First Love Sol

Replay the GPT-5.6 Sol Day-1 Codex harness instruction profile with a pinned, hash-verified, and reversible setup.

This project does not restore an older model snapshot or model weights. It runs the current model with the Codex instructions that shipped in the pinned Day-1 catalog record.

简体中文

What is included

• the verbatim Day-1 instruction profile for gpt-5.6-sol;
• a PowerShell extractor pinned to the original OpenAI Codex commit;
• source and output SHA-256 verification;
• safe Windows configuration and rollback instructions;
• sanitized cache and compaction observations.

Provenance

Item	Value
Repository	openai/codex
Commit	3380969a29134630d56feb6218e8e8dcc5e8196d
Commit date	2026-07-09
Source file	codex-rs/models-manager/models.json
Model	gpt-5.6-sol
Fields	model_messages.instructions_template, base_instructions
Extracted SHA-256	E9778714D505F3DD04D44DB4394024C5FAB5BF6554FC9FAA3CDF9CF776B63BB9

At the pinned commit, the two instruction fields are byte-for-byte identical. See THIRD_PARTY_NOTICES.md.

Install

For a one-command, backup-first installation on Windows, completely exit Codex Desktop and run:

powershell -ExecutionPolicy Bypass -File .\scripts\install-first-love-sol.ps1

The installer verifies the bundled prompt, backs up an existing config.toml before changing it, writes only the top-level model_instructions_file, and reports every backup path. It refuses to edit while Codex Desktop is running.

To remove the managed setting and verified prompt later:

powershell -ExecutionPolicy Bypass -File .\scripts\install-first-love-sol.ps1 -Uninstall

Manual installation

Extract a verified copy directly from the pinned upstream source:

.\scripts\extract-day1-instructions.ps1 `
  -OutputPath "$env:USERPROFILE\.codex\sol-day1-2026-07-09.md"

Back up %USERPROFILE%\.codex\config.toml, then add this as a top-level key before any TOML table headers:

model_instructions_file = "C:/Users/<you>/.codex/sol-day1-2026-07-09.md"

model_instructions_file is the current documented Codex configuration key for replacing built-in model instructions. See the official OpenAI configuration reference.

Use forward slashes in a Windows TOML basic string. Completely exit Codex Desktop, including its tray process, restart it, and create a new task.

Rollback

Remove model_instructions_file from config.toml, completely restart Codex, and create a new task. Codex will return to the instruction profile supplied by its active model catalog.

Expected behavior

Changing the instruction file changes the prompt prefix, so the first request after switching is expected to have a cold cache. With the file held constant, subsequent requests in the validation task recovered to a 97-99% cached-input ratio. The tested automatic compaction retained a normal post-compaction cache profile rather than returning to zero.

These observations are version- and workload-specific. See docs/validation.md for the exact scope.

Compatibility boundary

• project and user AGENTS.md instructions still apply;
• server-side policy, tools, plugins, and app-server behavior are not frozen;
• future Codex versions may change or remove model_instructions_file behavior;
• an old instruction profile can become incompatible with future tool protocols;
• OpenAI does not endorse this project.

License

Apache-2.0. The bundled instruction text is derived from the Apache-2.0 OpenAI Codex repository. Attribution and pinned source details are in THIRD_PARTY_NOTICES.md.