AGENT SKILL

discovery-catalog

Inventory the contents of this Discovery catalog repo. Lists publishers, agents, starter-kits, and the tools shipped under each agent. Use this skill whenever the user asks what's in the catalog, who publishes what, what agents/tools/starter-kits exist, lists or counts of any of those, "show me the catalog", "do we have an agent for X", "what starter kits do we have", "list the publishers", "list tools per agent", or any inventory/discovery question about repo-resident assets — even if the user doesn't explicitly say the word "list".

License 许可
MIT
最近更新
2026-11-14
Stars
★ 289
Forks
101

01 前言 Front Matter

name
discovery-catalog
description
Inventory the contents of this Discovery catalog repo. Lists publishers, agents, starter-kits, and the tools shipped under each agent. Use this skill whenever the user asks what's in the catalog, who publishes what, what agents/tools/starter-kits exist, lists or counts of any of those, "show me the catalog", "do we have an agent for X", "what starter kits do we have", "list the publishers", "list tools per agent", or any inventory/discovery question about repo-resident assets — even if the user doesn't explicitly say the word "list".
allowed-tools
shell

02 资源文件 1

03 SKILL.md 正文

Discovery Catalog

Compatibility

This skill is read-only and runs on Windows, macOS, and Linux through PowerShell. PowerShell 7+ (pwsh) is recommended for consistent table rendering, though the dispatcher also works in Windows PowerShell 5.1 with less polished wrapping.

Overview

Read-only inventory of the Discovery catalog repo. One dispatcher script (scripts/discovery-catalog.ps1) supports list and item-inspection invocations:

Invocation What it lists
discovery-catalog publishers Every publisher across agents/ and starter-kits/, with per-publisher agent and starter-kit counts
discovery-catalog agents Every agent in the catalog (one row per agent folder name)
discovery-catalog agents list-tools Every agent plus the tools shipped under each agent's tools/ folder
discovery-catalog starter-kits Every starter kit (one row per kit folder name)
discovery-catalog <agent-name> describe One agent with Name, Publisher, Version, HasTool, Description
discovery-catalog <starterkit-name> describe One starter kit with Name, Publisher, Version, Category, AgentCount, Description
discovery-catalog <agent-name> list-tools One agent with Agent, Publisher, Version, ToolCount, Tools, Description

Repo layout the script handles

The catalog uses a flat layout — one folder per agent or starter kit directly under the parent directory:

Asset Path pattern Display
Agent agents/<agent-name>/ <agent-name>
Starter kit starter-kits/<kit-name>/ <kit-name>
Tool under an agent agents/<agent>/tools/<tool-name>/ <tool-name>

Inclusion rules:

  • An "agent" is any folder containing an agent.yaml.
  • A "starter kit" is any folder containing a kit.json.
  • Folders without those files (shared helpers, README-only directories) are silently skipped.
  • The agents/tmp/ scratch directory used by the deployer skills is always skipped.

Publisher and party come from the asset's metadata, not from the folder location:

  • For an agent: metadata.yaml → publisher.name (publisher) and publisher.party (1p / 3p).
  • For a starter kit: kit.json → author.name (publisher) and the top-level party field.

If publisher.name / author.name is missing, the publisher is reported as (unspecified).

When to Use

Use this skill whenever the user wants to know what is in the catalog. Trigger phrases include but are not limited to:

  • "What agents are in this repo?" / "List the agents" / "Show me the catalog"
  • "Who are the publishers?" / "Which partners have agents?"
  • "What starter kits do we have?" / "List the starter-kits"
  • "Show me agents with their tools" / "List tools per agent"
  • "Do we have an agent for retrosynthesis?" (use -Tag retrosynthesis)
  • "Which agents are missing a tool?" (use -WithoutToolsOnly)
  • Before invoking discovery-services-agent-deployer, when the user isn't sure of the exact agent folder name

Inputs

The skill is fully read-only and needs no Azure config. All inputs are positional or named parameters on scripts/discovery-catalog.ps1.

Parameter Position Description Default
Command 0 (mandatory) publishers, agents, starter-kits, or an agent/starter-kit name —
SubCommand 1 (optional) list-tools (with agents or <agent-name>) or describe (with <agent-name> / <starterkit-name>) empty
-Format named Table, Markdown, Json, or Plain Table
-Publisher named Filter by publisher name (case-insensitive). Applies to agents and starter-kits all
-Tag named Filter agents whose metadata.yaml includes this tag (agents only) all
-WithToolsOnly named Only show agents that ship a tools/ folder (agents only) off
-WithoutToolsOnly named Only show agents missing a tools/ folder (agents only) off

Invalid combinations exit with a non-zero code and a friendly message (e.g. agents foo, publishers list-tools).

