Authoritative source, reproduced unchanged. Exact Markdown · Agent copy with resolved links. These instructions govern research runs, not checkpoint interviews.

Your platform: Codex

Read this once, at stage A1, after AGENT.md and OPENAI-RUNTIME.md. The method, the stage files, the tools and the checks are the same on both platforms; this file is where they differ. Generated by aggregator/handler/claude/gen_platform.py from aggregator/handler/claude/arm.codex.json; never edit it by hand.

Models and routes

aggregator/handler/claude/arm.codex.json sets each leg's model and effort; the controller (aggregator/handler/codex/run.py, aggregator/handler/codex/runtime/) records the native identities at every close. Never change the model to escape a failure.

stage	analyst	effort	readers
A1	gpt-6-astra	a medium; b medium; c high
B1	gpt-6-astra	a high; b high; c xhigh; d high; e xhigh
B2	gpt-6-astra	a medium; b high; c medium; d high	gpt-6-sol x3 via codex, effort medium
B3	gpt-6-astra	a medium; b high	gpt-6-sol x2 via codex, effort medium
B4	gpt-6-astra	a high; b medium; c high	gpt-6-sol x2 via codex, effort medium
C1	gpt-6-astra	a xhigh
C2	gpt-6-astra	a xhigh; b high; c high; d xhigh	gpt-6-astra x1 via codex, effort high
C3	gpt-6-astra	a high; b high; c xhigh
C4	gpt-6-astra	a high
C5	gpt-6-astra	a high; b medium; c high; d high	gpt-6-astra x1 via codex, effort high
D1	gpt-6-astra	a medium
D2	gpt-6-astra	a high	gpt-6-sol x2 via codex, effort medium
D3	gpt-6-astra	a high; b medium; c high
D4	gpt-6-astra	a xhigh
D5	gpt-6-astra	a high; b high
D6	gpt-6-astra	a high; b high
E1	gpt-6-astra	a high; b high; c medium; d high; e high	gpt-6-sol x1 via codex, effort medium
E2	gpt-6-astra	a high; b xhigh

The routine readers run on gpt-6-sol as independent native readers the controller owns, or on OpenRouter when the study selected that fleet arm (aggregator/handler/claude/fleet-arms.json, FLEET-ARMS.md). Wherever the stage files speak of fleet.py, read the controller's dispatch: it applies the same return validation (tools/fleet.py's validate).
The reviewers of C2 and C5 run on gpt-6-astra, each in a session of its own.
The labelers at D2 resume each room's original case lead in Codex (handoff/HANDLER.md).

Legs, closes and compaction

Each ordinary leg forks the preceding completed native checkpoint and its files. _head.md and the diary orient that continuation; they do not replace the history.
done.sh is your exit. The controller then runs the trusted checks and the recall probe, records the native boundary and commits the compressed checkpoint. Closing at packets-out asks the controller to dispatch; never run a fleet yourself.
The shell starts in out/: cd .. before running the tools. Scratch and cache paths are the controller's.
At a safe pause, log the working hypothesis and the next item with mem.py log --kind pause, then bash tools/done.sh STAGE --pause. The controller owns interruption and recovery.