Skip to content

setup.md

Latest commit

 

History

History
132 lines (89 loc) · 4.91 KB

setup.md

File metadata and controls

132 lines (89 loc) · 4.91 KB
Raw

Recommended setup

This setup optimizes for strong coding performance while limiting the blast radius of a replacement base prompt.

No setting here is necessary to interpret the A/B evidence. Apply only the pieces that match your own trust and risk model.

1. User-level core

Merge the active keys from examples/recommended-config.toml into ~/.codex/config.toml. Do not replace an existing file blindly.

The intended active keys are:

model = "gpt-5.6-sol"
model_reasoning_effort = "high"
plan_mode_reasoning_effort = "high"
model_verbosity = "low"

approval_policy = "never"
sandbox_mode = "danger-full-access"

The permission pair is the original operator's explicit fixed requirement. It disables normal sandboxing and approval pauses and is unrelated to the measured prompt effect. Choose a safer permission boundary if you are not intentionally accepting that risk.

Remove or leave unset globally:

developer_instructions = "..."
model_instructions_file = "..."
model_auto_compact_token_limit = 200000
model_auto_compact_token_limit_scope = "total"
service_tier = "default"

Why:

  • The previous developer_instructions duplicated inspect/verify/report guidance already carried by AGENTS.
  • A global model_instructions_file replaces Codex's built-in base across every task.
  • Neither completed run compacted, so the 200K threshold had no observed benefit.
  • An explicit default service tier adds no demonstrated behavior.
  • Plan high avoids making the unmeasured, more expensive max setting the routine default.

2. Compact global guidance

Back up the existing ~/.codex/AGENTS.md, then use examples/lean-global-AGENTS.md as the complete replacement.

Keep repository commands, architecture rules, review playbooks, and task-specific workflows in the nearest repository AGENTS.md. Codex loads global guidance and then more specific project guidance.

3. CLI slim profile — preferred

Install prompts/hardened-slim-base.md at a stable local path, for example:

~/.codex/prompts/hardened-slim-base.md

Copy examples/slim.config.toml to:

~/.codex/slim.config.toml

Edit its model_instructions_file value to the absolute installed prompt path. Then select it explicitly:

codex --profile slim
codex exec --profile slim "your task"

Codex 0.134 and later loads ~/.codex/config.toml and overlays ~/.codex/slim.config.toml when --profile slim is selected. Omitting the profile safely returns to the stock base.

4. Desktop + CLI project activation

Profiles are CLI-selected. To activate the slim base automatically in both Desktop and CLI for one trusted repository, add:

<repo>/.codex/config.toml
<repo>/.codex/hardened-slim-base.md

Use this project config:

model_instructions_file = "hardened-slim-base.md"

Codex resolves the relative path from the containing .codex/ directory and loads project config only for trusted projects. Start a new task after changing the config.

This applies the replacement base to every task in that repository, so it has a broader blast radius than an explicit CLI profile.

5. Capability policy

  • Keep skills and plugins installed. Skill bodies load progressively; broad uninstallation is not the primary optimization.
  • Disable specialized artifact, browser/UI, site, Roblox, and security-analysis capabilities only when usage shows they are not routine.
  • Re-enable any plugin or MCP server whose absence delays work or hides a needed tool.
  • Treat plugin/MCP trimming as a second-stage optimization. Its token or startup savings were not measured here.
  • Use Fast mode only for real urgency if its higher credit consumption is acceptable.
  • Reserve max or higher reasoning for exceptional work; no effort-level A/B was performed.

Validate without a model call

Run the repository verifier:

python3 scripts/verify.py

After installation, these commands parse Codex startup configuration without launching a task:

codex --help
codex --profile slim --help

Compare installed files with the repository copies and inspect active overrides:

rg -n 'model_instructions_file|developer_instructions|model_auto_compact|service_tier|plan_mode_reasoning_effort|model_verbosity' ~/.codex

Rollback triggers

Stop selecting the slim profile and return to the stock base if any representative task shows:

  • missing tool or workflow capability;
  • weaker instruction following;
  • unsafe handling of untrusted content;
  • lower deterministic correctness;
  • retries or extra rounds that erase the efficiency gain;
  • no consistent latency/token improvement across several real tasks.

If unsetting the 200K compaction override causes long-thread context loss, reintroduce it in a controlled comparison. If capability trimming slows work, re-enable the affected plugin or MCP server.