file the Codex GPT-6 prompt, which is not the GPT-5 one with a new name

[fengyat]

Orca-Code-Review — push 1

Severity	Count
P0	0
P1	0
P2	0
P3	0

✅ no blocking findings

prompt/CODEX/ held two prompts, gpt-5.6-sol and gpt-5.6-luna, both opening You are Codex, an agent based on GPT-5. This adds a third, gpt-6-astra-system-prompt.md at 21,269 bytes, which opens based on GPT-6 and reads as a rewrite rather than a revision.

The structure moved. It leads with "When to ask the user for permission" and "Autonomy and persistence", two sections neither GPT-5 prompt has, placed ahead of "Personality" — which is where both existing files start. sol carries its autonomy guidance far down, under "Rules for getting work done", and adds "File editing constraints" and "Destructive Actions" sections that this prompt drops entirely, folding the permission question into the opening instead. New on this side: "Apps (Connectors)", "Plugins", and a "Writing PR descriptions" subsection.

The tool surface moved with it. The prompt instructs batching through functions.exec with await Promise.allSettled([...]) and asks for clarification through functions.request_user_input_async — a JavaScript-shaped exec rather than the plain exec the two GPT-5 captures were recorded against.

Scrubbed, and checked rather than assumed: the skills list carries {{CODEX_HOME}} placeholders, and the file has no home directory, username, git identity, email, uuid, or long hex run left in it. LF endings and a trailing newline; .prettierignore excludes *.md, so fmt:check has nothing to say about it.

One gap, stated rather than papered over. The "What is here" table in capture/README.md gets no row here. Three of its columns — prompt as sent, tool count, prefix tokens — come from a capture's meta.json, and there is no capture folder behind this file. The README is explicit that the file on disk runs a few hundred characters shorter than the prompt as sent, so the byte count above is not that column's number, and writing it there would be wrong. The row goes in when the capture is rerun with node capture/capture.mjs codex --model gpt-6-astra.

Nothing reads prompt/ at runtime — it is the browsable collection, written by capture.mjs. mirrorPrompts repairs mirrors that have a capture behind them and leaves files without one alone rather than deleting them, so this sits safely until the table row lands.

What is missing, found after the first push

The file describes a section it does not contain, so the gap needs no comparison to establish. Line 130: Any skills available to you in the current session will be listed in the "## Skills" section under "### Available skills". There is no ## Skills heading anywhere in the file — the last heading is ## How to use plugins at line 163. Two more lines lean on that same absent catalog: line 132 explains what each of its entries holds (name, description, and a location for the SKILL.md), and line 148 says the catalog also carries the mapping from short aliases such as r0 to their filesystem roots.

Everything else in the file checks out on its own terms. It opens on the identity line, the heading tree descends without a skipped level, no XML-ish tag is left unclosed, and the last line ends on a complete sentence at the final bullet of "How to use plugins" — nothing is cut mid-word, and 21,269 bytes is not a truncation boundary.

The two neighbours say the same thing from outside: gpt-5.6-sol and gpt-5.6-luna are byte-identical through line 168 (17,767 bytes each) and diverge only in the session-injected tail — <skills_instructions> wrapping that skills catalog, then <permissions instructions> with sandbox_mode and the approval policy. That tail is 5,627 bytes on sol and 3,088 on luna, and this file has none of it.

What I cannot tell without the capture is why. The GPT-6 harness may send those blocks as a separate message rather than appending them to the system prompt, or they may have been lost on the way to this file. Settling it needs the capture's request.json — the same rerun that fills the README table row. The file is worth having either way; it is the static template, not a full recorded prompt, and the PR should be read that way.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GnesN1z7KmCMggnbSc9joJ

[orcacode-review]

🐳 OrcaCode Review

✅ No findings — nothing to flag in this PR. Great work!

_{OrcaCode Review — Route Smarter. Ship Safer. Spend Less.
Engine-reported: 131 calls · 3.8M tokens · 95% cached}

❤️ Share · Install OrcaCode Review

Free on GitHub — the review runs on your own OrcaRouter key. If it helped, a shout-out goes a long way.

Share: X · Reddit · LinkedIn
Follow: X · Discord · LinkedIn · OrcaRouter