# Typed-message commit comparison

App base: [d5f8a7573a](https://github.com/pingdotgg/t3code/commit/d5f8a7573a3817d8c20f7491e7672443355c4083).
App head: [3f2df04510](https://github.com/pingdotgg/t3code/commit/3f2df0451040ad275d260b30bba2977db701ee86).

One leased Windows Server 2025 Datacenter machine (10.0.26100), 4 logical CPUs, 16 GB RAM; Defender antivirus and real-time protection enabled. Both apps ran with `vp run dev`, fresh isolated T3 homes, seeding disabled (`T3CODE_DEV_SEED=0`), and trace level Trace. Chromium headless 156.0.8078.4, 1920×1080, dark theme. No paid provider: the thread's history came from Replay (dev), showcase.

The fixture was a local clone of T3 Code at [ca6e09c3e1](https://github.com/pingdotgg/t3code/commit/ca6e09c3e1191c8d46a9b97dd658ccb65dde8fcb), containing 32,168 tracked files and a 4,501,747-byte index. Set Git author to `Commit Proof <proof@example.invalid>`, remove its origin, and create branch `commit-proof`. Add it to each app as `Commit fixture`, then create a replay thread titled `Typed commit proof`.

Run the identical `measure.cjs` against base, reset the fixture to its initial seed commit, and run it against head. Six runs per side, in order. Before each run append `Commit proof line N` to README.md, CONTRIBUTING.md and LICENSE; reload the thread, open Git action options, wait for status refresh, choose Commit, type `proof: typed commit N`, and click Commit.

The driver's capturing pointerdown listener records the actual mouse press using the browser's Date.now. The WebSocket `action_finished` receipt records the end on the same Windows system clock. Playwright actionability waits are excluded. The subsequent status refresh is excluded. `driverDispatchToResultMs` is retained as diagnostic data, not the reported timing. Every run checks the resulting subject, changed files, clean worktree, tree hash, and full patch using Git after the timed action. Corresponding trees and patches matched for all six runs. There were preliminary interrupted/driver-debugging runs; none are included in the six-sample measurement series.

`analyze.py` selects each `runStackedAction.runAction` span within the recorded click/result window and walks its descendants. `gitProcesses` counts its `git.operation` process-execution spans, excluding subsequent refresh work even when that work shares a trace ID. Public trace excerpts retain span names, IDs, duration and outcome, with paths and other attributes omitted. `prepareCommitContext` and `GitVcsDriver.commit.addAll` counts come from the same descendants.

`comparison.json` holds all twelve observations, behavior checks and trace counts. `report.py` renders the measurement panels and chart from that file; those PNGs are measurement reports, not app screenshots. The recordings show the actual app. The short clips show run 2, without changing playback speed; full recordings include all six interactions and intervening reloads.

Results: base median 4,281 ms (range 4,082–4,989; spread 907); head median 3,054.5 ms (range 2,895–3,601; spread 706). Improvement 1,226.5 ms, 28.6%, greater than each side's full observed spread. Process-execution spans decreased 19→15, staging passes 2→1, temporary preparation passes 1→0 in every run. The reported 1.5-second absolute target was not reached.

The driver expects its synthetic fixture, isolated app homes and browser storage in its scratch root. Its only user-home lookup locates the preinstalled Playwright package; no credentials are included. Start each app and pair its browser before saving the storage state; never upload that state.
