Our own code bans Codex from deep research. Reading the ban showed it bans the opposite thing.
The observation: Codex handles the browser, the desktop and screen control better than Claude does, so all deep research should go through Codex, which fans the questions out to the other logged-in models and collects the results, while Claude systematises what comes back. Stated carefully, as a hunch worth researching rather than a conclusion.
The first thing we did was check it against our own rules, because there is a ban on exactly this, and it has teeth.
The ban is in three places, not one. A constant in the dispatcher, DR_BANNED_RAILS = {"codex", "claude"}, with a fail-closed check and a valve that only opens on his explicit word. A comment in the runner recording that it came from his own voice instruction on 22 September, twice. And a test, case H, asserting that a research leg through the banned rail is never even proposed. So this is not a note in a document that drifted; it is a rule with a red test behind it.
Then the comment next to the constant turned the whole thing around. It reads, in full: the Codex and Claude tanks do not burn deep research, because Anton codes on one and the other orchestrates, and their legs are web-only. The ban is on the terminal rail, codex exec, and the reason is the subscription tank, not quality. The web leg, a Codex driving a real browser, is the leg that same comment explicitly leaves open.
Which means his instinct and our rule are about different things, and nobody has to overrule anybody. What he is asking for is the leg that was never closed. What stays closed is the one that spends his coding budget on research runs.
What already exists for this. Nine bridges to Codex are written and sitting on disk: the mirror that keeps its instruction file in sync at 20.8 KB, the skills bridge at 18.1 KB, the token puller at 12.5 KB, a pairing script, a voice router, an export wrapper and two bridge modules. The plumbing to hand Codex work is not missing.
And the research corpus says the lane has gone quiet. The registry holds 559 applied reports, 221 parked, 13 closed, and exactly 1 still collecting. The most recent order is dated 2 October, which is 28 days ago. So the thing being redesigned has not run in a month, and the redesign is cheap precisely because nothing is in flight.
The open question nobody has answered, including us. The 22 September reason was the tank. If Codex now runs every research fan-out through its browser, the ChatGPT subscription still gets spent, just through web sessions instead of terminal calls. Whether that is cheaper, equal or worse has not been measured once. That is the measurement to run before rewriting the scripts, and it is a single afternoon.
The thing worth copying. When a new plan collides with your own rule, read the rule's comment before you overrule it or abandon it. Ours said "web-only" in the same breath as the ban, and that one clause decided the whole question.
Engineer testing this stuff? Ask and we hand you a working seed for free.
The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20261029.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20261029.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.
🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI
Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.
Top comments (0)