Squire

Doğrulama ve boşaltma işleri için uzaktan çalışma zamanları.

GitHub

Dokümantasyon

Squire

Squire is a local performance layer for AI coding agents.

The agent still chooses every command. Squire watches and warms the local workspace, then serves exact cached results for deterministic read-only operations only when it can prove the result is still valid. If proof is missing, stale, too expensive, or unsafe, the command runs natively.

Agent chooses. Squire serves.

Install

curl -fsSL https://raw.githubusercontent.com/reidgoodbar/squire/main/install.sh | bash

The installer downloads verified GitHub release archives and installs squire plus the matching squire-codex driver to ~/.local/bin by default. squire comes from reidgoodbar/squire; squire-codex comes from the real Codex fork in reidgoodbar/squire-codex. If a C compiler is available, the installer also builds the optional local hot library used by the driver.

Squire does not change Codex auth or model configuration. squire-codex uses the normal Codex home, so authenticate and configure Codex the same way you already would.

Install somewhere else:

curl -fsSL https://raw.githubusercontent.com/reidgoodbar/squire/main/install.sh | SQUIRE_INSTALL_DIR=/usr/local/bin bash

Start

From a repo:

squire-codex

That is the normal product path. Codex opens normally, the model sees the same tool surface, and it keeps emitting ordinary commands such as git status, sed -n '1,80p' file, or python --version. Squire runs underneath that execution boundary and only replaces a command result when the local proof says the bytes are exact.

There is no required setup step, no global command shim, no prompt change, no new agent tool, and no MCP injection.

What Squire Does

  • Starts or reuses a local maintainer for the workspace.
  • Warms repo/file/tool proofs in the background.
  • Serves exact stdout, stderr, and exit code for proven read-only hits.
  • Falls back to native execution on every miss, unsafe command, proof failure, corruption, or unavailable daemon.

Validation, builds, tests, edits, mutating Git commands, package installs, and unknown shell commands are native-only.

Why The Cache Is Valid

Squire does not trust a cached answer just because it exists. Every replay must pass a local proof against the current world:

  • normalized command, cwd, repo root, and tool identity match;
  • relevant Git epochs, config/index fingerprints, file hashes, selected environment inputs, and executable identity match;
  • Git-sensitive outputs also prove cwd-relative boundaries plus external Git inputs such as included config files, global ignore files, global attributes, core.excludesFile, and core.attributesFile;
  • cached stdout, stderr, and exit code come from an exact native observation or an exact hot-prepared observation;
  • the operation is still allowed by policy.

If any proof element is missing or stale, Squire runs the original command natively. Stale records may remain on disk, but they are not replayable without a fresh proof.

Replay Coverage

Current proven lanes include repeated local Git metadata, repo summaries, bounded source/config reads, line windows, fixed-string single-file search, tight directory listings, safe environment probes, common version probes, and selected read-only shell compositions.

Examples:

  • git rev-parse HEAD
  • git status --short
  • git ls-files
  • git diff --stat
  • cat src/app.js
  • sed -n '1,80p' src/app.js
  • head -n 20 src/app.js
  • tail -n 20 src/app.js
  • grep -F token src/app.js
  • rg -F token src/app.js
  • git status --short | head -n 5

Unsupported forms run natively.

Status

After a session:

squire boost status --short

Advanced diagnostics:

squire boost status --json
squire setup

Internal diagnostic namespaces remain for release tests, but the product is just Squire.

Benchmarks

Latest scoped preload panel:

  • 31/31 commands exact
  • direct command group: 52.8x native p50-sum speedup
  • composed shell group: 7.3x native p50-sum speedup
  • full panel: 13.1x native p50-sum speedup
  • representative direct p95:
    • git rev-parse HEAD: 0.097ms
    • git status --short: 0.379ms
    • git ls-files: 0.156ms
    • cat src/app.js: 0.129ms
    • grep -F token src/app.js: 0.122ms

These are local command-serving measurements only. They do not include model thinking time, API/network latency, or a broad task-speedup claim.

Full tables and fuzz results: docs/BENCHMARKS.md

Advanced architecture notes: docs/ADVANCED.md