Getting started

한국어 · Home

Start with keyword search. You can get a useful result without downloading Laya, connecting Codex, or supplying credentials.

1. Install the executable

Open the latest release. Choose darwin-arm64 for Apple Silicon macOS or linux-amd64 for x86-64 Linux. Intel Macs and Linux ARM do not currently have published binaries. Download the archive and SHA256SUMS, compare the archive's SHA-256 with its matching entry, then extract it.

For example, the published v0.3.0 Apple Silicon archive can be checked and opened with:

shasum -a 256 riidolaya-v0.3.0-darwin-arm64.tar.gz
# Compare the printed digest with the matching SHA256SUMS line before continuing.
tar -xzf riidolaya-v0.3.0-darwin-arm64.tar.gz
./riidolaya version

On Linux use sha256sum and the linux-amd64 archive. Keep LICENSE, NOTICE, and licenses/ with redistributed copies. You can run ./riidolaya directly; examples below assume you have placed the executable on PATH.

Alternatively, build from source with Go 1.27.1 and a C compiler (macOS Command Line Tools):

git clone https://github.com/teamswyg/laya-tools.git
cd laya-tools
go build -trimpath -o bin/riidolaya ./cmd/riidolaya
./bin/riidolaya version

For source builds, replace riidolaya below with ./bin/riidolaya. Python is not required for user execution.

2. Get a first result without a model

From a Git repository root:

riidolaya search --root . --lexical --json 'routing confidence'

In laya-tools this searches the implementation. In another project, use its identifiers. Expect paths, line ranges, and excerpts, or an empty result when terms do not match. This command does not send source to a provider. For non-root/subdirectory searches, install rg (ripgrep).

Flags must precede the query. Use a root you intend the tool to read.

3. Try repository preview or model planning

The following fixture paths are in a clone of laya-tools; they are not included in a binary-only installation.

riidolaya repo-preview --catalog examples/repositories/catalog.json --json 'refund invoices'
riidolaya plan --config examples/planner/config.json --request examples/planner/request.json --json

The first returns a fictional billing repository as a keyword candidate. The second calculates a plan over fictional models/prices. Neither performs work or calls a paid model. See Workflows before interpreting their statuses.

4. Add local inference only if needed

riidolaya setup
riidolaya doctor
riidolaya bench --iterations 10

Setup downloads about 446 MiB for the current model archive, plus the runtime. The installed model file is about 572 MiB; measured process RAM with the model is about 1.4 GiB. doctor reports paths/existence; it is not an accuracy test. bench actually loads and invokes the model.

Caches: ~/Library/Caches/laya-tools on macOS; ~/.cache/laya-tools on Linux. LAYA_CACHE changes the location. Current setup also installs corrected models-v2 notices and verifies checksums. Existing LAYA_* variables remain supported.

Next: Choose a workflow. You do not need to register MCP, start a daemon, or connect Codex just to use this tool.

Read the development evidence

Fair external Go comparison55 explains actual input and independent verification conditions. The v2 IDs version the existing humanize and UUID logical requests. Actual results55 contains four records, one per profile per request. Both ordinal candidates passed; only the Sol6-requested UUID candidate passed. Versions, repetitions and tests do not inflate distinct requests.

Two child plans shared four ordered reservations in one durable ledger. Reservations, starts and terminal receipts all reached4, with0 unknown starts and confirmed process/auth cleanup. Failures were not refunded or retried. The cap applies to that ledger, not the entire host or provider-internal calls. Inputs, original language conditions, plans, budget and native routing predictions were frozen before coding outcomes. Laya truncated and abstained on both full inputs; this does not establish the ultra-small resource or savings target.

The initial 24 probes and development records do not replace final performance evaluation. The golden-set scale guide targets at least 2,400 distinct protected final requests per domain. Read preparation, actual model attempts, training labels and final eligibility as separate counts.