writer-review
Guides LazyLLM writer artifact models, writer tools, revision/stream workflows, Feishu adapter boundaries, and review CLI planning.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- writer-review
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Guides LazyLLM writer artifact models, writer tools, revision/stream workflows, Feishu adapter boundaries, and review CLI planning.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/artifact-formats.md2.0 KB
- 参考文档references/troubleshooting.md1.7 KB
- 参考文档references/writer-review-workflows.md2.5 KB
- 脚本scripts/writer_artifact_smoke.py3.1 KB
03 SKILL.md 正文
LazyLLM Writer and Review
Use this sub-skill for LazyLLM writer workflows, writer intermediate representation (IR), artifact JSON envelopes, writer tools, revision/stream utilities, Feishu adapter boundaries, lazyllm review, review-local, and git review examples.
Start here when
- The task mentions writer data models such as
WriterDocument,WriterBlock,WriterSpan,WritingContext, orResourceProfile. - The user wants to run or adapt writer tools, revision tools, stream tools, or a writer pipeline.
- The task involves Feishu document bindings, provider payloads, or local artifact stores.
- The user wants to plan LazyLLM PR review commands or local code-review output without posting remotely.
Files to read
- writer-review-workflows.md for local writer artifact, tool, and review command workflows.
- artifact-formats.md for writer IR fields and artifact envelope expectations.
- troubleshooting.md for artifact, provider adapter, review CLI, and posting failures.
- scripts/writer_artifact_smoke.py for a no-network artifact round-trip check.
Safe workflow
- Start with local artifacts. Build a tiny
WriterDocumentandWritingContext; round-trip them through JSON before using provider adapters. - Keep provider bindings as metadata.
provider_bindingandprovider_payloadshould be preserved but not dereferenced unless the user supplies provider access. - Validate tool outputs. Writer tools should return artifact paths, context paths, schema names, counts, and summaries that downstream steps can inspect.
- Treat review commands as side-effecting.
lazyllm reviewmay post or call providers;review-localcan inspect git state and write reports. Ask before posting remote comments. - Route LLM/provider configuration to model-deployment when a writer pipeline calls an online model.
Key local facts
LazyLLM writer tests verify:
- nested
WriterDocument/WriterBlock/WriterSpanfields survive JSON round trips, iter_blocks()traverses depth-first andblock_by_id()finds nested blocks,- artifact envelopes contain
schema,schema_version,data, andmeta, save_artifact_json/load_artifact_jsonsupport models and lists,WriterToolBase._save_artifactsrecords primary artifact path, context path, schema names, counts, and metadata,- writer adapter, revision, stream, and tool tests are mostly local; provider-backed writer pipeline is optional.
Review CLI boundaries
Use review workflows for planning and local output unless the user explicitly asks to post:
lazyllm review --pr <number> [--repo owner/name] [--model ...] [--post] ...
lazyllm review-local [--repo-path .] [--base main] [--output review.json] ...
Do not post comments, call a provider, or inspect private repositories without user approval and credentials.
Handoff checklist
When you finish a writer/review task, include:
- artifact schema and file paths created or expected,
- whether provider bindings are preserved or used,
- local round-trip/tool smoke result,
- review command plan and side-effect status,
- model/provider/backend status if a writer pipeline uses an LLM.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).