菜单
Use this skill when updating SdkAgentTraining for a new public mainsequence SDK version, deciding whether a new authored case-set version is required, refreshing sdk/<version>/ snapshots, creating cases/vN, and mapping the SDK to the correct case set.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | sdk-agent-training-new-version-creation |
| description | Use this skill when updating SdkAgentTraining for a new public mainsequence SDK version, deciding whether a new authored case-set version is required, refreshing sdk/<version>/ snapshots, creating cases/vN, and mapping the SDK to the correct case set. |
Use this skill when the public mainsequence library changes and the training repo must decide how to represent that change.
The goal is to keep three things separate:
sdk/<sdk-version>/cases/<case-set-version>/runs/sdk/<sdk-version>/...Do not mix authored cases into SDK snapshots.
Before changing structure, read:
README.mddocs/structure.mddocs/conventions.mddocs/sdk-cli-notes.mdsdk/<latest-known-version>/case-map.yamlsdk/<latest-known-version>/source-of-truth.yamlcases/<latest-case-set-version>/manifest.yamlThe local structure docs define how this repo should evolve.
Before creating anything, decide:
sdk/<installed-version>/ already exist?Do not create cases/vN just because a patch version changed.
Create a new cases/vN when at least one is true:
Do not create a new case set when:
sdk/<version>/case-map.yamlRun:
git status --short
uv run python -c 'import importlib.metadata as m; print(m.version("mainsequence"))'
find sdk -maxdepth 2 -type f \( -name manifest.json -o -name case-map.yaml \) | sort
find cases -maxdepth 2 -type d | sort
If there are unrelated dirty files, do not revert them.
If a dirty file is in a path you need to regenerate, stop and decide whether to preserve, migrate, or intentionally overwrite it.
For latest public SDK:
uv lock --upgrade-package mainsequence
uv add 'mainsequence==<resolved-version>'
Use an exact pin once the version is known. The training repo should be reproducible.
Run:
uv run python scripts/populate_training_skills.py
The script must:
mainsequence versionagent_scaffold bundleAGENTS.md and each SKILL.md into sdk/<installed-version>/sdk/<installed-version>/manifest.jsonsdk/<installed-version>/source-of-truth.yamlsdk/<installed-version>/case-map.yamlThe copied SKILL.md files must remain exact snapshots of the installed package.
source-of-truth.yaml must record the public git repository and tag/ref. If the
commit has been verified, record the resolved commit SHA and verification
method. If it has not been verified, keep verification_status: unverified and
do not claim that the snapshot is audit-complete.
Compare old and new snapshots:
find sdk/<old-version>/skills -path '*/source/SKILL.md' | sed 's#^sdk/<old-version>/skills/##; s#/source/SKILL.md$##' | sort
find sdk/<new-version>/skills -path '*/source/SKILL.md' | sed 's#^sdk/<new-version>/skills/##; s#/source/SKILL.md$##' | sort
Classify changes:
This classification drives the case-set decision.
If a new case set is required:
v2, v3cases/<new-case-set>/manifest.yamlcases/<new-case-set>/skills/<skill-path>/ for every skill in the new SDK snapshotcases/ folders for new skill pathsMigrated cases must be marked:
case_set_version: vN
target_sdk_version: <new-sdk-version>
migrated_from_case_set: vN-1
migration_status: pending_revalidation
Do not mark migrated cases as current until they have been reviewed against the new SDK skill text and current docs.
For a new case set, sdk/<new-version>/case-map.yaml should usually map every new SDK skill to the new case set:
sdk_version: <new-sdk-version>
default_case_set: vN
skills:
platform_operations/orchestration_and_releases:
case_set: vN
Use a mixed map only when a specific skill can safely keep using an older case set.
When a new SDK version or case set is created, update:
README.mddocs/structure.mddocs/conventions.mddocs/ollama-workflow.md if examples mention a concrete SDK versionIf a spec has not been revalidated, say that directly:
Current checked SDK basis: mainsequence==<old-version>
SDK <new-version> is the active snapshot for cases/<new-case-set>; this spec is carried forward and must be revalidated.
For each migrated case:
SKILL.mdRecommended metadata after revalidation:
migration_status: revalidated
validated_against_sdk_version: <new-sdk-version>
validated_by: <agent-or-human-name>
validated_at: <ISO-8601 timestamp>
Before finishing, verify:
uv run python -c 'import importlib.metadata as m; print(m.version("mainsequence"))' prints the intended versionsdk/<new-version>/manifest.json existssdk/<new-version>/source-of-truth.yaml exists and includes the upstream git
repository and refsdk/<new-version>/case-map.yaml existsmanifest.json matches copied source/SKILL.md filescases/<new-case-set>/manifest.yaml exists if a new case set was createdsdk/<new-version>/case-map.yaml points to an existing case-set folderSuggested checks:
python3 -m py_compile scripts/populate_training_skills.py scripts/run_ollama_case.py scripts/evaluate_case.py
uv run python -c 'from pathlib import Path; import scripts.run_ollama_case as r; v=r.read_sdk_version(); s=r.locate_sdk_root(v); p=r.resolve_case("or-001-recurring-artifact-job", s); print(v); print(p.relative_to(Path.cwd()))'
If the known case does not exist in the new case set, use another migrated case id.
sdk/<sdk-version>/.When reporting the result, include:
Use this skill when creating, updating, reviewing, or planning evaluation cases in the SdkAgentTraining repository, including writing case.yaml, prompt.md, expected responses, rubrics, skill case-bank READMEs, and cases derived from SDK skill snapshots or training-source documents.
Use this skill when the task is about defining, changing, querying, or reviewing Main Sequence SimpleTables and SimpleTableUpdaters. This skill owns row schemas, backend-managed ids, insert versus overwrite behavior, filtering, foreign keys, and validation rules. It does not own DataNode producers, API route contracts, scheduling, or sharing policy.
Use this skill when the task is about producing, changing, validating, or reviewing Main Sequence DataNode update processes. This skill owns DataNode update configuration, dependencies, update logic, hashing, namespaces, and validation against a PlatformTimeIndexMetaTable storage contract. It does not own generic MetaTable governance, Alembic migration execution, API route contracts, scheduling, sharing policy, or storage registration internals.
Use this skill when the task is about defining, querying, or reviewing Main Sequence MetaTables. This skill owns SQLAlchemy table contracts, backend-managed table authoring, external table registration, governed compiled SQL operations, foreign keys, indexes, naming, cadence, and validation rules. For Alembic migration lifecycle work, use the mainsequence-metatable-migrations skill. It does not own DataNode producers, API route contracts, scheduling, releases, or sharing policy.
Use this skill when the task is about producing, changing, validating, or reviewing Main Sequence DataNodes. This skill owns DataNode contracts, hashing, namespaces, update logic, metadata, asset-indexed nodes, and DataNode validation. It does not own SimpleTable row modeling, API route contracts, scheduling, or sharing policy.
Use this skill when the task is about building or changing APIs in a Main Sequence repository. In a Main Sequence project, project APIs should be implemented as FastAPI project resources. This skill owns FastAPI structure, request and response contracts, request user binding, APIDataNode and SimpleTable consumption inside APIs, and exact widget-facing API response contracts. It does not own producer-side DataNode or SimpleTable design, workspace payloads, or scheduling and release workflows.