LingTai Workflow is a user-submitted, maintainer-reviewed catalog of workflow and daemon-tactic bundles for LingTai agents.
This repository is an advisory reference only:
- It is not part of the LingTai runtime or kernel contract.
- It is not a model leaderboard and not a fixed routing rulebook.
- It never overrides human instructions, standing rules, tool safety, repository facts, or the current task contract.
- Every workflow must be re-judged against the current task before use.
- Entries may age, fail, be superseded, or be removed.
Accepted workflows are installable LingTai skill bundles. Each workflow directory is self-contained: a SKILL.md with YAML frontmatter (name, description, and an advisory marker) plus any supporting assets/, references/, or scripts/ it needs, all kept inside the bundle directory. A human can copy one bundle into an agent's skill library and use it; nothing here reaches outside its own folder.
| Workflow | Use when | Avoid when | Evidence / benchmark case | Author | Last reviewed | Details |
|---|---|---|---|---|---|---|
| multi-model-daemon-orchestration | A task mixes 2+ of: architecture/decomposition, domain judgment, large-file construction, hard red-team review, human-facing delivery | A single short answer with no files/validation; or treating the table as a fixed ranking or standing authorization | Recurring practice across research/paper/code/eval deliverables; plus a preliminary single-day Grok routing probe (not a benchmark claim) | Runyuan Wang / 9s5bz2jvd2-lang | 2026-06-27 | workflows/multi-model-daemon-orchestration/ |
| roundtable-expert-panel | A plan, PR, release gate, research brief, or product decision needs temporary expert perspectives and structured disagreement around one executor-owned task | A short deterministic action; no real reachable LingTai agents; or using panel consensus as authorization for side effects | Public companion project rawpaper123/Roundtable-skill, shared in Lingtai-AI/lingtai#531; reference/adaptation only, not a benchmark claim |
rawpaper123 / Roundtable Skill | 2026-07-04 | workflows/roundtable-expert-panel/ |
| lingtai-release-blog-publication | Publishing LingTai release logs after an authorized release; required stats, canonical /releases archive format, contributor audit, and web publish path matter |
Private/local artifacts, no publish authorization, unsafe private contributor evidence, or ad-hoc blog posts pretending to be release logs | 2026-07-03 v0.16.1/v0.10.4 release correction: required commit/line/issue/PR stats; rejected issue/PR authors count; /zh/releases/ is canonical; GitHub push auto-deploys lingtai-web |
Jason H / lingtaidev3bot | 2026-07-03 | workflows/lingtai-release-blog-publication/ |
| log-sanitizer-workflow | Preparing LingTai logs, event traces, daemon reports, or exported evidence bundles for sharing | No authorization to share; need live runtime tracing/dashboard; verification still has hits | Recent LingTai log-sharing cleanup pain point; local text/SQLite self-test included (not a privacy guarantee) | Runyuan Wang / 9s5bz2jvd2-lang | 2026-06-30 | workflows/log-sanitizer/ |
| maintenance-pr-batch-triage | A contributor opens a burst of small behavior-preserving maintenance/refactor PRs and the maintainer needs a first-look inventory, risk tiering, and merge-order plan | Feature/security/schema/release changes; or when a full merge-gate review is already required for one PR | Distilled from TZZheng maintenance PR batches in lingtai-kernel (#577-#581, #591/#592/#599-#606), including closing #581 as over-deduplication |
mimo-1, with Jason H guidance and TZZheng example attribution | 2026-06-30 | workflows/maintenance-pr-batch-triage/ |
Each accepted workflow is a skill bundle you install by copying its directory, whole, into an agent's custom skill library:
<agent>/.library/custom/<workflow-slug>/
Copy the entire workflows/<slug>/ directory (its SKILL.md and every
assets/, references/, scripts/ file) so the bundle stays self-contained.
The agent then discovers it like any other skill via the name and
description in SKILL.md.
A future LingTai TUI may expose a community-workflow entry that performs this copy for you; until that exists, the install is a manual directory copy. Either way the bundle format is the same self-contained directory.
Refresh requirement. Installed bundles are point-in-time copies. Entries here
age, get corrected, or are removed. Re-copy the bundle when you want the current
version, and re-judge it against the current task before use — an installed copy
is not a standing rule or a default authorization. last_verified in each
SKILL.md tells you how fresh the source entry was at review time.
scripts/validate_workflow_skills.py checks that every workflows/*/SKILL.md is
a well-formed installable skill bundle. It is stdlib-only (no install step) and
advisory: it never touches the LingTai runtime or kernel and defines no
marketplace protocol.
python3 scripts/validate_workflow_skills.py
It fails (nonzero exit) on real packaging defects — missing or unparseable
SKILL.md frontmatter, a missing name/description/author, no advisory
marker, a bundle that references paths outside its own directory, or a bundle the
index never links. It warns (still exit 0) on portability nits such as an
over-long description or file, a missing last_verified, or a frontmatter name
that differs from the directory slug. Run it before opening a workflow PR.
Submit workflows by GitHub pull request. A workflow contribution should:
- Add a new directory under
workflows/<slug>/. - Include a
SKILL.mdwith YAML frontmatter (nameanddescriptionare required). - Preserve author attribution in the frontmatter and body.
- Describe when to use it, when not to use it, required evidence or benchmark case, validation, and failure signals.
- Remove secrets, private paths, internal mail IDs, raw tokens, and private user data.
- State that the workflow is advisory-only.
- Run
python3 scripts/validate_workflow_skills.pyand confirm it passes.
See templates/workflow-bundle/SKILL.md for the starting shape.
Agents may suggest creating a workflow PR only at a task boundary, only after a non-trivial reusable workflow has been demonstrated, and only with human consent. Do not interrupt active human work to ask. Do not ask more than once per day.