Squire

Production lanes are bounded but cover the common read-only command surface:

- Git metadata: supportedgit rev-parseforms and branch discovery.
- Repository reads: supportedgit status,git ls-files, andgit diffforms, including path-scoped diffs andgit diff --check, plus boundedgit log -N --oneline -- <literal paths>history.
- File and search reads: boundedcat, ordered single- or multi-rangesed -n,head,tail,nl -ba,file, fixed-stringgrep/rg, demand-prepared bounded repositoryrgsearches, and tightlsforms.
- Environment discovery: supported version probes,which/command -v, safeprintenv,whoami,id,hostname, andunameforms.
- Compositions: complete read-only plans over supported sources and filters, including pipes, sequences, redirection to/dev/null,head,tail, boundedsed -n,grep -F,wc -l, andsort.

No operation is removed. Builds, tests, edits, installs, mutating Git commands, expansions, unknown shell syntax, sensitive probes, and unsupported variants immediately follow Codex's unchanged native path. A safe cold miss does the same while requesting exact preparation in the background.rg --filesremains outside the bounded preparation policy and follows the native path.

Supported commands compile into typed bounded plans rather than exact command templates. Source proof and execution are separate: one proven file snapshot can serve different line selections, filters, and compositions without a cache entry for each command string. The same plan representation is implemented by the Go engine and native runtime, keeping future read operators additive while differential tests enforce parity at the ABI boundary.

Squire caches observations, not authority. A foreground replay either recomputes the inputs that can affect that command or reuses a cryptographic fingerprint while a completekqueue/inotifyguard reports no dependency change. The prepared epoch must still match. Proof inputs include the normalized command and cwd, Git refs/index/config and external behavior files, relevant workspace state, canonical paths, content hashes, command-specific environment proof, and executable identity. Guard failure always invalidates the resident proof.

For bounded file reads, an epoch mismatch may instead use the current-file lane: Squire retains the exact bytes read while computing the foreground SHA-256 proof and applies only its fixed byte grammar to those bytes. This requires no rewarm and does not persist the file or result. The cache may still contain stale records, but they are never replayed after a proof mismatch. Missing state, corruption, unsupported syntax, an ABI mismatch, or an unprofitable proof all become native fallback.

The invalidation suite changes file bytes without changing size or mtime, mutates the Git index and untracked set, changes same-size diffs, edits Git config, commits, renames branches, changes loose and packed object namespaces, changes environment inputs, and probes outside-workspace symlinks. Returning old bytes or an unsafe hit fails the release.

On July 16, 2026, a 500-command randomized production-ABI run recorded 421 exact hits, 79 safe fallbacks, 468 native comparisons, zero mismatches, and zero unsafe hits. Hit p50/p95/p99 was0.315/0.630/0.933ms; the same commands ran natively at8.084/27.692/51.820ms. A separate 500-query repository-search differential had 500 exact or order-equivalent hits and zero semantic mismatches. Bounded path history was 28/28 exact at0.333msp50 and0.492msp99 versus20.060msnative p50. All 2,048 steady calls were exact with0.433mswall p99. Under eight-way load, CPU p99 was0.334ms; scheduler- contended wall p99 was3.242ms.

Fresh livegpt-5.6-lunatreatments independently replayed 2/3 Express calls (66.7%), 5/5 Flask calls (100%), and 4/5 fmt calls (80%). Every treatment passed the50%all-terminal-call gate with valid accounting and zero diagnostic mismatches. These small live samples validate coverage, not causal whole-task wall time; model trajectories diverged before seeing tool results.

A deterministic 40-pair Codex attribution test held model responses, commands, workspace, and terminal payloads fixed. Six serial calls fell from374.876msto103.600ms, saving271.277ms(72.4%, paired 95% interval265.630-276.540ms). Two parallel batches saved52.6%. AB and BA orders both remained positive, while interleaved A/A and B/B intervals included zero.

This is a web browser that enables your coding agent, such as Claude Code, to visit websites on your behalf and assist you in identifying bugs or creating UI test cases.

Declarative workflow orchestration for Temporal using the CNCF Serverless Workflow specification.

All Azure MCP tools in a single server. The Azure MCP Server implements the MCP specification to create a seamless connection between AI agents and Azure services. Azure MCP Server can be used alone or with the GitHub Copilot for Azure extension in VS Code.

The Railway MCP Server enables natural language interaction with your Railway projects and infrastructure. Ask your IDE or AI assistant to create projects, deploy templates, manage environments, pull variables, redeploy services, and more.

The Shipyard CLI provides an MCP server for agents to manage Shipyard environments directly: by pulling logs, comparing branches, running tests, and stopping/starting environments..

Secure virtual machines for agents hosted by Superserve

Execute any LLM-generated code in the YepCode secure and scalable sandbox environment and create your own MCP tools using JavaScript or Python, with full support for NPM and PyPI packages

Provides tools to interact with the Ansible Automation Platform API for automation tasks.

A MCP server for managing Brev development environments using the Brev CLI.