docType, scope, status, authoritative, owner, language, whenToUse, whenToUpdate, checkPaths, lastReviewedAt, lastReviewedCommit

docType	scope	status	authoritative	owner	language	whenToUse	whenToUpdate	checkPaths	lastReviewedAt	lastReviewedCommit
guide	repo	current	true	skills	en	When installing, updating, or using the Tiangong AI reusable skills repository.	When install commands, target agents, scope behavior, environment variables, or available skill guidance changes.	AGENTS.md .docpact/config.yaml README.zh-CN.md .claude-plugin/** */SKILL.md	2026-09-13	7d8758dd19

Tiangong AI Skills

Repository: https://github.com/tiangong-ai/agent-skills

Use the skills CLI from https://github.com/vercel-labs/skills to install, update, and manage these skills.

Report a problem or suggest a capability

Use the feedback forms and contribution guide. Chinese and English reports are welcome. For CLI runtime problems or uncertain ownership, use the CLI forms. Auto Research can also prepare the same structured report from its installed reporting reference.

Atomic data skills

Twenty-one local candidate Skills—AirNow Hourly Observations, EPA EIS Records, Federal Register Documents, NASA FIRMS Active Fire, OpenAQ Air Quality, Regulations.gov Comments, Regulations.gov Comment Details, Regulations.gov Attachments, USBR Project Records, USBR RISE, USGS Water IV, three Open-Meteo sources, and GDELT DOC, Events, GKG, and Mentions, plus Bluesky Cascades, YouTube Video Search, and YouTube Comments—are thin semantic Skills over the Tiangong CLI TypeScript 7 data runtime. Each candidate records its stable capability and operation contract-major requirements plus any operation feature it actually depends on in references/tiangong-data-requirement.json; the caller or workspace runtime lock chooses the actual CLI build. The agent uses that same resolved CLI for data describe and data run. These Skills contain no second provider connector runtime and no per-Skill package lock.

This candidate set covers every item in the authoritative EcoCouncil source baseline, main@ac19289b4876d8a90595a0270721ef3f5ee7ced8, which contains 21 source-fetch Skills. All 21 have been revalidated against the source semantics, their requirements have been verified against the local CLI candidate, and the unified copy/symlink installation and data-specific gates pass. The qualifying CLI version and exact manifest/Schema digests live once in scripts/data-skill-migration-provenance.json; they are release evidence, not an installed Skill runtime lock. The repository-wide cold gate also passes with a separately scoped upstream filesystem-clock fix; that prerequisite is intentionally absent from this data migration branch.

See _docs/architecture/atomic-data-capabilities.md and _docs/runbooks/atomic-data-skill-migration.md for the ownership boundary, candidate inventory, source audit correction, staged migration order, and release gates. The audited RSS/full-text, Figshare-download, academic-paper, Tiangong/KB, and private-email candidates retain their existing content, artifact, product, research, or security boundaries instead of being narrowed into stateless data connectors.

The additional local candidate gdelt-web-ngrams-search is a post-migration extension, not a twenty-second EcoCouncil source Skill. It selects literal phrase-to-article discovery through the CLI's paired-minute NGrams/TOC operation. It does not silently replace DOC queries, reproduce DOC timelines, or join the three GDELT tables. Require data describe gdelt.web-ngrams from the same resolved CLI before using it; this does not assert that the extension is already published.

The current 0.0.61 candidate catalog keeps gdelt.doc-search, usbr.project-records, and usbr.rise discoverable but suspended after their production live gates failed; Auto Research therefore does not offer those operations for execution. Their thin Skills and stable requirements remain for diagnosis and later requalification. AirNow remains available: its connector now reads the same official files.airnowtech.org objects through the bucket's regional S3 endpoint after the public CloudFront route became unreachable.

Install the CLI

npm i skills -g

Install

• List available skills (no install):

  npx skills add https://github.com/tiangong-ai/agent-skills --list
  

• Install all skills (project scope by default):

  npx skills add https://github.com/tiangong-ai/agent-skills
  

• Install specific skills:

  npx skills add https://github.com/tiangong-ai/agent-skills --skill tiangong-auto-research --skill tiangong-kb-sci-search
  

• Install the Tsinghua graduate thesis LaTeX workflow:

  npx skills add https://github.com/tiangong-ai/agent-skills --skill tsinghua-graduate-thesis
  

• Install the Tiangong KB ingest workflow:

  npx skills add https://github.com/tiangong-ai/agent-skills --skill tiangong-kb-ingest
  

• For a WorkBuddy/CodeBuddy producer, install the thin adapter beside the canonical orchestrator:

  npx skills add https://github.com/tiangong-ai/agent-skills --skill tiangong-auto-research --skill tiangong-auto-research-workbuddy
  

Target agents and scope

• Target specific agents:

  npx skills add https://github.com/tiangong-ai/agent-skills -a codex -a claude-code
  

• Install globally (user scope):

  npx skills add https://github.com/tiangong-ai/agent-skills -g
  

• Scope notes:
  • Codex is a universal agent: project scope uses ./.agents/skills, and global scope uses $HOME/.agents/skills. CODEX_HOME does not change the skills@1.5.22 global destination.
  • Claude Code project scope uses ./.claude/skills; global scope uses $CLAUDE_CONFIG_DIR/skills when set, otherwise $HOME/.claude/skills.
  • Other agents have their own directories; inspect the exact path reported by the pinned skills CLI rather than deriving ~/<agent>/skills by analogy.

