§ 1 — WORK

From symptom to decision

Every entry reads symptom → cause → decision → result, with links you can check.

W01 · 2026.01 – 2026.03

MIT — Meeting-based organizational knowledge

SYMPTOM: Meeting notes drift from reality fast. We needed an agent that tracks meetings and decisions, but an agent that freely invites members or deletes meetings makes changes that are hard to undo.
CAUSE: A planner/executor split with an evaluator gate was slow and bloated the context. Autonomy came less from the model than from how tools were exposed and when the agent re-decides.
DECISION: Moved to ReAct and put a human-approval (HITL) gate only on mutating tool calls. Replaced the per-turn time lookup tool with the current time injected into the system prompt. Proposed a shared domain glossary for the team.
RESULT: Owned the agent part through the final demo, then evaluated prompts quantitatively with LLM-as-a-Judge after the project.

W02 · 2026.07

SYMPTOM: With the same setup, Codex subagent tools appeared in some sessions and vanished in others.
CAUSE: Two processes on the same binary received different model catalogs because of their originator, and kept overwriting one shared cache file. The cache key lacked the request's identity.
DECISION: Went past the symptom report (#33559) and narrowed the cause into its own issue (#33593), tracing the seven PRs that produced the state.
RESULT: An issue with repro, cause, and fix direction, plus a devlog write-up.

W03 · 2023.03 – now

SYMPTOM: I couldn't really know what a Node package manager does without building one.
CAUSE: Implementing install, resolution, and script running in Rust taught me that behavior drifts as changes pile up unless it is pinned down in a spec first.
DECISION: Define behavior such as the workspace lockfile in spec documents first, and use the repo to experiment with validation gates and agent-assisted development.
RESULT: My longest-running personal project, since 2023.

W04 · 2022 – 2026

SYMPTOM: Code that turned WebSocket messages into React state was scattered across components.
CAUSE: Mixing framework-agnostic message handling with React bindings in one package makes both hard to change.
DECISION: Split a core (socket-store) from a React adapter (react-socket-store), wired subscriptions through useSyncExternalStore, and documented version compatibility.
RESULT: Two libraries published on npm, with a docs site.

OPEN SOURCE

EXPERIMENTS