kernel-KBS
kernel-KBS is a read/query skill. It retrieves evidence-backed GPU kernel knowledge from a SQLite-centered corpus. It does not run kernels, profile kernels, benchmark kernels, or manage optimization experiments.
Default Authority
Default agent mode is read-only.
Allowed without asking first:
- Query and inspect KBS content.
- Run read-only commands such as
query, get, schema, stats, doctor, and check.
- Read Markdown pages, references, schemas, ledgers, and artifact metadata needed to answer.
Requires explicit human permission in the current task:
- Add, update, delete, restore, or export back to source.
- Run
build-db, reindex, optimize, or vacuum.
- Run PR refresh commands that apply changes or fetch artifacts.
- Mutate files, SQLite rows, indexes, state, ledgers, or artifacts.
delete --hard requires a separate explicit confirmation even when deletion has been approved.
First Command
All commands run from the kernel-KBS skill root. Use scripts/kbs.py as the single front door.
Start with SQLite query:
python3 scripts/kbs.py query "<terms>" --limit 10
Then narrow with filters:
python3 scripts/kbs.py query tcgen05 --architecture sm100 --limit 10
python3 scripts/kbs.py query --repo cutlass --language cute-dsl --limit 20
python3 scripts/kbs.py query --tag nvfp4 --has-code --limit 20
python3 scripts/kbs.py query --symptom memory-bound --limit 10
Fetch evidence after selecting likely rows:
python3 scripts/kbs.py get <doc-id> --follow-sources --include-code
python3 scripts/kbs.py get <doc-id> --json
Use schema and health commands when needed:
python3 scripts/kbs.py schema
python3 scripts/kbs.py schema --table docs
python3 scripts/kbs.py stats
python3 scripts/kbs.py doctor
python3 scripts/kbs.py check all
Do not scan Markdown recursively for retrieval. Retrieval goes through SQLite: docs, docs_fts, tags, doc_links, artifacts, perf_claims, kbs_meta, and schema_docs.
Query Workflow
- Start broad with
query <terms> --limit 10.
- Add filters only when useful:
--kind, --repo, --architecture, --language, --tag, --symptom, --confidence, or --has-code.
- Fetch top rows with
get <id> --follow-sources --include-code.
- Follow
source links before making evidence-backed claims.
- Cite page ids and canonical paths from SQLite result rows.
- Load references only if the query is too broad, the user asks for schema/query guidance, or you need a topic map.
Navigation
| Need |
Open / Run |
| Directory layout and storage boundaries |
store/store.md |
| Script responsibilities and allowed entry points |
scripts/scripts.md |
| SQLite schema and workflow notes |
python3 scripts/kbs.py schema |
| Topic map and canonical page ids |
references/primer.md |
| Frontmatter rules and controlled-vocabulary pointers |
references/schema.md |
| Worked query examples |
references/examples.md |
| Actual controlled tags and aliases |
store/schemas/tags.yaml, store/schemas/aliases.yaml |
| Current corpus counts |
python3 scripts/kbs.py stats |
Core storage map:
| Layer |
Path |
Role |
| Config |
store/config/ |
Layout, corpus manifest, PR refresh configuration. |
| Schemas |
store/schemas/ |
Page schemas, controlled tags, aliases. |
| Source docs |
store/docs/sources/ |
PR/blog/doc/contest source Markdown. |
| Knowledge pages |
store/docs/wiki/ |
Curated synthesis pages for answers. |
| Ledgers |
store/docs/ledgers/ |
Candidate/core/source ledgers and policy inputs. |
| Artifacts |
store/corpus/artifacts/ |
Diffs, code assets, provenance bundles. |
| Indexes |
store/indexes/kernel-KBS.sqlite |
SQLite metadata, relationships, and contentless FTS. |
| State |
store/state/ |
Structured maintenance state: refresh/, versions/, and audits/{content,refresh,validation}/. |
Full Markdown bodies remain under store/docs/. SQLite stores metadata, paths, short excerpts, relationships, artifact pointers, performance claims, and FTS terms. It does not store a second readable copy of every Markdown body.
