Support per-project Fast Mode / speed-tier configuration

Summary

Please add a supported way to enable Fast Mode / speed-tier behavior for only selected repositories or projects, without changing the global Codex default for every workspace.

A concrete use case: keep the normal thorough/default Codex behavior globally, but allow a specific repository such as /home/thijs/Documents/werk_tue/aifaa/epv_possession_sequence to start in Fast Mode by default.

Current behavior

Codex currently supports project entries in ~/.codex/config.toml, but the project-specific config surface is limited to trust metadata. In the installed CLI source for codex-cli 0.125.0, ProjectConfig only contains trust_level:

pub struct ProjectConfig {
    pub trust_level: Option<TrustLevel>,
}

The active project is resolved from the current working directory or git root, but that project object is only used as project metadata. Model, reasoning effort, and verbosity are resolved from CLI/profile/global config instead:

let active_profile_name = config_profile_key
    .as_ref()
    .or(cfg.profile.as_ref())
    .cloned();

model_reasoning_effort: config_profile
    .model_reasoning_effort
    .or(cfg.model_reasoning_effort),
model_verbosity: config_profile.model_verbosity.or(cfg.model_verbosity),

There is a workaround using a repo-local .codex/config.toml with profile = "...", but that is not the same as a first-class per-project Fast Mode setting. It also means users need to model Fast Mode indirectly through model/reasoning/profile choices, and it does not expose the same product-level speed-tier behavior as the Fast Mode UI.

Desired behavior

Allow Fast Mode to be configured per project/repository. For example, one of these shapes would be enough:

[projects."/path/to/repo"]
trust_level = "trusted"
fast_mode = true

or:

[projects."/path/to/repo"]
trust_level = "trusted"
speed_tier = "fast"

or a project profile binding in the user config:

[projects."/path/to/repo"]
trust_level = "trusted"
profile = "fast"

The key requirement is that a user can opt one repo into Fast Mode while leaving all other repos on the normal default.

Impact

Users often have a small number of repositories where latency matters more than maximum depth, and other repositories where the opposite is true. A global Fast Mode toggle forces users to repeatedly switch modes or accept the wrong default for part of their work.

Per-project Fast Mode would make Codex defaults match the actual risk/latency profile of each codebase. It would also make the existing [projects] table more useful without requiring users to create shell aliases or remember --profile flags.

Acceptance criteria

• Codex supports a documented per-project way to opt into Fast Mode or the fast speed tier.
• The setting applies automatically when the session cwd is inside the configured project or its git worktree/root.
• The setting does not affect other repositories.
• CLI flags and explicit model/profile selections still take precedence over project defaults.
• The config schema rejects misspelled/unknown keys with a clear error, consistent with the existing config behavior.
• Documentation explains the relationship between global defaults, project defaults, profiles, and explicit CLI/TUI selections.

Environment / evidence

• Codex CLI version observed locally: codex-cli 0.125.0.
• Installed package metadata points to openai/codex, package @openai/codex.
• Relevant local source observations:
• codex-rs/core/src/config/mod.rs: ProjectConfig only contains trust_level.
• codex-rs/core/src/config/mod.rs: active project lookup is separate from profile/model resolution.
• codex-rs/core/src/config/profile.rs: profiles support model, sandbox, reasoning effort, verbosity, and related config, but project entries do not bind to those settings directly.