lattice

<p align="center"> <img src="docs/assets/lattice-mark.png" alt="Lattice mark: a bounded repository slice ending in a verified result" width="128"> </p> <h1 align="center">Lattice</h1> <p align="center"><strong>An optimized execution system for repository-scale coding tasks.</strong></p> <p align="center"> Lattice moves deterministic repository work out of the model loop so strong coding models can focus on solving the task. </p>  One package includes both the **Codex** and **Claude Code** adapters. Live paired results on the same reset-token task, plain agent versus Lattice. These are separate single-task runs, not a model ranking: | Pair | Fresh input + output | Cost | End-to-end time | Acceptance | | --- | ---: | ---: | ---: | --- | | **Claude Opus 5.5** / high, v2.0.1 | 37,279 → 4,747 (**87.3% less**) | $0.362 → $0.048 (**86.8% less**) | 35.0 s → 10.3 s (**70.4% less**) | 4/4 both arms | | **GPT-6 Astra** / medium, v1.0.0 | 16,872 → 4,005 (**76.3% less**) | Not measured | 70.2 s → 27.0 s (**61.6% less**) | 4/4 both arms | > **Evidence boundary:** each row is one owner-run pair on one small, > maintainer-authored fixture (five source files), not independent task > selection or a general savings claim. The ratio will differ on larger > repositories and harder tasks. > > - **Opus 5.5:** both arms received the same task text, and the patches are > functionally equivalent. Most of the difference comes from Lattice finishing > in one model turn where Claude Code used eight. > [Record](docs/evidence/owner-run-claude-opus-5-5.md). > - **Astra:** measured with v1.0.0, whose Lattice arm also received five > hand-written acceptance criteria that the plain arm did not, so it is not > like-for-like. Arm time excludes setup, checkpoint and cleanup. > [Record](docs/evidence/owner-run-gpt-6-astra.md). > > Historical v1.0.0 records for > [Sol](docs/evidence/owner-run-gpt-5.6-sol.md), > [Luna](docs/evidence/owner-run-gpt-5.6-luna.md) and the community-run > [Opus 5](docs/evidence/community-run-claude-opus-5.md) pair remain published. [](https://github.com/moulwyse/lattice/actions/workflows/ci.yml) [](https://github.com/moulwyse/lattice/actions/workflows/quality.yml) [](LICENSE) [](package.json) [](https://github.com/moulwyse/lattice/releases) ## Install both integrations and verify in 30 seconds You need [Git](https://git-scm.com/downloads) and Node.js `20.19+` or `22.12+`. The local verification below needs no model account and makes no model call. Run it from the Git repository where you want to use Lattice. ```powershell # Windows (PowerShell): irm https://raw.githubusercontent.com/moulwyse/lattice/main/scripts/install.ps1 | iex ``` ```sh # Linux / macOS (Bash): curl -fsSL https://raw.githubusercontent.com/moulwyse/lattice/main/scripts/install.sh | bash ``` Or install the prebuilt release package: ```sh npm install --global https://github.com/moulwyse/lattice/releases/download/v2.1.0/lattice-v2-2.1.0.tgz lattice benchmark --worker mock ``` Or build from canonical source: ```sh git clone https://github.com/moulwyse/lattice.git cd lattice npm ci npm run build npm link lattice doctor --workspace . lattice benchmark --worker mock ``` Installing straight from Git (`github:moulwyse/lattice#v2.1.0`) fails on npm 11 and later: npm prepares a global Git dependency without its dev dependencies, so the TypeScript build cannot find `tsc`. A healthy verification ends with `Status: passed`. The unified package includes the Codex and Claude Code adapters; each integration remains opt-in so Lattice does not silently change either agent. Continue with the [Codex setup](docs/quick-start.md) or [Claude Code setup](docs/claude-code.md). Created and led by **[Moulwyse](https://github.com/moulwyse)**. This repository is the original and canonical home of Lattice. > **v2.1.0 scope:** review the [limitations](docs/limitations.md) and > [security model](SECURITY.md) before using Lattice on a sensitive repository. > Install the prebuilt release package or use the installer; `@moulwyse/lattice` is not yet available in the public npm registry. Transparent Codex > hooks stay experimental. > See the [release and upgrade guide](docs/release-v2.0.0.md). ## How it works Lattice indexes a local repository, selects bounded task-relevant context, coordinates an agent run, validates edits against repository fingerprints, and records local execution state. The optimization comes from moving deterministic repository operations outside the model loop while keeping context selection, edit authority, and verification visible.  ## Why Lattice - **Bounded context:** the agent receives explicit task-relevant pages instead of an unrestricted repository dump. - **Verified edits:** stale or out-of-scope patches are rejected against edit grants and repository fingerprints. - **Visible state:** local artifacts record what was selected, changed, and verified without hiding the execution path. - **Cross-provider:** one installation includes tested Codex and Claude Code paths with explicit stability labels, RAW bypasses, and separate evidence. Lattice was originally created and developed by Moulwyse. ## Unified installation: Codex + Claude Code You need [Git](https://git-scm.com/downloads) and a supported [Node.js](https://nodejs.org/en/download) version (`20.19+` or `22.12+`). You do not need an API key to install Lattice or run its local demo. ### Run the unified v2.1.0 release Install the prebuilt release package globally through npm on Windows, macOS, or Linux: ```sh npm install --global https://github.com/moulwyse/lattice/releases/download/v2.1.0/lattice-v2-2.1.0.tgz lattice --version lattice benchmark --worker mock ``` The package contains both provider adapters. Codex is the default worker, and the Claude Code integration is available alongside it. Running the package enables neither integration until you choose it for your environment or repository. The benchmark is local, deterministic, credential-free, and makes no model call. It is a functional smoke test, not evidence of general quality or token savings. ### Build from source #### Windows (PowerShell) Copy and run these commands: ```powershell git clone https://github.com/moulwyse/lattice.git Set-Location lattice npm ci npm run build npm link lattice --version lattice benchmark --worker mock ``` The final command is a credential-free self-test. A successful installation ends with a passed benchmark and creates no model charges. To connect Lattice to an already installed and authenticated Codex environment: ```powershell codex login status lattice integration codex doctor --workspace . lattice integration codex enable lattice integration codex status --workspace . ``` The Windows integration registers the Lattice MCP server and installs its Lattice-owned launcher and synchronization hooks. Restart Codex after enabling it. Lattice never asks you to paste a Codex API key into its configuration. #### macOS or Linux Install and run the same local self-test: ```sh git clone https://github.com/moulwyse/lattice.git cd lattice npm ci npm run build npm link lattice --version lattice benchmark --worker mock ``` On Linux (Bash, Zsh, Fish), enable the launcher, MCP bridge and hooks: ```sh codex login status lattice integration codex enable # Open a new terminal and restart Codex. lattice integration codex doctor ``` Linux setup includes shell PATH configuration and supports Arch/Omarchy. A native Linux lifecycle smoke test is included in CI; live Omarchy validation remains pending. For macOS or manual MCP setup, see [platform support](docs/installation.md#platform-support). If `npm link` is unavailable or requires global permissions, skip it and run the CLI from the cloned directory as `node dist/cli.js <command>`. ### Use it on a repository Open a terminal in the repository you want to work on and run: ```sh lattice doctor --workspace . ``` Resolve any reported error, then open that repository in Codex. On Windows and Linux, the enabled integration installs the launcher, MCP bridge and session synchronization hooks. Hook activation still depends on the Codex client. ### Undo the integration Windows: ```powershell lattice integration codex disable npm uninstall --global lattice-v2 # after npm link: npm unlink --global lattice-v2 ``` Linux automatic integration: ```sh lattice integration codex disable npm uninstall --global lattice-v2 # after npm link: npm unlink --global lattice-v2 ``` The disable command removes only integration state that Lattice recognizes as its own. Full installation, troubleshooting, and safety notes are in the [installation guide](docs/installation.md). ## What is included | Capability | Status | Notes | | --- | --- | --- | | Local repository discovery and index | Available | Respects repository boundaries and ignore rules. | | Bounded context pages and edit grants | Available | Local deterministic controls; covered by tests. | | Fingerprint-checked patch application | Available | Rejects stale or out-of-scope edits; a verified patch is applied to the workspace (`--no-apply` only verifies). | | Mock worker and deterministic fixture benchmark | Available | Runs without a model account or API credential. | | Manual handoff workflow | Available | The operator transfers a bounded request and response. | | Direct Codex SDK worker | Beta | Requires an authenticated Codex environment; exercised by published owner-run paired smoke tests. | | Transparent Codex launcher, hooks, sidecar, and MCP bridge | Experimental | Alters user-level integration state when explicitly enabled; inspect before use. | | Adaptive model selection and verified-patch cache | Experimental | Opt-in; exact behavior and limits are documented. | | Claude Code | [Available](docs/claude-code.md) | Included in the same package; locally tested and exercised by published Opus 5.5 and Opus 5 pairs. | | Gemini, Cursor, Grok, or other providers | Not implemented | No adapter for these providers is included in this repository. | “Available” describes implemented and locally tested behavior, not a production support