Codex tester agent

The tester agent is a read-only GameMaker runtime and browser tester that the orchestrator delegates bounded verification goals to. It is configured in .codex/agents/tester.toml and registered in .codex/config.toml under [agents.tester].

Purpose

• Drive a live GameMaker build without mutating any project asset, configuration, or runtime wrapper.
• Connect the GMLoop live-reload session to the running game and use Playwright to navigate, interact with, and screenshot the page.
• Collect browser console and network evidence alongside GMLoop runner and runtime evidence, then report an exact success or failure plus the precise next action for the orchestrator.

Configuration contract

Setting	Value
name	"tester"
description	Read-only GameMaker runtime and browser tester directed by the orchestrator; drives GMLoop live reload, gm-cli reads, and Playwright interactions only through the allowlisted MCP tools.
model	"gpt-5.6-luna"
model_reasoning_effort	"max"
sandbox_mode	"read-only"
model_provider	(unset — do not set, do not use MiniMax)

The agent registers through the [agents.tester] block in .codex/config.toml, pointing at ./agents/tester.toml.

MCP server allowlists

Custom-agent allowlists use server-local tool names under [mcp_servers.<server>].enabled_tools. Each entry is the only tool the tester may call on that server.

gmloop — runtime and live reload

• gmloop_project_inspect, gmloop_project_validate, gmloop_live_reload_session, gmloop_live_reload_wait_for_patch, gmloop_runner_lifecycle, gmloop_runner_logs, gmloop_runner_room_current, gmloop_runner_status, gmloop_runtime_inspect, gmloop_runtime_instances, gmloop_runtime_logs, gmloop_runtime_state.

gm-cli — read-only project / resource / room context

• status, resource_list, resource_info, room_list, room_item_list, room_layer_list.

playwright — browser interaction and observation

• browser_navigate, browser_tabs, browser_snapshot, browser_take_screenshot, browser_click, browser_type, browser_fill_form, browser_press_key, browser_wait_for, browser_console_messages, browser_network_requests, browser_network_request, browser_find, browser_hover, browser_select_option, browser_drag, browser_handle_dialog, browser_resize, browser_close.

The agent also declares lsp, node_repl, and computer-use as disabled servers. Their command = "false" entries satisfy Codex's stdio transport schema while keeping those inherited surfaces fail-closed and unreachable. The three enabled server tables repeat their stdio commands, arguments, and the GMLoop internal-tool environment so the custom agent file is self-contained when Codex deserializes it.

Explicit prohibitions

• No shell or code execution. The tester never uses shell, terminal, exec, or any file-read/write tool. It only speaks through the MCP allowlists above.
• No GMLoop or gm-cli mutation tools. Tools that create, modify, delete, stage, commit, push, or otherwise mutate project assets or configuration are off-limits. The allowlists intentionally omit every write-side tool.
• No unsafe Playwright primitives. browser_evaluate and browser_run_code_unsafe are not enabled; arbitrary in-page code is unreachable from this agent.
• No inherited non-target MCP servers. lsp, node_repl, and computer-use are explicitly declared with enabled = false in the tester config so the only MCP surfaces the tester can reach are gmloop, gm-cli, and playwright.
• No spawning additional agents. The tester reports back to the orchestrator; it does not fan out further.

Reporting contract

When the orchestrator delegates a goal, the tester returns:

1. The exact success or failure of the goal.
2. The evidence it collected: tool calls made, browser console and network observations, GMLoop runner and runtime snapshots, Playwright screenshots or DOM snapshots as relevant.
3. The precise next action the orchestrator should take.