verification, and structured handoff contracts, aimed at preventing design-by-committee compromise in multi-agent work.

One lane (Terra) is production-viable for private architect-led work. The other (Luna, meant for multi-task orchestration) exists only as documentation, not executable code, which blocks any enterprise adoption that needs code-backed enforcement rather than a prose contract. CPO and CISO also disagree on how reviewer isolation should work: shared context re-inspected later, or a fresh thread with no fork at all.

This one's distillx itself, seen from the outside — the same multi-lens analysis engine and 6-axis Scorecard that produces these teardowns, bundled with the Ponytail plugin ecosystem: a lazy-dev coding skill with adapters across Claude, Codex, OpenCode, Cursor, and Qoder, reporting a 54% median code reduction in its own benchmarks.

The panel's read: deterministic scoring (temperature pinned to 0) plus a sha256-logged egress ledger is a real differentiator against ChatGPT one-shots and consultants, neither of which can produce reproducible, auditable comparisons. The gap is operational rather than architectural — no CI/CD on the provider backends or egress ledger, no telemetry on trending batch runs, so nobody can measure whether the growth loop is actually working.

An Electron desktop app that records screen video, voice narration, window/app metadata, clipboard, and browser URLs, then runs multi-turn Copilot CLI agents over that capture to infer task intent and generate a production-ready skill or automation, preferring native tools over UI replay.

The capture pipeline is production-hardened and the agent integration is designed for scale, with bounded session pools and graceful degradation. Supply-chain discipline is notably tight: installs are pinned to exact commit SHAs for cryptographic integrity. Remaining gaps (Windows ARM64 CI, broader scenario coverage) are backlog items, not architectural blockers.