feat(codex): experimental GPT-6 setup and profile-aligned evals

[teamchong]

Summary

• Reuse request-local factsheet preparation between history gates and emission without changing outgoing bytes.

• Default GPT-6 to original image detail.

• Deduplicate native exact-spelling factsheets within a request, preserving repeated-count hints and opaque-value excerpts.

• Add a GPT-6 rendering/history profile and cache-aware savings accounting for Responses requests.

• Improve history grouping, opaque-identifier preservation, and usage handling for interrupted streams.

• Align model-quality evaluations with runtime profiles and shared source fixtures.

Evaluation

Historical high-detail baseline: GPT-6 Astra, profile-aligned rendering, reasoning low. These are not results for the current original-detail profile:

Arithmetic	Gist	State	Never-stated confabulations	Hex association
98/100 (high baseline)	25/98	6/18	0/16	5/15

Methodology and results.

Verification

• node scripts/build.mjs: passed.
• pnpm test: 1,267 tests passed.
• pnpm typecheck: passed.
• node --test eval/model-quality/*.test.mjs: 25 tests passed.

Matched hex pilot (N=2 per variant)

Variant	Correct	Input tokens	Output tokens	Mean response	Estimated billed token-equivalents
native	2/2	8,614	21	1.730s	10,871.0
high	2/2	2,656	443	6.395s	5,533.5
original	2/2	2,656	124	2.622s	3,938.5
compact-native	2/2	6,380	21	2.114s	8,078.5

Adds shared profile-controlled imageDetail and a lossless compact-JSONL research representation. Pilot results.

Production-path replay (original detail)

Three public coding/state fixtures, five native/production pairs including two
repeated-prefix pairs. Same questions, model, low reasoning, and 768-token output
cap. Production uses transformOpenAIResponses, including completed custom-tool
history and native factsheet deduplication.

Arm	Correct	Median end-to-end	Reported input + output	Estimated billed token-equivalents
native	24/24	2.078s	35,726	27,198.70
production	16/24	6.942s	21,167	20,068.15

Estimated cost fell 26%, but accuracy and latency regressed. This does not meet
the cost/accuracy/latency objective. Token-equivalents use cached-read 0.1,
cache-write 1.25, and output 5 relative to uncached input, not provider invoices.
This is a history-recall replay, not an autonomous coding-agent benchmark.

A guide painted inside every PNG was tested on the first case without changing
image dimensions or input-token count:

First case	Correct	End-to-end	Input	Output
native	6/6	1.979s	4,360	29
production	3/6	2.781s	2,999	69
per-image guide candidate	5/6	6.727s	2,999	186

The candidate still missed a stated package decision and increased latency.
Validation stopped after this case. The guide is not included in runtime code.

Local preprocessing

Request-local preparation reuses extraction, not coverage-dependent output.
The cache is bounded to 64 entries and 1,048,576 source characters.

Image cache	Baseline median	Prepared factsheets
warm	1,068.145ms	685.695ms
cold	3,492.155ms	3,110.305ms

Outgoing request bytes were identical across both arms of the 96-round offline
replay. No model calls were made. This reduces local processing time, not API
billing. Measurements.