Process

Mapping /discovery-catalog ... requests to invocations

When the user types a slash-style invocation like /discovery-catalog publishers, run the dispatcher with the matching arguments:

User says Run
/discovery-catalog publishers pwsh -NoProfile -ExecutionPolicy Bypass -File .github/skills/discovery-catalog/scripts/discovery-catalog.ps1 publishers
/discovery-catalog agents ... discovery-catalog.ps1 agents
/discovery-catalog agents list-tools ... discovery-catalog.ps1 agents list-tools
/discovery-catalog starter-kits ... discovery-catalog.ps1 starter-kits
/discovery-catalog <agent-name> describe ... discovery-catalog.ps1 <agent-name> describe
/discovery-catalog <starterkit-name> describe ... discovery-catalog.ps1 <starterkit-name> describe
/discovery-catalog <agent-name> list-tools ... discovery-catalog.ps1 <agent-name> list-tools

Use -Format Table for all catalog responses. Only add -Format Markdown if the user explicitly asks for a paste-ready markdown table. When the dispatcher output is long, gather the full result before replying; never mention internal transport details, truncation, temp files, or resource files to the user. Always return the complete table for the requested command; do not shorten it to a partial list or a narrative summary. In chat responses, present tabular results as Markdown tables for readability, while keeping the count anchor line (*_COUNT=) above the table.

How the dispatcher works (internal — for skill maintainers)

  1. Resolve repo root via git rev-parse --show-toplevel.
  2. Walk agents/ and/or starter-kits/ one level deep, skipping tmp/.
  3. Inclusion: agent folders need agent.yaml; starter-kit folders need kit.json.
  4. For each entry, parse the appropriate metadata file:
    • Agent: metadata.yaml for name, version, description, tags, plus the nested publisher.name and publisher.party.
    • Starter kit: kit.json for name, version, description, category, party, author.name, plus top-level agentRefs; treat the role: primary entry as the launch agent.
    • Tool (when agents list-tools): tool.yaml for name; folder name as fallback.
  5. Apply filters: -Publisher, -Tag, -WithToolsOnly, -WithoutToolsOnly.
  6. Sort by name (and publisher where applicable).
  7. Render in the requested format.

The first line of Table and Markdown output is a deterministic count anchor (AGENT_COUNT=, PUBLISHER_COUNT=, STARTER_KIT_COUNT=) so callers can grep for it.

Output Format

discovery-catalog publishers

PUBLISHER_COUNT=2

Publisher              Party AgentCount StarterKits
---------              ----- ---------- -----------
Contoso Legal Tech     3p             1           0
Microsoft              1p            41           2

discovery-catalog agents

AGENT_COUNT=41

Agent                 Publisher  Party
----                 ---------  -----
aizynthfinder        Microsoft  1p
online-researcher    Microsoft  1p
...

discovery-catalog agents list-tools

AGENT_COUNT=41

Agent                  Publisher  Tools
----                  ---------  -----
aizynthfinder         Microsoft  tools/aizynthfinder
bookshelf-researcher  Microsoft  (no tools)
...

discovery-catalog starter-kits

STARTER_KIT_COUNT=2

Starter-Kit                  Publisher  Category
----                         ---------  --------
drug-discovery               Microsoft  Chemistry
protein-structure-analysis   Microsoft  Biology

discovery-catalog <agent-name> describe

Name           Publisher  Version HasTool Description
----           ---------  ------- ------- -----------
aizynthfinder  Microsoft  1.0.0      True Expert agent for retrosynthetic route planning using AiZynthFinder...

discovery-catalog <starterkit-name> describe

Name            Publisher  Version Category  AgentCount Description
----            ---------  ------- --------  ---------- -----------
drug-discovery  Microsoft  1.0.0   Chemistry          5 Accelerate small-molecule drug discovery...

discovery-catalog <agent-name> list-tools

Agent          Publisher  Version ToolCount Tools         Description
-----          ---------  ------- --------- -----         -----------
aizynthfinder  Microsoft  1.0.0           1 aizynthFinder Expert agent for retrosynthetic route planning using AiZynthFinder...

