Agent Configuration¶

Pick who runs the agent loop with one flag:

pyrualean play --runtime {codex,claude,openai} ...

Every runtime gets the same prompt, the same python (or run_program) and finish tools and the same budgets, and the episode is judged the same way. They differ in which product drives the conversation and what it adds: Codex keeps its own shell, sandboxed to the workspace; Claude Code its read-only Read, Glob and Grep; the openai loop adds nothing.

`--runtime`	What it is	When to pick it
`codex` (default)	The Codex CLI runs the loop; PyRUA-Lean serves it the tools.	GPT models, with a Codex login or an OpenAI Responses endpoint.
`claude`	Claude Code, through the Claude Agent SDK.	Claude models.
`openai`	A plain loop in PyRUA-Lean against any OpenAI-compatible chat-completions endpoint: a gateway, or a local SGLang or vLLM server.	Open models such as PyRUA-Lean Coder; RoboDojo.

--model names the model as the endpoint knows it. --reasoning sets the reasoning effort of Codex and Claude Code (default xhigh); with openai, put it in --openai-extra-body.

Codex¶

npm install -g @openai/codex@0.155.1

Without further options the code arm uses whatever provider the Codex home is logged in to.
--base-url https://host/v1 --api-key-env MY_KEY sends it to an endpoint instead (the key is read from the variable MY_KEY). Codex speaks the Responses API there; a server that serves only chat completions needs --runtime openai.
For a run nothing else may leak into, give it a clean Codex home with --codex-home: a directory with only the credentials and a config.toml that sets approval_policy = "never" and memories = false under [features]. --codex-bin should point at the real Codex binary, not a wrapper.
Long episodes: with gpt-6-astra Codex assumes a 272k-token window and compacts the conversation at a threshold derived from it. The paper's runs set model_auto_compact_token_limit = 945000 in the Codex home's config.toml (or CODEX_AUTO_COMPACT_TOKEN_LIMIT=945000), so no episode was compacted; result.json records the setting (host.codex_context) and any compaction (generation.compactions).

GPT-6 Astra on Codex 0.155.1

Codex 0.155.1 runs gpt-6-astra in "code mode", which hides every tool behind one JavaScript tool. PyRUA-Lean passes Codex a model catalog that turns this off (direct function calls, as in the paper's runs), so nothing needs to be done for play and generate.

Claude Code¶

pip install claude-agent-sdk          # into the robot stack's environment
pyrualean play --runtime claude --model <claude model> --claude-cli /path/to/claude ...

Claude Code reaches the model the way it is configured to (ANTHROPIC_* variables). Opus 5.5 needs Claude Code 2.1.280 or newer. --reasoning none turns thinking off.

OpenAI-compatible endpoints¶

pyrualean play --runtime openai --base-url http://127.0.0.1:30000/v1 --api-key-env MY_KEY \
    --model <model name> ...

Without --base-url and --api-key-env, the loop reads OPENAI_BASE_URL and OPENAI_API_KEY.

Option	Meaning (default)
`--openai-max-tokens`	the longest reply (`4096`)
`--openai-extra-body`	JSON merged into every request, e.g. `{"reasoning_effort": "high"}`
`--openai-nudges`	replies without a tool call tolerated before the episode ends (`2`)
`--openai-tool-choice`	`auto` or `required` (`auto`)
`--openai-temperature`, `--openai-timeout-s`	sampling temperature, request timeout (`900` s)
`--image-max-side`	downscale images to this longest side (`0`: keep)

A rate-limited endpoint (HTTP 429) is waited out, not counted as a failure.