guarantee. See [provider status](docs/providers.md) for the precise boundary. ## Claude Code Claude Code is an opt-in integration inside the main Lattice package. There is one package, one installation, and one CLI: `lattice`. Existing Codex commands and defaults remain unchanged. Install the same unified release through npm: ```sh npm install --global https://github.com/moulwyse/lattice/releases/download/v2.1.0/lattice-v2-2.1.0.tgz lattice --version ``` Enable it only in the repository where you want Claude Code to use Lattice: ```sh lattice integration claude enable --workspace . lattice integration claude status --workspace . lattice claude ``` Use `lattice claude --raw` to launch the same bundled Claude Code while bypassing Lattice for that child process. Undo only the project integration with `lattice integration claude disable --workspace .`; uninstall the unified package with `npm uninstall --global lattice-v2`. The integration has passed local build and contract tests with Claude Agent SDK `0.3.281` and its bundled Claude Code `2.1.281`. An owner-run Opus 5.5 pair on the public fixture observed 87.3% less fresh input plus output, 86.8% lower provider-reported cost, and 70.4% lower end-to-end time, with 4/4 pristine acceptance tests in both arms. This is one task-specific signal, not a universal Claude claim. Claude Code, Agent SDK, hook, and MCP behavior may change. Read the [install, RAW bypass, and removal guide](docs/claude-code.md) before enabling it. Maintainers and reviewers can use the concise [Claude Code OSS project brief](docs/claude-for-oss.md). ## Requirements - Node.js `^20.19.0` or `>=22.12.0`, matching the locked development toolchain; - Git for repository and worktree features; - Windows, macOS, or Linux with a filesystem accessible to Node.js; - Codex authentication only when using the Codex worker; - Claude authentication or API access only when using Claude Code or the direct Claude worker. The final local release audit ran on Windows. GitHub Actions now builds and tests the public repository on Ubuntu, Windows, and macOS with Node.js 20 and 22. Dependency resolution was also checked for Linux x64 and Darwin ARM64. Hosted CI is valuable compatibility evidence, but it is not the same as a full interactive Codex integration test on every platform. ## Core commands ```text lattice lattice run "<task>" --worker mock lattice run "<task>" --worker manual lattice run "<task>" --worker codex lattice run "<task>" --worker claude lattice run "<task>" --worker codex --no-apply lattice continue <task-id> lattice handoff validate <task-id> lattice session new|show|reset lattice doctor lattice benchmark --worker mock lattice integration codex status|doctor|enable|disable lattice integration claude status|enable|disable lattice claude [--raw] lattice sidecar status|stop lattice --version lattice --about ``` Use `lattice <command> --help` for command-specific options. A direct Codex run can inherit the active Codex model settings, or accept explicit `--model`, `--reasoning-effort`, and `--model-policy` options. Model identifiers are passed to the provider; availability depends on the installed provider and account. ## Configuration Configuration is repository-local in `lattice.config.json`. Start from [`examples/lattice.config.example.json`](examples/lattice.config.example.json): ```json { "model": "inherit", "reasoningEffort": "inherit", "modelPolicy": "inherit" } ``` Do not commit credentials or provider session state. Lattice does not require an API key field in this file. Configuration precedence and experimental adaptive behavior are documented in [`docs/configuration.md`](docs/configuration.md). ## How it works 1. Lattice discovers a safe repository root and builds a local structural index. 2. A task compiler converts the goal into acceptance criteria and context needs. 3. The context kernel returns bounded pages instead of an unrestricted repository dump. 4. An agent or manual operator proposes edits against explicit edit grants. 5. Lattice checks fingerprints, applies the transaction in an isolated Git worktree that reproduces your uncommitted work and installed dependencies, and runs allowlisted verification commands. A rejected patch or a failed verification is returned to the worker for up to two corrections. 6. A verified patch is applied to your workspace with `git apply` after every source fingerprint is re-checked (`--no-apply` only verifies). 