# Live Codex plan verification

Revision: 2e536497b1e4c8a8027a7eebb31ccb4fea9afa6e. Linux x86_64, T3 collaborative Chromium preview, 1280x800, normal DPR/font scale. Development web app with real server and Codex GPT-6.1-Sol High in legacy Plan mode. Isolated checkout and copied development state; no live user database was opened for writing.

Apply live-probes.patch only in a throwaway checkout. Run `vp install` and `vp run dev --no-seed`, pair the browser using the private startup URL, and open a thread with completed history. Enable legacy Plan mode in General settings. Ask for a detailed 20-step TypeScript offline chess tactics trainer plan, then submit this feedback:

> Refine the proposed plan into 25 detailed numbered steps, with three sentences per step. Add explicit acceptance criteria for every step. Do not use tools, inspect files, implement, or ask questions. Return only the revised proposed plan.

The probes wrap the existing entry/row incremental functions. `reused: false` means the function took its full derivation fallback. `reused: true` means it returned the incremental result. The organic plan probe does not record caller labels. It observes two entry calls per accepted plan text update; accepted update counts use distinct growing item lengths rather than treating those calls as separate provider updates. A later attributed fixture probe distinguishes ChatView and ThreadTimeline callers. Rows are measured before stable-row identity recovery. Compact exported records retain item type/id/text length, plan-artifact length, timestamp and reuse decision. They contain neither provider credentials nor unrelated data.

This probe holds input references and is for branch acceptance counts only. It is not a latency, CPU, memory or frame-time benchmark. Input event spacing is not derivation execution duration. Completion and other structural transitions are expected to take full fallbacks.