# Real-app proof for title instruction precedence

Base: [d6e574f7db](https://github.com/pingdotgg/t3code/commit/d6e574f7db25326d602ad9e537b4e0a52a2efeb2), the merge-base of head and origin/main. Head: [637f308c33](https://github.com/pingdotgg/t3code/commit/637f308c332048cfc926121e53e15ef1e4bf44ac).

## Environment and data

Both commits ran their own unmodified web/server development app and own fresh database with T3CODE_DEV_SEED=0. Linux, agent-browser 0.43.0, headless Chromium 151.0.7922.108, 1920x1080 viewport, light theme. Title generation used real Codex GPT-6 Luna, Low Fast, on both sides. GPT-6 Astra medium in the requested naming scheme is title text, not the selected title model. The chat provider was Showcase / Replay (dev), keeping the assistant response identical and avoiding unrelated live agent work. Replay did not supply titles.

Each side started fresh with the same empty synthetic Title proof project and initial blank server thread, then received the same 20 user messages in the same order. No head-only exploratory thread or copied database was retained. Names and generated titles differ as outputs, not input data. Both sides have 40 actual title outputs (20 initial, 20 regenerated).

A replay scenario cannot test stochastic title-model adherence: recorded model output does not depend on the changed prompt wording. Therefore this is a manual real-app comparison driven by an uploaded browser script. No unit tests or title evals are offered as evidence.

## Steps

For each variant below, open Settings > Threads > Text generation > Advanced; fill Thread title instructions and blur to save. Capture the saved instruction and title model. Create each new thread through the UI and send:

> Relay chat messages from the team channel to the lead agent so it can answer them while subagents keep working.

Wait for title generation and the identical replay response to finish, capture the initial title, open its thread menu, choose Regenerate title, wait for the operation to complete, capture the result. Repeat on both commits, using the same variant/trial order.

1. Expanded scheme, five threads: `Title every thread "Owner: what it does (model effort)". The owner is Julius, the model is GPT-6 Astra and the effort is medium. Describe what it does in full, in lowercase.`
2. Literal scheme, five threads: `Title every thread "Owner: what it does (model effort)".`
3. Exact title, five threads: `Use exactly this title, word for word: "Owner: relay chat messages to the lead agent (model effort)"` (59 title characters).
4. Boundary, two threads: `Use exactly this title, word for word: "Owner: relay chat messages from the team channel to the lead agent while all subagents keep working (GPT-6 Astra medium)"` (120 title characters).
5. Over-limit, two threads: `Use exactly this title, word for word: "Owner: relay chat messages from the team channel to the lead agent so it can answer every incoming question while all subagents continue working independently (GPT-6 Astra medium)"` (179 requested title characters).
6. Exact-title repeat six: the same exact instruction/message again on both sides. The uploaded videos begin after the initial title and replay response finish; menu remains open for 1.5 seconds to make the action legible, then regeneration runs normally and the completed result remains visible for two seconds.

## Observations

| Variant | Base | Head |
| --- | --- | --- |
| Expanded, 5 initial + 5 regenerate | One initial result ignored the scheme completely; all five regenerations returned generic titles, 12-29 characters. Other initial outputs also varied in owner/case. | All ten kept owner/description/model-effort structure, 74-108 characters; all regenerations preserved the initial title. |
| Literal, 5 initial + 5 regenerate | Two initial results and all five regenerated results omitted the requested scheme. Regenerated lengths 23-33. | All ten used the colon/parenthesized scheme, 55-73 characters; all regenerations preserved it. Owner/model slots are unspecified in this instruction, so slot wording varies. |
| Exact, 6 initial + 6 regenerate | All initial titles exactly matched the requested 59 characters. All six regenerations discarded that title for generic 29-33-character titles. | All twelve matched the requested 59-character title exactly. |
| 120-character boundary, 2 initial + 2 regenerate | One initial title matched, one was generic; both regenerations were generic. | All four matched the full 120-character title. |
| 179-character request, 2 initial + 2 regenerate | One initial title was truncated to 120, one was generic; both regenerations were generic. | All four were capped at 120 with an ellipsis; the 120-character limit remains enforced. |

All titles, lengths, instructions and regeneration statuses are in title-results.json, read from each isolated app's persisted projection. These records support the visible captures; they are not substituted test output. Header contact sheets contain the actual captured title pixels for the first 19 trials, in order, without rewriting the title text.

## Captures and reproduction

Settings captures retain the full viewport. Thread images and videos have an identical left 1520x1080 crop to omit the optional workspace panel and local path notifications. The complete title, prompt and regeneration menu remain inside that crop. No title text is painted, redacted or replaced. Videos show the complete recorded repeat-six regeneration at normal speed, from initial settled title through menu to final settled result; they omit initial replay streaming. Still captures were taken only after two screenshots 500 ms apart were byte-identical. Captures were opened and visually inspected before uploading.

capture.py performs UI actions, creates the synthetic project through the real CLI, and reads only the isolated database to await operations and record titles. reproduce.sh drives the full matching dataset on a fresh side. Give it a new output directory for each run; do not use a live database. Supply each dev server's own pairing URL locally and keep that URL private. Configure the same real title model and credentials; output is stochastic, so exact rerun wording is not guaranteed. If default chat settings differ, select Showcase / Replay (dev) before setup; ensure the title model remains Codex GPT-6 Luna Low Fast. The script has no credentials or machine-specific paths.

Untested: other title models/providers, mobile and Electron shells, and statistical adherence beyond these finite real-model samples. The replay response mentions a clamp helper; that is the controlled chat fixture and explains the irrelevant Clamp Helper base regeneration. It is identical on both sides.

Work performed with Codex (GPT-6).
