Skip to content

Biweekly Update Note: 2026-06-28 to 2026-07-11

Source Boundary

This note summarizes public repository history from 2026-06-28 through 2026-07-11. It uses public commits, pull requests, releases, shipped docs, examples, and smoke tests. It excludes private operator state, raw benchmark evidence, private links, local paths, and credentials.

Highlights

  • Released LoopX v0.2.0, completing the peer-agent runtime cutover and defining claims as soft routing and writeback signals rather than runtime authority.
  • Extended issue-fix from patch generation into a maintainer loop with repository context, feasibility, CI and acceptance evidence, reviewer routing, PR lifecycle observation, and durable outcome writeback.
  • Made Explore and Explore Harness more useful for real long-running work: resumable experiments, focused graph exports, independent lanes, semantic topic affinity, and resource-aware portfolio planning.
  • Turned Auto Research into a visible evidence loop over a real workspace, with role-based rounds, automatic wakeups, summarized evidence writeback, and explicit promotion or continuation frontiers.
  • Reduced control-plane ambiguity through typed todo continuations, peer routing, compact quota payloads, selected-todo projection, and more reliable monitor, vision, and quota-spend transitions.

What Shipped

LoopX v0.2: Peer Agents, Not Runtime Hierarchies

v0.2.0 made the peer-agent model the public control-plane contract. A claim now answers who is expected to take the next step and write back the result; quota, gates, capabilities, exclusions, and write scope still decide whether execution is allowed. The release also replaced special reviewer-lane assumptions with typed continuation and independent-handoff contracts.

Representative changes:

  • #1787 cut the runtime over to peer agents.
  • #1773 added typed todo continuation policies.
  • #1867 preserved due-monitor routing across peer reassignment.
  • #1852 validated and shipped the v0.2.0 release.

Issue-Fix Became a Maintainer Loop

The issue-fix path now carries work beyond the first code change. It can bring repository context into planning, capture first-push CI and acceptance evidence, route a reviewer request with fallback, observe merge state, and write the outcome back into the domain lifecycle. Repository memory was also narrowed toward decision-useful context rather than undifferentiated history.

Representative changes:

  • #1780 added caller-repository context to issue-fix workflows.
  • #1835 added an optional OpenViking context provider without making it a control-plane dependency.
  • #1865 improved reviewer context, while #1876 captured first-push CI evidence.
  • #1883 projected merged-PR events so dependent work can resume.
  • #1887 made repository memory decision-useful for issue fixing.

Explore and Auto Research Became Resumable

Explore gained graph projections that preserve real topology while supporting focused owner views. Independent experiment lanes and semantic topic affinity survive projection, and portfolio planning can account for limited resources instead of ranking every branch as if capacity were unlimited. Explore Harness also became configurable and resumable rather than a one-shot experiment.

Auto Research moved from a scripted demonstration toward a visible multi-agent loop over a real KNN workspace. Fake metrics were removed; rounds now retain role ownership, summarized evidence, automatic wakeups, and an explicit next frontier.

Representative changes:

  • #1785 made Explore Harness configurable and resumable.
  • #1804 added focused Explore graph exports; #1859 preserved independent lanes; #1863 added resource-aware portfolio planning.
  • #1448 removed fake Auto Research metrics, and #1455 moved the KNN demo to a real benchmark workspace.
  • #1590 enabled automatic wakeups, while #1677 simplified the visible frontier projection.

Control-Plane State Became Easier to Read and Trust

This window tightened the state machine around todos, monitors, vision, quota, and scheduler decisions. Agent-facing hot paths became smaller without losing the selected todo or its lineage, monitor transitions stopped consuming quota before material writeback, and terminal todo/vision states became more explicit. Several large CLI modules were split along existing ownership boundaries rather than replaced with a parallel framework.

Representative changes:

  • #1694 compacted quota todo summaries, and #1731 exposed the selected todo directly.
  • #1764 preserved due monitors through quota compaction; #1769 canonicalized agent vision lifecycle states.
  • #1774 reduced the todo CLI module budget without changing the public state machine.
  • #1855 fixed quota spending after material monitor transitions.

Validation and Failure Attribution Improved

Full-public smoke sweeps became a regular release gate, with parallel shards and explicit handling for git-heavy checks. Benchmark and host paths also gained more fail-fast attribution for unavailable sandboxes, remote runtimes, setup dependencies, and missing score evidence. These checks do not replace real long-running validation, but they make a failed run easier to classify before another expensive attempt.

Representative changes:

  • #1727 repaired and parallelized the full-public smoke runner.
  • #1743 restored reliable full-public smoke coverage.
  • #1803 failed fast when the Codex sandbox could not execute.
  • The v0.2.0 release passed all six public smoke shards, a 539-check local public sweep, and a 17-surface premerge canary.

Validation And Public Boundary

The changes summarized here were validated through focused contract smokes, full-public sweeps, release checks, and risk-based premerge canaries. This note links only public repository evidence and does not copy private state, raw benchmark logs or trajectories, local paths, credentials, or operator-only material into the publication surface.

Next Window

The next expected window is 2026-07-12 to 2026-07-25. The publication workflow will continue to generate a deterministic, reviewable artifact from public history; a maintainer or LoopX agent remains responsible for narrative quality and the final pull request.