Examples

Example 1: List publishers

User: "Who publishes agents in this repo?"

pwsh -NoProfile -ExecutionPolicy Bypass -File .github/skills/discovery-catalog/scripts/discovery-catalog.ps1 publishers

Example 2: List all agents

User: "What agents do we have?"

pwsh -NoProfile -ExecutionPolicy Bypass -File .github/skills/discovery-catalog/scripts/discovery-catalog.ps1 agents

Example 3: List agents with their tools

User: "Show me each agent and the tools it ships."

pwsh -NoProfile -ExecutionPolicy Bypass -File .github/skills/discovery-catalog/scripts/discovery-catalog.ps1 agents list-tools

Example 4: List starter kits

User: "What starter kits are available?"

pwsh -NoProfile -ExecutionPolicy Bypass -File .github/skills/discovery-catalog/scripts/discovery-catalog.ps1 starter-kits

Example 5: Describe a specific item

User: "Describe the drug-discovery starter kit"

pwsh -NoProfile -ExecutionPolicy Bypass -File .github/skills/discovery-catalog/scripts/discovery-catalog.ps1 drug-discovery describe

User: "Describe aizynthfinder"

pwsh -NoProfile -ExecutionPolicy Bypass -File .github/skills/discovery-catalog/scripts/discovery-catalog.ps1 aizynthfinder describe

Example 6: Show tools for one agent

User: "List tools for aizynthfinder"

pwsh -NoProfile -ExecutionPolicy Bypass -File .github/skills/discovery-catalog/scripts/discovery-catalog.ps1 aizynthfinder list-tools

Example 7: Filtered queries

User: "Which agents have a retrosynthesis tag?"

pwsh -NoProfile -ExecutionPolicy Bypass -File .github/skills/discovery-catalog/scripts/discovery-catalog.ps1 agents -Tag retrosynthesis

User: "Which agents don't have a tool yet?"

pwsh -NoProfile -ExecutionPolicy Bypass -File .github/skills/discovery-catalog/scripts/discovery-catalog.ps1 agents -WithoutToolsOnly

User: "Which agents are published by Microsoft?"

pwsh -NoProfile -ExecutionPolicy Bypass -File .github/skills/discovery-catalog/scripts/discovery-catalog.ps1 agents -Publisher Microsoft

Example 8: Markdown for a doc

User: "Give me a markdown table of all starter kits I can paste into a README."

pwsh -NoProfile -ExecutionPolicy Bypass -File .github/skills/discovery-catalog/scripts/discovery-catalog.ps1 starter-kits -Format Markdown

Guidelines

  • Read-only: never writes to the repo or to Azure. Safe to run repeatedly.
  • Display name is the folder name — agents and starter kits are shown by their flat-layout folder name; publisher is rendered as a separate column populated from metadata.yaml/kit.json.
  • Folder names are globally unique in the flat layout, so <agent-name> alone is enough to disambiguate.
  • Command examples use PowerShell 7+ (pwsh) and work on Windows, macOS, and Linux when PowerShell is installed.
  • Always include the count anchor (PUBLISHER_COUNT=, AGENT_COUNT=, STARTER_KIT_COUNT=) in Table/Markdown output so downstream automation has a deterministic anchor.
  • Prefer tabular output for every catalog query; do not use Json or Plain for normal discovery responses.
  • Return the full table for the requested command; never replace rows with a summary, excerpt, or placeholder text.
  • Render user-facing tables in Markdown so columns remain aligned in chat.
  • Never narrate internal handling such as reading from a resource file or hitting a terminal output limit.
  • Don't crash on missing metadata — entries with missing metadata.yaml or unparseable kit.json are silently skipped (kits) or shown with empty fields (agents). Missing publisher.name is rendered as (unspecified).
  • Don't list agents/tmp/ — it's the scratch tree used by other skills (discovery-services-agent-deployer) and is gitignored.
  • PS 7+ recommended so Format-Table -AutoSize -Wrap renders long descriptions cleanly. The script works in Windows PowerShell 5.1 too but wrapping is uglier.
  • Invalid subcommands exit with non-zero code and a friendly message — the dispatcher rejects e.g. agents foo or publishers list-tools.

04 同仓库更多 Skills

microsoft
microsoft/discovery

token-estimator-mvp

Estimate AI planning ranges from a work description or explicitly authorized safe project context. Provides coarse class ranges or conditional planning bases without requiring token guesses.

