Agent security · research · OKRs · maintenance

Chores

Proactive security hygiene for people using AI agents. Routine checks can rotate; every agent must immediately report suspected unauthorized activity directly to the human owner. The same evidence-linked status model can track personal Objectives and Key Results (OKRs) and scoped maintainer assistance, including private Linux kernel morning briefings. Our adaptation also supports research tasks, with two completed CPU Prefix Integrity Analysis case studies.

Current boundary Loading the published claim scope…

Adopt Chores in your own project

Self-maintenance with company-approved AI

Have the agent you already use investigate a maintenance problem, prepare and test a repair, and carry it to your selected completion point: a reviewed patch, a submission, or an eligible merge checked afterward. Reuse your project's tests and contribution process, with explicit action permissions and useful human handoffs.

Start here: request, guide and templates → Portable maintenance skill

These are adoption materials, not an installed autonomous runner. Model access, test execution and merge rights remain separate. Security and direct owner reporting remain the foundation.

Optional chore families

Four starting points

One event model; distinct operating and privacy rules.

Primary motivation

Agent security workflows

Rotate bounded routine checks and record their evidence. Suspected unauthorized activity bypasses the rotation and goes directly to the human owner; this dashboard is not that reporting channel.

Reusable extension

Personal OKR tracking

Treat each measurable key result as a workstream. Append progress, performer, reviewer, evidence, and the next review time without publishing a private goal set.

profile template example events

Maintainer workflow

Open-source maintainer assistance

Let agents triage assigned issues, draft pull-request reviews, coordinate bounded work, and prepare human handoffs. Public project content is untrusted input, and review scope is not merge authority.

Future direction: knlp R&D plans to release an open-source project that explores much more extensive AI maintenance.

private profile example events

Kernel maintainer workflow

Linux kernel morning briefing

Rank in-scope mailing-list patch series, syzbot reports, public CVE announcements, and lore discussions for a maintainer's morning review. Exact files and recursive directories bound the analysis; the maintainer retains every review and write decision.

private profile example briefing

A working adaptation

From a paper queue to tested cache contracts

A research use case within a broader task framework.

We have successfully used an adaptation of Chores for Prefix Integrity Analysis (PIA): checking whether a transformed KV-cache object remains safe to restore and share. The AI Post Transformers paper queue supplies candidate research. Scoped Chores tasks turn relevant papers into primary-source reviews and small, reproducible CPU experiments.

  1. DiscoverMatch a queued paper to an active question and retain its source revision.
  2. InspectRead the mechanism, define the saved object and check prior evidence.
  3. Test on CPUUse Cartridge geometry and a block manifest to check identity, selection and layout.
  4. ReviewKeep inputs and verdicts; identify any model or serving question that remains.

Finding · prefix identity

SnapStream

Query-dependent selection makes a prefix-only cache key unsafe for cross-query reuse. Adding query identity resolves the collision; compact layout, logical positions, padding and mutable ring state still require a method-aware connector.

Read the finding and sources →

Finding · backing history versus selected reads

HiSparse

The complete backing prefix passes the CPU contract model. Selected reads are request-time routing; saving only them under a prefix-only key fails the misuse control. That is a reuse mistake, not a defect in the complete-history design.

Read the finding and sources →

CPU-first testing makes it practical to screen more candidates without waiting for a GPU or loading a model. These two tests use metadata-only Cartridge fixtures; they establish the modeled contract. Quality and serving-performance claims require separate model work. The broader pipeline is documented in AI Auto Research.

This is a case study of our adapted workflow. The public dashboard below remains a status view and reference implementation; its implementation boundary is reported separately. Research tasks keep explicit scope, evidence and human review.

Project loading
State —
Updated —
Trace digest —

Implementation boundary

Workstreams

State and implementation coverage are reported separately.

Perfetto TrackEvent trace

Inspect the same activity on a time axis

The status document and trace come from the same profile and event stream. Perfetto shows each workstream on its own lane with state counters below the activity markers.

Open in Perfetto Download trace Status JSON

Published project record

Priority briefing and activity

waiting for status