What is fixed per session
When a session is created, paddock composes its session settings from the agent’s effective configuration at that moment: the system prompt, the environment, the permissions, and Claude Code’s small model. They never change for that session. The harnesses read these values only when a process or conversation starts, so fixing them per session means every restart and resume of a session runs on the same values, and no change ever needs a harness restart. Create a new session to pick up new values.
A session’s working directory (
cwd) is fixed at creation too; see Agents and sessions.
System prompt
The system prompt is text appended to the harness’s own system prompt in every turn; it is never shown as a message. It comes from three places:- The template’s
system_prompt. - The agent’s own, which replaces the template’s when set.
- The
systemPromptgiven tosessions.create, appended after a blank line.
systemPrompt (null when all are empty), at most 64 KiB.
get, set, and delete each return the effective value with its source, template or agent.
Environment
On Claude Code, environment variables reach the harness and every command it runs. On Codex, they reach every command Codex runs but not the Codex process itself.- Names match
^[A-Za-z_][A-Za-z0-9_]*$, up to 128 bytes; values are UTF-8 up to 16 KiB. - These names are reserved, because paddock sets them for the harness:
PADDOCK_*,ANTHROPIC_API_KEY,ANTHROPIC_AUTH_TOKEN,ANTHROPIC_BASE_URL,OPENAI_API_KEY,OPENAI_BASE_URL,CLAUDE_CONFIG_DIR,CODEX_HOME,HOME, andPATH. Using one is400. - The collection rules of Capabilities apply:
addof a template’s name creates the agent’s override, anddeleterestores the template’s value.
Permissions
Permissions areallow and deny rules in the syntax of the agent’s harness. paddock passes them to the harness as they are and does not interpret them: on Claude Code they become the allow and deny lists of its settings, such as Bash(npm run test:*).
set replaces the agent’s own permissions and returns the effective ones: the effective deny is the template’s and the agent’s together, and the effective allow is the agent’s once it has set one, otherwise the template’s.
Model and small model
The model is the one setting of this page that reaches running sessions: a change is applied between turns, so a turn always finishes on the model it started with. Claude Code’s small model, used for background requests and for subagents that ask for a smaller model, is a session setting.model field is always the effective model. delete() restores the template’s value. Which names you can use is on Models.