7. State and diagnostics are written beneath the repository-local `.lattice/` directory, which Lattice adds to `.git/info/exclude`; keep it private. See [architecture](docs/architecture.md), [protocol](docs/protocol.md), and [persistence schemas](docs/persistence-schemas.md). ## Tests and quality checks ```sh npm test npm run lint npm run format:check npm run scan:public npm run package:check ``` `npm test` builds the project before running the Vitest suite. The public-export scanner checks source files while ignoring expected generated directories in a working clone (`.git`, `node_modules`, `dist`, and `.lattice`). CI and release preparation invoke the scanner directly in strict export mode. Both modes report suspicious artifacts and exit non-zero; neither deletes files. The scanner is defense in depth, not proof that a repository is safe. ## Security and privacy Lattice reads source code in the repository you point it at. Context sent to a remote model is subject to that provider's terms, account settings, and retention policy. Local metadata can contain source excerpts, diffs, goals, and diagnostics. Treat `.lattice/` as sensitive: Lattice adds it to the clone-local `.git/info/exclude` so it is not committed, but remove it before sharing a repository copy. The optional transparent Codex integration can create Lattice-owned launch shims, an MCP registration, and Codex hooks in user-level configuration. It is never enabled by installation. Run `lattice integration codex doctor`, review the reported paths, and keep a configuration backup before enabling it. The disable command removes only state that Lattice recognizes as its own. Automatic persistent-PATH setup supports Windows and Linux (Bash, Zsh, Fish). Linux live Omarchy sessions and macOS automatic integration remain unverified. See [installation](docs/installation.md) and [provider status](docs/providers.md). Read [SECURITY.md](SECURITY.md) and [`docs/security.md`](docs/security.md) before real use. ## Limitations - This is an early public release, not a hosted service or security boundary. - A smaller context is not automatically a correct context. - Verification is only as strong as the repository's tests and the configured command allowlist. - Token and latency savings vary by task, repository, model, cache state, and provider accounting. - Local integration tests do not substitute for a live provider evaluation. - The credential-free reset-token benchmark is a deterministic functional fixture, not evidence of model quality or savings. - The Opus 5.5 result is one owner-run pair on one small, maintainer-authored fixture; it is not independent task selection. - The Astra, Luna, Sol and Opus 5 pairs were measured with v1.0.0, whose Lattice arm received extra fixture-specific acceptance criteria; re-measure before citing them. The complete list is in [`docs/limitations.md`](docs/limitations.md). ## Evaluation Performance claims should come from paired, isolated runs with provider-reported usage, evaluator-owned tasks, disclosed failures, and a predeclared acceptance rule. The proposed protocol is documented in [`docs/evaluation.md`](docs/evaluation.md). The release includes a sanitized, reproducible [credential-free smoke-test result](docs/evidence/mock-benchmark-v0.1.0.json). The current source also publishes a CI-checked [240-case safety-frontier result](docs/evidence/economy-frontier.json) covering risk classification, fail-closed protocol handling, path confinement, exact patch lowering, and the live-evaluation budget guard. Reproduce it with `npm run evidence:frontier`. These are deterministic local checks, not live model results or evidence of token savings. The source also includes sanitized paired records for [Claude Opus 5.5](docs/evidence/owner-run-claude-opus-5-5.md), [GPT-5.6 Luna](docs/evidence/owner-run-gpt-5.6-luna.md), [GPT-5.6 Sol](docs/evidence/owner-run-gpt-5.6-sol.md), and a [community-run Claude Opus 5 reproduction](docs/evidence/community-run-claude-opus-5.md), plus the [spend-gated public drivers](benchmarks/README.md). Each record is one RAW run and one Lattice run on one fixed fixture. None is population-level or independent task-selection evidence. Raw provider sessions, private configuration, local paths, and unreviewed transcripts remain excluded. ## Contributing Read [CONTRIBUTING.md](CONTRIBUTING.md) and the [Code of Conduct](CODE_OF_CONDUCT.md). Bug reports and pull requests must not contain secrets, personal paths, private source, model transcripts, or provider session data. ## Support See [SUPPORT.md](SUPPORT.md). Security vulnerabilities belong in the private reporting path described by [SECURITY.md](SECURITY.md), not in public issues. Lattice is independently built and maintained. To support its development, discuss sponsorship, or help with access to testing infrastructure, contact [ptech1500@gmail.com](mailto:ptech1500@gmail.com?subject=Lattice%20support). You can also support the project through [Patreon](https://www.patreon.com/c/moulwyse). Patreon support helps cover cross-platform testing, CI, reproducible benchmarks, and model access. Please do not send credentials, API keys, or private source code by email. ## License and authorship Licensed under the [Apache License 2.0](LICENSE). Attribution and provenance are recorded in [NOTICE](NOTICE), [AUTHORS.md](AUTHORS.md), and [CITATION.cff](CITATION.cff). Project-name guidance for forks is in [TRADEMARKS.md](TRADEMARKS.md). Lattice was originally created and developed by Moulwyse. - Original author: <https://github.com/moulwyse> - Canonical repository: <https://github.com/moulwyse/lattice>