Skip to content

Biweekly Update Note: 2026-06-14 to 2026-06-27

Source Boundary

This note summarizes public repository history from 2026-06-14 through 2026-06-27. It uses public commit and PR history, shipped docs, examples, and smoke tests. It does not include private operator state, raw benchmark evidence, private links, local paths, or credentials.

Highlights

  • Renamed the project to LoopX and clarified the product shape: a local control plane for long-running agents, not a hidden executor.
  • Hardened benchmark and skills workflow support across route contracts, launch blockers, lifecycle closeout, and developer-facing evidence rules.
  • Added Codex App and CLI-oriented entry points for /loopx, /loopx <goal>, global summary commands, local launch dry-runs, and host integration plans.
  • Advanced issue-fix from a concept into a workflow with issue metadata intake, ordered todo previews, validation evidence, and PR review packet readiness.
  • Expanded multi-agent control-plane projection with scoped gates, task graph projection, cold-path todo detail, event-sourced state planning, and contributor task refreshes.
  • Improved scheduler safety with reset policy, repeated monitor detection, wait cadence simplification, status/quota budget checks, and an integrated control-plane canary.

What Shipped

Benchmark Workflow Hardening

LoopX added more explicit benchmark route and lifecycle contracts for SkillsBench, Terminal-Bench, ALE, SWE-style routes, and benchmark developer workflows. The work concentrated on making benchmark runs attributable: launch blockers are classified, runner prerequisites are preserved, case-local lifecycle discipline is measurable, and closeout requires real solver or validation activity.

Host Commands And Agent Entry Points

The product gained a clearer command surface for Codex App and CLI users: /loopx for project bootstrap, /loopx <goal> for explicit goal-start planning, /loopx-global-* commands for global summaries and gates, and local launch dry-run protocols. The near-term path remains skill-level fallback; the longer-term path is host command registration with explicit project root, agent id, and permission boundaries.

Issue-Fix Workflow

Issue-fix work moved toward a complete maintainer workflow: classify an issue, turn it into ordered owner/todo work, run validation, capture fail/pass evidence, and produce a PR review packet. The workflow is designed to keep human gates explicit when an issue requires maintainer judgment.

Task Projection And State Reliability

The control plane added richer read surfaces around todos, gates, claims, recent runs, and blockers. Task graph projection provides a compact "why is this stuck" view, while cold-path todo detail keeps heavy per-todo data out of hot status payloads. Event-sourcing work began moving canonical todo/history toward append-only records while keeping Markdown active state as the human and agent work surface.

Product Surface And Public Docs

README and docs work clarified the LoopX positioning, showcase catalog, host integration, onboarding help, contributor task board, and public/private boundary. Some visual showcase work was intentionally reverted when it was not ready to remain on the main branch, preserving the public surface as a more conservative docs-first entry point.

Scheduler And Control-Plane Reliability

Scheduler work tightened reset policy, repeated-monitor detection, per-agent frontier hints, wait cadence, user-gate precision, and canary coverage across status, quota, review packet, event, and read-path behavior. These changes make long-running agent loops less likely to wait forever, clear the wrong gate, or spend compute without a verifiable transition.

Validation And Public Boundary

This window added or refined focused smokes around benchmark boundaries, status/quota behavior, host command contracts, task graph projection, cold-path detail, and integrated control-plane canaries. Public docs continue to exclude private local state, raw benchmark logs, trajectories, private URLs, and credentials.

Next Window

The next window should turn the new surfaces into a repeatable release rhythm: keep issue-fix and benchmark workflows small enough to validate, publish two-week update notes through a reviewable PR, and continue moving canonical todo/history toward append-only event records without making the operator surface heavier.