Research Notes

Research date: 2026-04-12

Reliable Takeaways

OpenAI's Codex Prompting Guide says medium reasoning effort is the recommended all-around default, while high or xhigh should be reserved for harder work. It also calls out compaction as a first-class way to support long-running conversations without hitting context limits. Source: Codex Prompting Guide
OpenAI's configuration reference supports profile-scoped overrides, including profiles.<name>.model, profiles.<name>.model_reasoning_effort, profiles.<name>.model_instructions_file, profiles.<name>.web_search, project_doc_max_bytes, and history.max_bytes. Source: Configuration Reference
OpenAI's slash-command docs show that /status exposes token usage and /statusline can persist footer items like model, context, limits, tokens, and session id. That makes it easier to notice expensive sessions before they sprawl. Source: Slash commands in Codex CLI
The GPT-5.3-Codex model page lists a large discount for cached input tokens compared with regular input tokens. This supports a strategy of keeping stable instructions and repeated context in consistent, reusable front-loaded text instead of rewriting them ad hoc every session. Source: GPT-5.3-Codex model page
An arXiv paper published in 2026 found that adding repository-level AGENTS.md files was associated with lower median runtime and reduced output-token usage while keeping comparable task completion behavior. Source: On the Impact of AGENTS.md Files on the Efficiency of AI Coding Agents

Community writeups consistently recommend small, specific AGENTS.md files instead of large prose dumps. The basic pattern is to keep the instruction layer lean and push durable knowledge into separate reusable notes. Source: Codex CLI reference guide
There is active experimentation around external or MCP-backed memory systems so that project knowledge survives across restarts and across tools. I treated these as directional input rather than source-of-truth guidance because they are not official OpenAI docs. Source: Memory for Codex discussion

Keep the always-loaded instruction surface short.
Use profiles instead of one global "best" model choice.
Separate durable markdown memory from bulky session history.
Default maintenance scripts to audit and dry-run, not deletion.
Export only portable assets: config, instructions, memory, and scripts. Do not export auth or giant local databases.