Install method

• Interactive installs let you choose:
  • Symlink (recommended)
  • Copy

Update and verify

• List installed skills:

  npx skills list
  

• Check for updates:

  npx skills check
  

• Update all skills:

  npx skills update
  

Environment Variables

Environment requirements live with each skill. Before using a skill that calls an external service, read that skill's references/env.md when present. npx skills add installs or links skill files; it does not provision language runtimes or execute post-install hooks. When a skill provides a locked runtime bootstrap and smoke command, run those explicit steps from its own instructions.

Tiangong KB Ingest Compatibility

tiangong-kb-ingest uses the exact reviewed CLI 0.0.48 distribution. After installing or updating the Skill, run the credential-free compatibility smoke:

npx --yes --package "@tiangong-ai/cli@0.0.48" -- tiangong-ai --version
npx --yes --package "@tiangong-ai/cli@0.0.48" -- tiangong-ai kb --help

The version command must print 0.0.48; KB help must list the bulk scan, metadata dry-run, collection list/schema, ingest, and status surfaces. These checks do not call or mutate the KB backend.

Auto Research External Skill Setup

Use npx skills directly for ordinary Skill management. For an Auto Research workspace, prefer the CLI's guarded setup layer: it still uses the exact pinned skills CLI underneath, while also binding source commits, tree hashes, destinations, license choices, safe credential bindings, and audit state. The Wizard lets ordinary users enter each selected key with hidden TTY input; named environment variables and bounded stdin/password-manager input remain explicit alternatives. Start with the read-only catalog or the guided Wizard:

REVIEWED_BOOTSTRAP_CLI_VERSION=X.Y.Z # replace with one reviewed exact stable release
npx --yes --package "@tiangong-ai/cli@$REVIEWED_BOOTSTRAP_CLI_VERSION" -- \
  tiangong-ai research setup catalog \
  --workspace /absolute/path/to/workspace --json
npx --yes --package "@tiangong-ai/cli@$REVIEWED_BOOTSTRAP_CLI_VERSION" -- \
  tiangong-ai research setup \
  --workspace /absolute/path/to/workspace

For repeatable non-interactive provisioning, run research setup init first. It creates a no-overwrite schema-v2 .tiangong-research/setup.yaml template plus setup.env.example. The YAML explicitly lists every current catalog Skill, credential, and setting, including disabled optional entries; the env example lists every credential variable with an empty placeholder. After the user reviews the current catalog, enabled states, licenses, models, pricing, checks, and confirmations, bare research setup detects only that workspace-local YAML and bypasses the Wizard. The optional real setup.env must be owner-only; disabled credentials stay empty and a non-empty disabled value is rejected. The removed v1 declaration and all invalid or incomplete declarations fail closed; they never trigger an interactive fallback or parent-directory search. Setup returns success only after every selected dependency, provider live check, Policy compatibility check, and independent reviewer smoke reaches complete readiness. Use explicit research setup wizard to choose the interactive path. See tiangong-auto-research/references/setup.md and tiangong-auto-research/references/env.md.

The bootstrap version is an explicit new-workspace choice, never latest, a tag, or a range. After apply creates runtime-lock.json, the installed orchestrator's bundled resolver runs exactly that locked version for ordinary workspace operations. A reviewed upgrade-capable candidate is invoked explicitly when the old installation cannot plan its own upgrade; follow the candidate's returned apply/recovery commands and return to the activated lock. Installed command support and release metadata are separate, and unavailable metadata is not evidence that no update exists. See the installed setup reference.

For acquisition-heavy work, the orchestrator uses the CLI's read-only artifact size and role-coverage forecasts, bounded atomic content batches, and explicit prepared scientific-review execution. Recovery reuses verified discovery and downloaded artifacts. A compatible locked runtime can revise acquisition in the same project before analysis, explicitly reopening discovery only for new sources; a compatible runtime also allows the owner to approve pre-analysis amendments to planned lifecycle declarations and existing model/parameter links. Substantive design changes and later revisions retain the formal successor workflow. Original and currently approved task requirements, native check records, independent review and portable audit stay linked without extra fixed reviewer rounds. Workflow closure is reported separately from task completion. See tiangong-auto-research/references/execution-assurance.md; frozen history and scientific requirements are never silently rewritten to make a gate pass.

