YoraiLevi

Artifact-grounded review: a local-first coordination layer for humans and agents

Research report — verified 2026-09-09

1. Executive summary

The clearest category name is artifact-grounded review layer: software that keeps a discussion, work request, and its evidence attached to the particular content being reviewed, while presenting the same underlying record visually to a human and structurally to an agent. “Annotation system” is too narrow because it omits assignment and action; “agent task manager” is too broad and loses the anchor; “collaborative editor” implies owning the artifact; and “code review” excludes the document, slide, spreadsheet, PDF, and image cases.

Finding: no complete product was verified. Existing systems cluster into four strong but incomplete groups. Google Workspace, Microsoft Office, Figma, GitHub, Gerrit, and Frame.io provide excellent native human review inside their own artifact boundaries but are proprietary or remote, unevenly expose anchors and thread opera

YoraiLevi

Agent Identity Prior Art: Tools, Methods, and Mechanisms

An engineering survey of how autonomous coding agents are given an identity distinct from the human who runs them, with a focus on Git and GitHub. Written 2026-09-08.

Why This Document Exists

Coding agents (Claude Code, Codex CLI, Cursor, Copilot, and home-grown harnesses) run Git and platform commands on your behalf. By default they inherit your identity: your user.email on every commit, your token or SSH key on every push, your account on every pull-request comment. That default erases the human/agent boundary, so you cannot filter agent work in git log, apply different branch-protection rules to bot pull requests, or scope and revoke an agent's blast radius independently of your own access.

This survey catalogs the prior art: the tools people have built, the methods those tools use, and the underlying mechanisms. It closes with a user story for each method that states who benefits, what the method provides, and why that method is the r

YoraiLevi

Figma Agent Tooling: A Prior-Art Digest and Rating Framework

Survey window: 2026-08-24 to 2026-09-07 Method: GitHub REST search (gh api search/repositories), Figma developer documentation, figma/mcp-server-guide, and the figma/community-resources agent-skill index. Star counts are as observed on 2026-08-24 unless a later date is noted. They drift.

---

1. Why This Digest Exists

YoraiLevi

OMP Live Is Not Voice Mode

ChatGPT-flavored speech in the terminal, with a coding model that never speaks and a voice model that never touches the repo.

If you have pressed Ctrl+L in Oh My Pi and talked at the prompt, the illusion is complete. A voice answers. Code gets written. It feels like ChatGPT Advanced Voice grew a repo.

It did not. Live is a duplex call glued onto the session you already had. Two models. One transcript. A small labeled handoff in between.

The Thing You Think It Is

YoraiLevi

Evidence-based lecture design — deliverable guide

This package turns the supplied deep-research protocol into a report, a standalone toolkit, and five structured research tables.

Files

1. evidence-based-lectures-deep-research-report.md — research synthesis, methodology, evidence labels, mechanisms, context modifiers, contradictions, myths, accessibility requirements, evaluation, and references.
2. lecture-design-toolkit.md — planning canvas, 18 conditional rules, seven scenario blueprints, the ten requested supporting tools, eight annotated redesigns, two decision matrices, and a pilot/rehearsal process.
3. evidence-register.csv — compact 71-source register for easy scanning.
4. evidence-register-detailed.csv — the protocol’s field-complete source-table view. Where a synthesis does not report one uniform audience, duration, delay, or format, the cell says varied/not reported instead of inventing precision.

YoraiLevi

Running multiple Orca dev builds side by side (one per git worktree)

How to dogfood Orca while developing it: keep your stable release for real work, run several dev builds — one per worktree/branch — and let an agent session drive them without ever hitting the wrong one.

Upstream tracking: issue stablyai/orca#15647 · fix stablyai/orca#15648

---

TL;DR

YoraiLevi

Draft, then freeze

A manifesto for software that grows with the person using it.

We are not building an operating system. We are not installing another plugin and restarting. We are not asking a model to hallucinate HTML with nothing behind it.

We are building a growth landscape: visualization and interactivity that appear when needed, stay if they earn it, and remain visible to both the human and the agent — without hiccups, without restarts.

YoraiLevi

Trusting a Claude Code workspace without the interactive dialog

Runbook. Writes a workspace's trust record into Claude Code's global config so the CLI stops asking "do you trust this folder?" and starts honoring that project's .claude/settings.json.

Verified against Claude Code 2.1.234, native binary install, on Windows 11.

---

Verdict

YoraiLevi

Claude Code — Every Automatic File Cleanup, and How to Configure It

Scope: Claude Code CLI version 2.1.234.

Method: Derived by reading the shipped binary directly, not from documentation. See "How this was verified" to reproduce it on your own install.

Applies to: Local files on your machine only. Server-side retention is a separate policy and is not covered here.

---

YoraiLevi

Prompt-refinement brief — changelog and experiment log

The brief converts a terse, fragmentary user reply into an explicit self-contained prompt. It is tested by handing it to fresh subagents with real conversation turns and comparing their output against what was produced by hand.

x

Test cases