Steve

Steve is a local coding agent forked from OpenAI Codex. It retains the upstream terminal UI, tools and harness while separating its public command and automatic state discovery from Codex. Internal Rust crate names remain upstream-compatible.

Development status

Native packages are available as GitHub Actions artifacts for Linux, macOS and Windows on x86-64 and ARM64. See installing a binary package for downloads, checksum verification and installation. Installers, npm packages and Python packages are not yet published; upstream Codex installers do not install Steve. The linked network inventory describes the original pinned baseline, not this branch.

To build from source, run this from the repository root. It creates a local package containing the CLI, code-mode host, executor, and platform sandbox helpers:

STEVE_REPO_ROOT="$PWD" CARGO_BUILD_JOBS=8 \
  python3 scripts/build_codex_package.py \
  --target x86_64-unknown-linux-gnu --package-dir dist/steve
./dist/steve/bin/steve

The example targets x86-64 Linux with glibc, including Ubuntu. On another platform, pass its Rust target triple (see --help). Install Rust through rustup, Python 3.11 or newer, and the native build dependencies first; Ubuntu needs build-essential, pkg-config, libcap-dev, and libasound2-dev.

The code-mode host requires sandbox-enabled V8. Plain Cargo builds try Deno's default archive location, which does not provide the required sandbox pair. The package builder instead downloads the version-pinned archive and matching Rust bindings from openai/codex releases, verifies their checksum manifest against this checkout, then verifies both files. These are build inputs, not an inference service. This remains an OpenAI-published binary dependency. To build V8 from source instead, set V8_FROM_SOURCE=1; that uses rusty_v8's separate native toolchain and source-download path and takes substantially more time and disk. See V8 build provenance.

The Steve binary build workflow runs after updates to main and can also be started manually for an exact candidate commit. Each job checks the linked V8 sandbox and extracted package, then uploads an archive, SHA-256 checksum and build provenance. All six platforms passed for candidate c5ba040781 in the initial successful package run. These checks cover package construction and startup, not a complete interactive session or the full test suite. Packages are not release-signed or notarized.

Separate configuration

Steve defaults to ~/.steve; STEVE_HOME selects another home. Project configuration is discovered under .steve. It does not automatically import Codex's home or project configuration.

The recommended Cambridge configuration uses Kimi for coding, blocking Guardian review and baseline risk classification, with D3 alongside for decision comparison. Put this in ~/.steve/config.toml:

model = "moonshotai/Kimi-K3"
model_provider = "cambridge"
web_search = "disabled"
check_for_update_on_startup = false
approval_policy = "on-request"
approvals_reviewer = "auto_review"

[auto_review]
model_provider = "cambridge"
model = "moonshotai/Kimi-K3"

[features]
guardian_approval = true
guardianv2_decisions_comparison = true

[features.guardianv2]
enabled = true
model_provider = "cambridge"
model = "moonshotai/Kimi-K3"
comparison_model_provider = "cambridge"
comparison_model = "vllm-sr/d3"

[model_providers.cambridge]
name = "Cambridge HPC"
default_model = "moonshotai/Kimi-K3"
base_url = "https://llm.hpc.cam.ac.uk/v1"
env_key = "CAMBRIDGE_LLM_API_KEY"
wire_api = "responses"

[analytics]
enabled = false

[feedback]
enabled = false

Supply CAMBRIDGE_LLM_API_KEY in the process environment. Keep credentials out of this repository. D3 uses the configured provider's /decisions endpoint; its comparison result does not approve actions or replace the baseline classifier. The tests establish protocol compatibility, not comparative safety quality.

OpenAI account login and managed account-token renewal are removed; supplied API keys and configured authentication commands remain available. Every inference provider needs an explicit base_url, including OpenAI providers. Set model_provider and either model or that provider's default_model. Catalogue ordering never selects a model.

Model discovery polls the explicitly configured providers at models, or their full model_catalog_url when supplied. It accepts Codex-native catalogs and OpenAI-compatible data/id lists. steve models and the app server's model/list expose qualified choices such as cambridge/moonshotai/Kimi-K3, explicit global/provider defaults, and provider discovery errors. Provider defaults come from default_model, never from list order. Pass a qualified choice with -m or a model-only API request. Explicit provider/model request pairs use the native model ID; saved configuration retains that pair. Public thread model selections, including settings acknowledgements, use the qualified choice.

model_catalog_json supplies metadata for live IDs, including context limits and instructions missing from discovery. Metadata slugs must use the canonical provider/native-model-id spelling. A local-only entry does not make a model available. Metadata can mark an otherwise unclassified model with "non_chat": true; advertised non-chat modes are excluded from coding choices. features.api_key_model_discovery = false disables discovery instead of advertising a static catalog as live.

The native provider transport uses the Responses API. A provider exposing another protocol needs a compatible gateway; changing its URL or token alone does not translate Anthropic Messages or Chat Completions requests.

Selecting Cambridge and disabling analytics/feedback does not prove that every optional integration avoids OpenAI. See the baseline network inventory for destinations, outgoing data, controls and remaining verification.

Guardian configuration

Automatic review can use a separate entry in model_providers:

approval_policy = "on-request"
approvals_reviewer = "auto_review"

[auto_review]
model_provider = "local_reviewer"
model = "your-reviewer-model"

[model_providers.local_reviewer]
name = "Local reviewer"
base_url = "http://localhost:8000/v1"
wire_api = "responses"

Set the reviewer model explicitly. A user-supplied local model catalogue may also provide an auto_review_model_override; downloaded model metadata does not choose a reviewer. The selected provider supplies model capabilities and prompt metadata. Choosing a provider does not establish model or tool compatibility.

The asynchronous classifier has independent provider and model settings:

[features.guardianv2]
enabled = true
model_provider = "local_reviewer"
model = "your-classifier-model"

Its model must satisfy the classifier's output contract. The optional Decisions comparison requires both comparison_model_provider and comparison_model; its service uses the Decisions protocol rather than ordinary chat inference.

To disable Guardian completely, including reviewer/classifier prewarming, use:

[features]
guardian_approval = false

Use approvals_reviewer = "user" for manual approval. Managed settings requiring Guardian review are incompatible with turning it off and fail visibly.

Documentation and provenance

• TypeScript SDK
• Python SDK
• Upstream Codex documentation, including useful harness and MCP guidance; upstream command names may differ.

Steve retains the upstream Apache-2.0 license and NOTICE.