Repository navigation
Workflows EN
Examples with examples/… paths run from a laya-tools clone. Flags precede queries. Synthetic prices and catalogs are examples, not your real configuration.
riidolaya search --root . --lexical --json 'redirect authorization'
riidolaya search --root . --candidates 8 --limit 3 'where are redirect headers removed?'Use keywords/identifiers first. The second command additionally uses Laya if installed, or reports a warning and falls back to keywords. Results contain paths, line ranges, and excerpts for inspection. Missing candidates cannot be recovered by reranking. Reranking can add substantial latency; compare it with --lexical on your own tasks.
For the optional code-relevance checkpoint:
riidolaya setup --checkpoint code
riidolaya search --checkpoint code 'redirect authentication'Use base, not code, for model/repository choice experiments. Korean retrieval is unvalidated; --candidate-query 'English identifiers' can help keyword candidate generation.
riidolaya plan --config examples/planner/config.json --request examples/planner/request.json --jsonThink of this as a calculator: the catalog describes capability/context/prices; the request supplies estimated tokens, assessment, current model, and switching history. These inputs are not discovered from your subscription.
plan.status |
What to do |
|---|---|
recommend |
Inspect the proposed model and rejection reasons for alternatives. Nothing ran. |
hold |
The current model satisfies constraints and a switch is withheld. |
blocked |
No permissible recommendation. Check required capabilities, prices, budget usage, and reasons. Do not execute an empty model ID. |
A valid blocked plan exits with code 0. Invalid JSON/input exits nonzero. A trailing task adds local Laya classification; it does not call a paid coding model. This plan is not automatically applied by codex.
riidolaya repo-preview --catalog examples/repositories/catalog.json --json 'refund invoices'
riidolaya repo-preview --catalog examples/repositories/catalog.json --laya --json 'refund invoices'Supply a catalog filtered to repositories the caller may access. Use short role summaries and aliases; do not include whole source trees. See the example catalog.
status |
Meaning |
|---|---|
candidate |
Keyword candidate only; not a calibrated selection. Inspect reason. |
suggest |
Experimental Laya suggestion passed thresholds; still not execution authority. |
abstain |
No safe suggestion; inspect candidates/reason, or resolve manually. |
Every result has preview: true. No GitHub access, clone, repository modification, or work assignment occurs. No permission is granted. --laya is optional and English-only in the tested configuration. Current measurements did not establish an accuracy benefit.
Replace placeholder IDs with models available in your Codex installation:
riidolaya route --fast-model YOUR_FAST_MODEL --standard-model YOUR_STANDARD_MODEL --strong-model YOUR_STRONG_MODEL --json 'Fix a spelling mistake'
riidolaya codex --model YOUR_MODEL --dry-run 'Implement the feature'The first only recommends. The second shows the proposed command. Removing --dry-run starts the installed Codex and can consume your normal Codex usage. Existing login/permissions/approval settings remain in force. No active conversation is switched.
In route output, suggested_tier is the model's proposal; tier is the applied policy; abstained: true explains retaining stronger capability. A high classification probability is not a coding-success guarantee. An explicit --model override wins.
Next: Agent/Go integration or troubleshooting.
Easy work can move down; difficult work can move to a stronger model. Set the current model with request JSON current and capability ordering with catalog rank. A higher price alone does not mean an upgrade.
riidolaya plan --config examples/planner/config.json --request examples/planner/upgrade.json --jsonThis example recommends example-fast → example-strong with direction: upgrade and reason: quality_upgrade. The existing request.json demonstrates a downgrade. This example uses a supplied assessment without invoking Laya. Append an English task description to let local Laya assess complexity instead.
direction is upgrade, downgrade, lateral (same rank), initial (first selection), or unchanged (hold); blocked plans omit it. Upgrades prioritize quality over cache payback and cooldown, but still respect budgets, capabilities, context limits, and manual pins. Uncertain Laya output is not upgrade evidence; if the current model also fails the constraints, the result is blocked. With default configuration, catalog confidence must also reach 0.9: the switch upgrade threshold of 0.5 alone is insufficient.
These are recommendations only. No active Codex conversation is switched and no failure is automatically detected or retried. An integrating agent must update the current model and assessment at each task stage.
Decomposition is not an executable feature yet. A generative agent proposes a plan; Laya will be evaluated/trained to choose keep atomic, split sequentially, split with parallel work, or request more context. Go checks dependencies, declared shared-resource conflicts, and budgets. Laya does not generate subtask descriptions or code.
For example, establishing a contract before implementation needs ordering; independent modules with fixed contracts may have concurrent work. Different files alone do not establish independence. Measure savings using success rates and total cost including planning, integration, and retries.
Decomposition design and external review · MPS training preparation · Tracking issue #11
riidolaya is experimental. No demonstrated Codex cost savings; repository selection is preview-only.
riidolaya는 실험 단계입니다. Codex 절감은 미입증이며 저장소 선택은 preview입니다.
English README · 한국어 README · Versioned Wiki sources / Wiki 원본
이전 연구 안내 펼치기 / Earlier research
- 실제 이어 학습 V3 · 한국어 · Actual V3 training · English
- 진행·완료 보고·질문100건 결과 KO · 100-case results EN
- 라벨·이모지·상태 힌트 KO · Content/state hints EN
- 개발 진행·완료 보고·질문 pilot KO · Development pilot EN
- 작은 검증 기록 KO · Compact evidence EN
- 현재 개발 자료 35개·다음 학습 조건 · 35-row data and next training gates
- 개발용 학습 자료 2개 · Two qualified training requests
- 첫 두 과제 실제 관측 · First two actual observations
- 전체 사용법 KO · Full usage EN
- 작은 동작 힌트 KO · Small hints EN
- 설명 보강 비교76 · Caption diagnostic76
- Go 압축 가중치78 · Packed Go weights78
- 공개 함수 동작 검증 · Public function verification
- 다음 학습 자료 준비 · Training data preparation
- 동작에서 학습 정답·실제 학습79 · Grounded labels and actual fit79
- 소스 동작 관찰 · Source observations
- 저장 크기와 RAM · Storage and RAM
- 라이선스 · Licensing
- 시점별 기록 원문 / Chronological history
- 전체 연구 링크 / Full research index
- 문서 목록 / Documentation