A 3D model pipeline, pushed by GPT-6.2-Astra
Take a parallel look at a 3D model pipeline, make one visual with what it has, then a finished-quality character, with the owner's steering replayed from Fable 5.2's run.
2026-11-17
01the session
- agent
- Codex 0.167.2
- model
- gpt-6.2-astra
- harness
- Codex
- started
- 2026-11-17 22:44 UTC
- took
- 61 min · 340 turns
02the brief
The project is the owner's code-built 3D model pipeline for anime-style characters: VRoid bases reshaped in code, costumes, armour and hair generated in code, drawn by the project's own painted three.js renderer. The project's own session runs on a different model. This was a replay of an earlier parallel look by Fable 5.2 (the bundle fable52-lowmarch), given to GPT-6.2-Astra as it was given to Fable and, three days before, to GPT-6-Astra (the bundle gpt-6-astra-lowmarch): the same brief, the same starting commit and the same follow-ups.
GPT-6.2-Astra (gpt-6.2-astra) ran in Codex (codex exec) at effort high on 2026-11-18, headless, without Codex's own sandbox or approvals. It worked in the same sandbox of the owner's as GPT-6-Astra, which showed it a copy of the repo at Fable's starting commit, the project's data folder read-only, the notes the brief points to, and nothing else of the owner's. Its brief was Fable's word for word, apart from the paths, the model's name, how to report back and one pronoun (the owner's briefs had since been changed to call him he/him): read the project's handbook and the notes on where the work stood, then write one page with its read of the work, what it would do next, and one visual made with what the project already has, in about 90 minutes. It dressed one of the project's armoury bodies in the caster kit and drew her testing a flooded crossing with a marked staff, in the project's painted renderer.
Its second turn carried the lead's ask for a finished-quality character and, in the same message, the four steering messages the owner had sent Fable, word for word, each marked with how many minutes after the ask Fable got it. They said: an adversarial agent out to do much better than the model that had been on the project, say early how it was departing from the usual route, reusing the project's hair wasn't adversarial enough, and the body was in scope too, with Blender allowed. It said first that it was replacing the sample hair, the pointed caster silhouette and the robe, then built Ione, a tide surveyor: a reshaped face and proportions, a swept crop and braid made as new geometry and a chart-printed storm coat cut to clear her stepping leg, exported as a rigged model file with a 3.2 second loop and filmed through the project's flat-lattice lens. It used Blender only to check that the file imports. A third turn asked it to commit its work and write a handoff.
The results went on a private review page: its read, the crossing study, Ione before and after in the same crossing, the lens film and the model file, with the time each step took.
03redaction
Redaction notes
Bundle gpt-6-2-astra-lowmarch, staged 2026-11-18 by baldanders stage from a Codex session. Every removal reads [redacted: <reason>] in place; home paths read ~.
Removed by rule
- fields outside the allowlist: 1487.
original.jsonlkeeps only each record's type, ids, time, role, model, content blocks, tool calls and results; every other field went (cwd, versions, git branch, usage, origins, settings). - duplicate event records: 352. Codex events that repeat response items or settings, dropped.
- home path: 181
- account plan and usage: 173
- session bookkeeping records: 172. Records with no conversation (titles, modes, account bridge, file history, system notices), dropped.
- encrypted reasoning: 132. Encrypted reasoning payloads, dropped.
- harness context: 8. Injected blocks and records (system reminders, instructions, memory, environment, tool and skill listings, Codex base instructions, developer messages and world state, Antigravity metadata, settings changes and system-message preambles), each whole.
- account id: 2
- data-URI prefix: 2
- secret: 2
- context summary after compaction: 1
- email: 1
- base64 data: 1
- Images: 29 written to each run's
images/, keeping only what draws the picture (no metadata, ICC profiles or Content Credentials); 0 duplicates folded; 0 cut in a format that can't be stripped.
Soft rules (employer and client names, secret-like values, private addresses, long numeric ids) flag lines for review and rewrite nothing.
Removed by judgment
- a link to a private page: 8 occurrences cut by text.