The installed orchestrator owns ordinary research planning, evidence work and validation. On continuation it restores original/current obligations and current checks before selecting local work, reuses exact artifacts and honors independent workflow overrides. Valid prior goals need no repeated contract, and routine continuation does not create another complete audit bundle.

Before large acquisition that depends on institutional or other authorized full text, the orchestrator checks a small relevant sample through the actual planned route and reuses applicable exact successes. It distinguishes article identity and readable content from HTTP success or a downloaded binary. Sufficient lawful OA evidence needs no campus-network gate; unknown failures do not diagnose VPN or subscription coverage. See references/fulltext-access-preflight.md inside the installed Auto Research Skill.

For repeated diagnostic calculations, a compatible locked CLI provides a bounded investigation tied to one computational requirement. The native host can test approved configurations without per-trial confirmation, inspect immutable history and remaining quotas, then request separate exact candidate promotion. Candidates remain exploratory until frozen and rerun as a fresh certification Canary with the existing scientific reviews. Interrupted unknown results remain unresolved. See tiangong-auto-research/references/bounded-investigation.md.

The installed assurance reference connects already-fetched windows, closest literature and discriminative pilots to concrete research decisions and concise researcher explanations. Review findings lead to permitted repairs before a new bound review. Publication guidance keeps map conventions project-specific and visual claims conditional on actual image delivery to a capable independent reviewer. These instructions reuse existing records and do not add a lifecycle ledger or an automatic paid explanation/review round.

Compatible runtimes also fulfill predeclared pending scientific objects in place, preserve original-request provenance, and bind computational acceptance to an observed ordinary native calculation. Complete packet artifacts are available on demand without an arbitrary total context-length gate. A command report alone remains unverified execution; file hashes and read receipts do not certify scientific validity. These are CLI-owned contracts, not a second Skill executor.

The orchestrator gates a research question before setup or any tool call. A request that presupposes its conclusion or excludes contrary evidence is paused with one testable rewrite and resumes only after explicit user confirmation. Controversial topics and directional hypotheses remain allowed when null results, alternatives, and counterevidence are testable; evidence fabrication or concealment is refused.

The catalog also offers the tiangong-auto-research workflow orchestrator, default-baseline Brave internet evidence, optional Tiangong SCI/document/paper companions, and optional Anthropic or PPT Master post-closure authoring Skills. The workspace can be any user-selected directory. Every entry is external, separately licensed, pinned, and explicitly confirmed/selected; nothing is bundled or installed by a research package. See tiangong-auto-research/references/setup.md and external-skills.md. Built-in CLI data connectors are different: the native discover packet projects their current catalog dynamically, and Auto Research invokes the shared TypeScript runtime through its Research evidence command. The Skill does not duplicate provider adapters or keep a fixed connector list; Research adds only budget, owner-only credential, immutable receipt/ledger, artifact, and review bindings around the unchanged core result. When that result is larger than one Agent context, Auto Research keeps the full immutable evidence and exposes receipt-bound continuation cursors. Exhaustive row review follows those local pages to completion; summary work may stop early only with the presented fraction disclosed. Provider gaps, operation limits, and Agent-context projection remain separate signals. tiangong-auto-research-workbuddy is only a sandboxed-IDE adapter. It routes back to the canonical orchestrator and its signed reviewer-bridge reference; it does not define a second research workflow. For PPT creation, prefer PPT Master; Anthropic PPTX remains compatible and may be selected alongside it when its workflow fits the task.

For a top-journal goal, the orchestrator includes a conservative Research Policy template pack for article type, field, journal class, project brief, and four independent final-review roles. The CLI Policy Wizard copies the selected Markdown into the research workspace for human review; it reports when generic defaults remain, requires explicit approval of the exact content hash, and invalidates approval after any edit or expiry. Exact-journal readiness requires current official guidance and substantive human customization. These gates can produce a reviewable submission candidate, never a promise of editorial acceptance.

Before discovery, the current native Codex or Claude host must also provide a closed, target-specific scientific design. The CLI validates and freezes the design. Frozen model implementations and environment locks must first enter the workspace through research scientific object register; the Skill never asks the user to hand-copy them into .tiangong-research. The CLI requires independent research-design review before discovery, then real-record evidence-construct and pilot-methods reviews after acquisition and before analysis. After acquisition it also requires exact decomposition records, evidence atoms, a typed-content snapshot, a passing inference snapshot, a reproduced analysis, and a mechanically generated Claim-Evidence Graph. Publication freeze requires the complete manuscript sections and explicit cover/title/checklist/availability/source-data files; four fresh reviews must use the configured agent family that differs from the native producer. It reserves the complete early/final-review and revision lifecycle, requires reapproval for every authoritative recovery generation, and exports a semantic-chain-verified portable audit directory containing exact formal evidence rather than host-local pointers. The CLI remains the deterministic control plane; it never launches a nested producer to invent the science. See tiangong-auto-research/references/publication-policy.md and tiangong-auto-research/references/scientific-design.md.