Permissioned Maintenance
Only run these after explicit human approval:
python3 scripts/kbs.py put <markdown-path>
python3 scripts/kbs.py update <doc-id> ...
python3 scripts/kbs.py delete <doc-id>
python3 scripts/kbs.py restore <doc-id>
python3 scripts/kbs.py export <doc-id> --write-source
python3 scripts/kbs.py build-db --reset
python3 scripts/kbs.py reindex
python3 scripts/kbs.py optimize
python3 scripts/kbs.py vacuum
python3 scripts/kbs.py update-prs --since <YYYY-MM-DD> --apply
Dry-run PR discovery is read-only from the corpus perspective, but it may require network and credentials:
python3 scripts/kbs.py update-prs --since <YYYY-MM-DD> --max-new 20
If approved maintenance changes Markdown or artifacts, run the approved rebuild/check sequence appropriate to the change:
python3 scripts/kbs.py build-db --reset
python3 scripts/kbs.py doctor
python3 scripts/kbs.py check all
Answer Contract
When answering from this skill:
- Cite concrete page ids and canonical paths, for example
kernel-flash-attention-4 at store/docs/wiki/kernels/flash-attention-4.md.
- Use
sources, doc_links, artifact_dir, and artifacts metadata before making evidence-backed claims.
- Preserve confidence labels exactly:
verified, source-reported, inferred, or experimental.
- For performance claims, include available
gpu, dtype, shape, metric, value, and source_id.
- Keep profiling, benchmarking, environment checks, and correctness workflows outside this skill.
1---2name: kernel-kbs3description: Corpus-backed GPU kernel knowledge base for CUDA, Triton, CuTe, CUTLASS, and Ampere/Hopper/Blackwell kernel research. Use when the user needs to search merged kernel PR pages, inspect PR diff/provenance artifacts, find KernelWiki synthesis pages, query blog/doc/contest notes, or retrieve evidence-backed implementation patterns by hardware feature, technique, repo, language, or kernel type. Do not use for environment checks, correctness checks, Nsight Compute profiling, benchmarking, or iterative optimization bookkeeping.4---56# kernel-KBS78`kernel-KBS` is a read/query skill. It retrieves evidence-backed GPU kernel knowledge from a SQLite-centered corpus. It does not run kernels, profile kernels, benchmark kernels, or manage optimization experiments.910## Default Authority1112Default agent mode is **read-only**.1314Allowed without asking first:15- Query and inspect KBS content.16- Run read-only commands such as `query`, `get`, `schema`, `stats`, `doctor`, and `check`.17- Read Markdown pages, references, schemas, ledgers, and artifact metadata needed to answer.1819Requires explicit human permission in the current task:20- Add, update, delete, restore, or export back to source.21- Run `build-db`, `reindex`, `optimize`, or `vacuum`.22- Run PR refresh commands that apply changes or fetch artifacts.23- Mutate files, SQLite rows, indexes, state, ledgers, or artifacts.2425`delete --hard` requires a separate explicit confirmation even when deletion has been approved.2627## First Command2829All commands run from the `kernel-KBS` skill root. Use `scripts/kbs.py` as the single front door.3031Start with SQLite query:3233```bash34python3 scripts/kbs.py query "<terms>" --limit 1035```3637Then narrow with filters:3839```bash40python3 scripts/kbs.py query tcgen05 --architecture sm100 --limit 1041python3 scripts/kbs.py query --repo cutlass --language cute-dsl --limit 2042python3 scripts/kbs.py query --tag nvfp4 --has-code --limit 2043python3 scripts/kbs.py query --symptom memory-bound --limit 1044```4546Fetch evidence after selecting likely rows:4748```bash49python3 scripts/kbs.py get <doc-id> --follow-sources --include-code50python3 scripts/kbs.py get <doc-id> --json51```5253Use schema and health commands when needed:5455```bash56python3 scripts/kbs.py schema57python3 scripts/kbs.py schema --table docs58python3 scripts/kbs.py stats59python3 scripts/kbs.py doctor60python3 scripts/kbs.py check all61```6263Do not scan Markdown recursively