company-knowledge-fetch

Company Knowledge — Fetch Into Context

Goal

Pull documentation, examples, or guidelines from a registered company knowledge source into the current conversation so you can reason over it, generate code from it, or answer questions about it.

---

Step 1: Read Config Silently

Read #file:_superml/config.yml silently.

Check for company_knowledge.sources[]. If the array is empty or missing:

"No company knowledge sources are registered yet. Run company-knowledge-connect to add your first source."

---

Step 2: Identify the Source

If the user specified a source key or topic in the argument, match it against company_knowledge.sources[].key or company_knowledge.sources[].name.

If no argument was given, or no match found, show the available sources:

Available company knowledge sources:

  {i}. {display_name}  [{key}]
       {description}
       Access: {url | mcp}

Which source would you like to fetch? (enter number or key)

⏸️ STOP — wait for the user to select a source.

---

Step 3: Fetch Based on Access Type

Read sources[selected].access.type.

If type: url

A. Ask what to fetch (if the user hasn't already specified a topic):

"What would you like to fetch from {display_name}? (e.g., 'getting started', 'security module', 'auth annotations', 'all examples', or leave blank to load the root endpoint)"

B. Construct the URL

Use access.base_url plus any query params or path segments from the user's topic. Apply the notes hint if present (e.g., append ?format=markdown).

C. Fetch the content

curl -s {auth_flags} "{constructed_url}"

Where {auth_flags} depends on access.auth.type:

• none → no flags
• bearer → -H "Authorization: Bearer {token}"
• basic → -u "{username}:{password}"
• header → -H "{header_name}: {header_value}"

D. Format for context

• If response_format: json → parse and summarise key fields, then quote full JSON
• If response_format: markdown → render as Markdown
• If response_format: html → extract text (strip tags), render as plain text
• If response_format: text → render as-is

If the response is large (>200 lines), summarise the top-level structure first, then ask:

"This source is {n} lines long. Should I load it all, or focus on a specific section?"

If type: mcp

A. Ask what to fetch (if not specified):

"What would you like to know from {display_name}? (e.g., 'show examples for auth module', 'list available components', 'how to use {ComponentName}')"

B. Use MCP tool calls

Address the server using its server_name from config. Example prompts to use (adapt based on user's question):

• @{server_name} get documentation for {topic}
• @{server_name} show examples for {topic}
• @{server_name} list available modules
• @{server_name} search {topic}

If the MCP server name is unknown or the tools aren't responding, suggest:

"Check that {server_name} is listed in .vscode/mcp.json and the server is running. You can verify with @{server_name} help."

---

Step 4: Summarise What Was Loaded

After fetching, briefly summarise what is now in context:

📚 Loaded: {display_name}
   Source: {url or mcp:server_name}
   Topic: {fetched_topic_or_root}
   Content: {n} lines / {n} sections / {brief description}

You can now ask questions about this content, generate code based on it,
or reference it in other skills. For example:
  - "Generate a Spring Boot service using the platform starter pattern"
  - "What annotations does the internal security module provide?"
  - "Show me an example of {framework_concept} from these docs"

---

Step 5: Offer Next Actions

Suggest what the user can do with the loaded knowledge:

"What would you like to do with this?

1. Generate code following these conventions
2. Answer a question about this content
3. Compare this with the existing codebase
4. Add examples to a story or architecture doc
5. Fetch a different topic from the same source"

Proceed with whichever the user selects, or free-form if they ask directly.

---

Notes for AI

• When using URL sources, never log or display Bearer tokens or passwords — only show the constructed URL with auth headers redacted (e.g., Authorization: Bearer ***).
• When content is proprietary (marked with internal or confidential in the response), do not paste it verbatim into git-tracked files. Use it only for AI reasoning in the current session.
• If a URL returns an error (4xx/5xx), show the status code, suggest checking auth config, and offer to re-run company-knowledge-connect to update credentials.