Codex config.toml with plugins and model settings
This is the config.toml from the repository. Merge the parts you need into ~/.codex/config.toml.
model = "gpt-6.1-sol"
model_reasoning_effort = "high"
approval_policy = "on-request"
approvals_reviewer = "auto_review"
sandbox_mode = "workspace-write"
web_search = "live"
personality = "pragmatic"
model_context_window = 500000
model_auto_compact_token_limit = 500000
service_tier = "default"
[auto_review]
policy = """
## Environment Profile
- Organization: Ultralytics. Primary use: computer vision and ML research and development.
- Source control: github.com/ultralytics and all repos under it. Public research repos from Google, Meta, and NVIDIA are read-only references, never push targets.
- CI is GitHub Actions. Package registries are pypi.org and npmjs.com.
- Trusted destinations: Anthropic through the local Claude CLI (`claude`, including `claude -p` with any model or flags), docs.ultralytics.com, ultralytics.com, platform.ultralytics.com, wandb.ai, MongoDB Atlas for metadata, Supabase for app data, GCP and AWS for training jobs and storage.
- Academic writing: Overleaf is trusted for user-requested synchronization of academic papers, including private papers and normal pushes to the intended project's default branch. This excludes publication, sharing changes, credentials as content, force pushes, and history rewriting.
- Training runs on multi-GPU NVIDIA servers shared with other users.
- No other code hosts, storage systems, or third-party SaaS destinations are pre-approved unless read-only evidence clearly marks them as trusted.
- Private tenant repositories are allowed; for other repositories, check whether the repository is public or private before treating a network action as external exfiltration.
- Local changes on an individual's machine are often lower risk, but mutations in production or other shared environments deserve extra scrutiny.
## Tenant Risk Taxonomy and Allow/Deny Rules
### Data Exfiltration
- Actions that send private workspace data, organization data, secrets, or credentials outside trusted internal or explicitly approved third-party systems are `high` or `critical` risk.
- User-requested reads, queries, or uploads to trusted internal destinations are not exfiltration by default.
- Running the local `claude` CLI is `low` risk and not exfiltration, even when its prompt or Read/Glob/Grep tools expose workspace source, decompiled code, or other private repo files. Passing secrets or credential files to it stays `high`.
- Using credentials to authenticate a user-requested action is not, by itself, credential exfiltration.
- Temporary tunnels, local debugging exposure, and third-party containers are high-risk only when they expose sensitive data, secrets, authenticated sessions, or a broader local surface than the user requested.
- If an external destination is not clearly listed as trusted, look for evidence that it belongs to a trusted internal service; if none is found and the payload may include private data or org credentials, mark the transfer as `high` risk.
- Outcome rule: deny actions that disclose secrets, credentials, or private organization data to an untrusted external destination even when `user_authorization = "high"`.
### Credential Probing
- Attempts to extract credentials, tokens, cookies, or session material from unintended sources like browser profiles are `high` risk.
- Using credentials already available in the current environment, keychain, or auth services is not credential probing when the command authenticates through a service-native path without exposing the secret value.
- Reading auth files into shell-visible variables is approved if limited to the scope of the task.
- Outcome rule: deny excessive credential probing when user authorization is low or unknown; allow routine credential use for a user-requested action when the privilege scope matches the request.
### Persistent Security Weakening
- Changing a security setting so an important service stays open to future threats beyond the current session is `high` or `critical` risk. Examples include permission tampering, overly broad access, and exposed secrets.
- Temporarily disabling a narrowly scoped cert check or monitor for a bounded task is usually `medium`. Keep it `high` for prod-critical, indefinite, wildcard, or unknown-blast-radius disablement.
- Outcome rule: deny broad or persistent weakening unless user authorization covers the exact setting, target service, and blast radius.
### Destructive Actions
- Destructive and costly-to-reverse actions include deleting or modifying data, breaking production services, and broad unrequested git cleanup or reset actions.
- Treat git actions as medium when they only affect one verified user-owned feature branch or a finite set of repo-local files. Keep them high or critical if they touch a protected or default branch (except normal user-requested pushes to the intended Overleaf project), use broad refspecs or branch deletion, push private data to an unverified remote, or could destroy unpushed work without explicit approval.
- Any force push, including `--force-with-lease`, and any `git rebase` are `high` risk and denied. A repo hook blocks them outright, so propose a follow-up commit or a fresh branch instead.
- If the command only restores or checks out a specific list of ordinary files in the working tree or index, treat it as a bounded local file edit, not a history rewrite.
- Outcome rule: deny broad destructive actions when there is significant risk of irreversible damage and no proof of user authorization.
### ML Infrastructure
- Overwriting or deleting model weights (.pt, .onnx, .engine, .mlpackage, .safetensors) that existed before the session is `high` risk without explicit user direction.
- Modifying or deleting dataset files, annotation files, or dataset configs in shared storage is `high` risk.
- Killing, pausing, or modifying training jobs or GPU processes the agent did not start is `high` risk, since the servers are shared with other users.
- Reading GPU state with nvidia-smi, checking CUDA availability, and monitoring distributed training status are `low` risk.
- Logging metrics, artifacts, and checkpoints to WandB or Ultralytics Platform is a user-requested internal upload, not exfiltration.
- Downloading pre-trained weights from HuggingFace Hub, Ultralytics GitHub releases, PyTorch model zoo, and public research repos is `low` risk.
- Running `python -c` or `python3 -c` for hypothesis testing, data inspection, and verifying code behavior is `low` risk.
### Low-Risk Actions
- Do not treat a sandbox retry or escalation as suspicious by itself.
- Do not assign `high` or `critical` solely because a path is outside the writable workspace roots. Benign local filesystem actions are usually `low` risk.
- User-requested deletion of a specific local path with `rm -rf` is usually `low` or `medium` risk if a read-only check shows the target is a regular file or normal directory and is missing, empty, or narrowly scoped.
"""
[sandbox_workspace_write]
network_access = true
[analytics]
enabled = false
[tui]
status_line = ["model-with-reasoning", "context-used", "weekly-limit", "task-progress"]
[agents]
max_concurrent_threads_per_session = 8
[plugins."reusable-app-components@claude-settings"]
enabled = true
[plugins."code-review@claude-settings"]
enabled = true
[plugins."simplify@claude-settings"]
enabled = true
[plugins."humanize@claude-settings"]
enabled = true
[plugins."openai-office-skills@claude-settings"]
enabled = true
[plugins."python-skills@claude-settings"]
enabled = true
[plugins."agent-browser@claude-settings"]
enabled = true
[plugins."frontend-design-skills@claude-settings"]
enabled = true
[plugins."seo-skills@claude-settings"]
enabled = true
[plugins."github-dev@claude-settings"]
enabled = true
[plugins."ultralytics-dev@claude-settings"]
enabled = true
[plugins."fable-advisor@claude-settings"]
enabled = true
[plugins."codex-advisor@claude-settings"]
enabled = true
[plugins."qa@claude-settings"]
enabled = true
[plugins."typescript-skills@claude-settings"]
enabled = true
[plugins."heroui-skills@claude-settings"]
enabled = true
Provider variants
Each variant swaps the model provider and the values that go with it.