for retrieval. Retrieval goes through SQLite: `docs`, `docs_fts`, `tags`, `doc_links`, `artifacts`, `perf_claims`, `kbs_meta`, and `schema_docs`.6465## Query Workflow66671. Start broad with `query <terms> --limit 10`.682. Add filters only when useful: `--kind`, `--repo`, `--architecture`, `--language`, `--tag`, `--symptom`, `--confidence`, or `--has-code`.693. Fetch top rows with `get <id> --follow-sources --include-code`.704. Follow `source` links before making evidence-backed claims.715. Cite page ids and canonical paths from SQLite result rows.726. Load references only if the query is too broad, the user asks for schema/query guidance, or you need a topic map.7374## Navigation7576| Need | Open / Run |77|---|---|78| Directory layout and storage boundaries | `store/store.md` |79| Script responsibilities and allowed entry points | `scripts/scripts.md` |80| SQLite schema and workflow notes | `python3 scripts/kbs.py schema` |81| Topic map and canonical page ids | `references/primer.md` |82| Frontmatter rules and controlled-vocabulary pointers | `references/schema.md` |83| Worked query examples | `references/examples.md` |84| Actual controlled tags and aliases | `store/schemas/tags.yaml`, `store/schemas/aliases.yaml` |85| Current corpus counts | `python3 scripts/kbs.py stats` |8687Core storage map:8889| Layer | Path | Role |90|---|---|---|91| Config | `store/config/` | Layout, corpus manifest, PR refresh configuration. |92| Schemas | `store/schemas/` | Page schemas, controlled tags, aliases. |93| Source docs | `store/docs/sources/` | PR/blog/doc/contest source Markdown. |94| Knowledge pages | `store/docs/wiki/` | Curated synthesis pages for answers. |95| Ledgers | `store/docs/ledgers/` | Candidate/core/source ledgers and policy inputs. |96| Artifacts | `store/corpus/artifacts/` | Diffs, code assets, provenance bundles. |97| Indexes | `store/indexes/kernel-KBS.sqlite` | SQLite metadata, relationships, and contentless FTS. |98| State | `store/state/` | Structured maintenance state: `refresh/`, `versions/`, and `audits/{content,refresh,validation}/`. |99100Full Markdown bodies remain under `store/docs/`. SQLite stores metadata, paths, short excerpts, relationships, artifact pointers, performance claims, and FTS terms. It does not store a second readable copy of every Markdown body.101102## Permissioned Maintenance103104Only run these after explicit human approval:105106```bash107python3 scripts/kbs.py put <markdown-path>108python3 scripts/kbs.py update <doc-id> ...109python3 scripts/kbs.py delete <doc-id>110python3 scripts/kbs.py restore <doc-id>111python3 scripts/kbs.py export <doc-id> --write-source112python3 scripts/kbs.py build-db --reset113python3 scripts/kbs.py reindex114python3 scripts/kbs.py optimize115python3 scripts/kbs.py vacuum116python3 scripts/kbs.py update-prs --since <YYYY-MM-DD> --apply117```118119Dry-run PR discovery is read-only from the corpus perspective, but it may require network and credentials:120121```bash122python3 scripts/kbs.py update-prs --since <YYYY-MM-DD> --max-new 20123```124125If approved maintenance changes Markdown or artifacts, run the approved rebuild/check sequence appropriate to the change:126127```bash128python3 scripts/kbs.py build-db --reset129python3 scripts/kbs.py doctor130python3 scripts/kbs.py check all131```132133## Answer Contract134135When answering from this skill:1361371. Cite concrete page ids and canonical paths, for example `kernel-flash-attention-4` at `store/docs/wiki/kernels/flash-attention-4.md`.1382. Use `sources`, `doc_links`, `artifact_dir`, and `artifacts` metadata before making evidence-backed claims.1393. Preserve confidence labels exactly: `verified`, `source-reported`, `inferred`, or `experimental`.1404. For performance claims, include available `gpu`, `dtype`, `shape`, `metric`, `value`, and `source_id`.1415. Keep profiling, benchmarking, environment checks, and correctness workflows outside this skill.
fmh66