glean-run

name	glean_run
description	Discover and run Glean skills for enterprise app tasks
argument-hint	<task description>
allowed-tools	["Read(path=\"//**/glean-skills-cache/**\")"]

Glean Run

Discover and use Glean skills to help with enterprise app tasks (Jira, Slack, Google Workspace, Salesforce, etc.) or actions you don't already have a tool for. Where possible, aim to complete the user's request end-to-end rather than just listing available skills.

Authentication

Authentication is handled exclusively by the setup tool. If any other tool returns a response containing [SETUP_REQUIRED], the user needs to (re-)authenticate via setup.

When this happens:

1. Call setup (no arguments).
   • If no Server URL is configured, setup returns [SETUP_REQUIRED] with instructions. Relay them, ask the user for their work email, then call setup again with email set to what they provided.
   • Once a Server URL is configured, setup opens the Glean sign-in page in the browser and waits for sign-in.
2. Once setup returns "Glean setup is complete", retry the original tool call.

Do not treat [SETUP_REQUIRED] as an error or try to work around it any other way.

Step 0: Verify Setup

Call setup (with no arguments). If the connection isn't ready, setup returns instructions — follow them and call setup again; it guides the whole flow. Once it returns "Glean setup is complete", proceed to Step 1.

Step 1: Plan tool usage

A small set of popular tools is directly available, and no discovery is needed to use them. Discover is complementary and recommended if the direct tools cannot satisfy the user request end to end.

Calling find_skills

If no arguments were provided and the task can't be inferred from conversation context, ask the user what they'd like to do before proceeding.

Call find_skills with the task descriptions.

The first entry in queries MUST be the user's prompt verbatim (the raw, unmodified task description as the user phrased it). This ensures end-to-end skills that match on the overall intent are discovered before you fragment the request into pieces that only match low-level capability skills. After the verbatim prompt, you may append additional atomic sub-tasks broken down from the request.

find_skills({
  queries: [
    "<user's prompt verbatim>",
    "<atomic sub-task 1>",
    "<atomic sub-task 2>"
  ]
})

The response is an XML index of discovered skills with file paths.

You can call find_skills multiple times — e.g. to discover skills for individual sub-tasks as you work through a broad request.

Step 2: Read Skill Instructions

Browse the returned skills and select the one most relevant to the user's request. Read its SKILL.md file for detailed instructions. Skills typically contain guidance on how to use their tools, but the tools can also be called as independent units.

Step 3: Read Tool Schemas

Read each tool's JSON file (e.g. tools/TOOL_NAME.json) to get the exact server_id, name, and inputSchema with parameter names and types.

Never guess parameter names - always read the tool JSON file first.

Step 4: Execute Tools

Call run_tool with the server_id, tool_name (from the name field in the JSON), and arguments matching the inputSchema exactly.

run_tool({
  server_id: "composio/jira-pack",
  tool_name: "jirasearch",
  arguments: { query: "project = PROJ AND status = Open" }
})

Long-form arguments via file_args

For long-form content — drafted Slack messages, Confluence pages, doc bodies, etc. — write the draft to a local file first, then reference it via file_args instead of passing it as a huge inline string. The plugin reads each file and substitutes its UTF-8 contents into the named key in arguments before calling the remote tool.

run_tool({
  server_id: "...",
  tool_name: "slack_post_message",
  arguments: { channel: "C123" },
  file_args: { text: "/tmp/glean-drafts/announce.md" }
})

Constraints:

• Paths must be absolute.
• A key in file_args must not also appear in arguments.
• Each file must be ≤ 1 MB (override via GLEAN_FILE_ARG_MAX_BYTES).

Rules

• Always read tool JSON files before calling run_tool - never guess parameters
• On find_skills, the first queries entry must always be the user's prompt verbatim; only append atomic sub-task decompositions after it
• If discovery returns no relevant skills, tell the user what was searched