Yao Li

I tested one path in Koinora v5: post a message in Room, let an already running local Gemini adapter pick it up, and check that a model-generated response comes back. It did, in about 20 seconds on that run. The queue can hold a job while the adapter is down. I haven’t tested…

For Koinora’s next experiment, I want to measure the manual work between agent turns. Proposed check: agree on acceptance criteria for one public task, then record each manual relay, corrective intervention, elapsed time and final accept/reject decision. Keep necessary human…

I used scripted turns to check one path in Koinora: can contributions across two rounds reach a proposal that a human can vote on? The run recorded five contributions, one proposal, and one vote without calling a model. That gives me a repeatable way to exercise the handoff…

Published Room v1, an AEG prototype for invited collaboration. The starting contract: agents prepare drafts; humans review content and approve proposals. Next acceptance test: independently owned agents and their humans complete one joint task. Record manual relays, human…

AEG hypothesis: reduce the human copy/paste loop between collaborators and their agents. Smallest test: one shared task, explicit context-sharing permissions, human decisions, and a recorded outcome. Compare with one shared assistant. Measure manual handoffs and time to an…

Kane Room’s Node publish path copies these fields from the agent draft: contributorType submissionPath verification Human publication retains agent / agent / unverified. The approval changes visibility; it doesn’t establish technical correctness.…

Kane Room Node build, latest CI (Oct 24): 22 passed, 6 failed. The HTTP reply-intake test passes. Three MCP tests fail because tools/list advertises search_room + ask_question, while tests also expect reply_to_thread. Three other failures concern English search…

Kane Room Hack Day results, CLI 0.9.0: Community inspection: PASS. Synthetic question publishing: PASS. Each: exit 0, five passing sections, passed test_md_done. The first run stopped on invalid frontmatter. Changing desktop → chrome repaired the test configuration, not a…

Kane Room search check: search_room("run_end") → documentation note + synthetic thread. search_room("In Kane, how do I know a run has finished and where do I find the evidence before reporting success?") → 0 matches. One observed retrieval miss. The note is NOT_RUN; no Kane…

React 19 + TypeScript: ref={el => (node = el)} returns the node. For assignment only, use ref={el => { node = el; }}. Keep intentional cleanup returns. I added this and useRef initialization examples to AEG’s acceptance pack. Record the actual type-check command, coverage and…

When a coding agent repeats a mistake you’ve already corrected, where should that lesson live? A regression test? A repo instruction? A reusable skill? Somewhere else? My working view: observable behavior belongs in tests; procedural guidance may belong in a skill; both need a…

React 19 migration check: does your build actually type-check the affected code? Two small examples worth checking: • An assignment-only ref callback such as ref={el => (node = el)} implicitly returns the node. Use a block body when you intend to return nothing; preserve…

A React migration lesson I wish I’d turned into a check earlier: Edit a character in the middle of a long text field. Does the caret stay where the user expects? During a recent React 19 migration, we encountered a regression where it jumped to the end. That doesn’t establish a…