★ 289
microsoft
microsoft/discovery

build-research-paper

Cross-domain / Scientific writing and publication — Build beautifully-formatted, arxiv-ready research paper PDFs from a mix of hand-authored LaTeX and/or markdown + PNG figures, using `tectonic` (default) or `pandoc` (escape hatch). Captures the reusable pattern that research projects keep re-discovering: `report/main.tex` + `references.bib` + `generate_figures.py` + iterative feedback/response scaffolding, all buildable with a single command. WHEN: write a research paper, arxiv submission, prepub, generate PDF from markdown, LaTeX paper, build main.tex, journal manuscript, referee revision, response to referee, resubmission package, publication-quality figures, matplotlib publication style, tectonic build, pandoc build, generate arxiv-ready PDF, bibliography from BibTeX.

内容创作者文档处理写作与文案
★ 289
microsoft
microsoft/discovery

add-example-agents

Add agents from the public microsoft/discovery GitHub repository to this workspace, either (a) locally for the Copilot app by emitting a `.agent.md` companion per agent (so each is invocable via `runSubagent`), (b) into a Microsoft Discovery project via `UpsertAgent` sourced fresh from upstream (binding pre-existing tool ARM resources when present, or building + pushing + registering them in Azure first when missing), or (c) promoting an agent already added to the local Copilot app (sourced from its `.github/agents/<name>.agent.md` companion; tool-bearing agents also carry a `discovery_catalog/<name>/` folder plus a locally-built tool container, while toolless agents have only the companion) up to a specific Microsoft Discovery project + Azure without re-fetching or refreshing from upstream. The user picks exactly one of the three targets per run. USE FOR: 'add the discovery agents', 'download discovery agents', 'install all Microsoft discovery agents', 'add agents from microsoft/discovery', 'create copilot agents from microsoft/discovery', 'deploy discovery agents to Foundry', 'deploy discovery agents to Azure', 'promote a local agent to Microsoft Discovery', 'push my local agent and container to a Discovery project', 'copy a local app agent to Azure'. Always refreshes every selected subfolder under `agents/<name>/` of the upstream repo into `discovery_catalog/<name>/` so the local catalog stays current (new agents appear, removed agents are surfaced, updated content is applied). The presence check for what is 'already locally added' uses `.github/agents/<name>.agent.md` (the Copilot companion), not the `discovery_catalog/` folder. For local deployment, each generated `.agent.md` carries a lazy Docker-bootstrap preamble so the agent builds its tool image on first use only if absent. For Discovery + Azure deployment, pre-flights every declared tool via `GetDiscoveryTool` against the workspace's RG and **skips Docker build / ACR push / AcrPull RBAC entirely when the tool already exists** (common for shared catalog images on registries like `mdqacr`); only builds + registers tools that are actually missing. Always creates the agent in the chosen Discovery project via `UpsertAgent` and polls **`GetAgent` for `provisioningState=Succeeded`** (the MCP `GetAgentOperation` tool returns an opaque error in practice; never rely on it). All MCP calls flow through PowerShell scripts embedded in this SKILL.md and materialized into `.vscode/` during Step 4d — the skill does not depend on sibling helper files or on the VS Code MCP runtime surfacing `mcp_microsoft-dis_*` tools mid-session. Local Discovery runtime registration via `agents action=create` is optional and skipped by default. DO NOT USE FOR: running an already-deployed agent (use runSubagent), or editing a single agent's prompt (edit `agent.yaml` directly).

开发工程师运维与部署图像与多媒体
★ 289
microsoft
microsoft/discovery

migrate-podman-containerfile

Migrate a Podman Containerfile to the canonical Dockerfile required by the Discovery catalog. Use when a PR fails POL-022, a tool contains Containerfile or Dockerfile.in, or a contributor asks to convert Podman build files for Docker. Creates the Dockerfile directly only when the migration is mechanically safe and refuses ambiguous preprocessing or multi-file builds.

★ 289
microsoft
microsoft/discovery

discovery-services-starter-kit-deployer

Deploy Discovery catalog starter-kits end-to-end: validate kit.json agentRefs, build and deploy every referenced agent tool, deploy each referenced agent individually, and summarize deployed agents/tools plus customer-ready sample prompts. Use this skill whenever the user asks to deploy a Discovery starter-kit or deploy all agents listed in starter-kits/*/kit.json.

Skill与智能体开发
★ 289
microsoft
microsoft/discovery

discovery-services-agent-deployer

Deploy Discovery catalog agents end-to-end with one command: discover the agent, build and bind its tool when present, skip tool stages for agents without tools, deploy the agent, and optionally run validation. Use this skill whenever the user asks to deploy a Discovery agent, publish or rebuild an agent tool image, register a Discovery tool from a repo Dockerfile/tool.yaml, resume a failed deployment, or produce a dry-run deployment plan for an agent in this catalog.

人力资源运维与部署Skill与智能体开发
★ 289