[titanwings]

Requested by Tianyi · project thread

Stacked on #197 and #201, and already merged with dot-skill (including #202). Until #197 and #201 merge, this PR also shows their commits. Only the last commit, feat: discover the owner's existing sources at profile setup, is new here.

Summary

Before: the owner flow described a fixed set of source types (calendar, mail, docs, code), and the agent had to work out on its own whether it could reach any of them.

After: at setup, the agent first takes stock of what it can already reach:

• the tools, connectors, MCP servers and memory in its own session;
• the platform it runs on (for example Grok and the owner's X account);
• what tools/owner_sources.py discover finds on the machine: host user and memory files, configured MCP servers, related skills, and personal exports in ~/Downloads.

It then proposes one small, read-only slice per source and asks the owner which sources to use and whether to add anything else. Nothing is read until the owner picks it.

Changes

• tools/owner_sources.py (new):
  • discover reports local sources by name only. It never prints commands, args, env values or tokens, and never opens exports.
  • It reads the config locations of Claude Code, Claude Desktop, Codex, Hermes and Cursor, plus OpenClaw's workspace and skills.
  • classify maps any source or tool name to a category hint (mail, calendar, social, docs, code, chat, media, ai_memory) with the slice to read and what it teaches. Health and finance are flagged as sensitive, and unmatched names come back as unknown for the agent to judge from the tool description.
• prompts/owner/intake.md: new Step 0 (discover), Step 1 (propose a plan and ask "which ones, anything else?") and Step 2 (read only the approved sources). The interview then asks only what the sources did not answer.
• prompts/owner/profile_builder.md: source labels are now <category>: <name>.
• SKILL.md: the owner flow describes discovery instead of a fixed list, in both languages.
• tests/test_owner_sources.py: 8 unit tests, including one checking that config secrets never appear in the report.
• tests/e2e/owner_discovery/ and docs/examples/owner-discovery-e2e.md: a harness and a recorded run in a real Claude Code CLI.

Motivation

Personal agents already hold the owner's connections and memory, and which ones differ from host to host. Discovering them works across hosts without maintaining a list of every app.

Testing

• python -m unittest discover -s tests -p 'test_*.py' passed (109 tests)
• python -m compileall tools/ passed
• node bin/distilly.mjs --check-package passed
• Manually tested end to end in Claude Code 2.1.296. I used a fresh HOME with mock gmail, google-calendar, x and weather MCP connectors serving fictional data, plus an X archive in ~/Downloads. See docs/examples/owner-discovery-e2e.md.
  • The agent found all four connectors and the archive, and skipped weather.
  • It asked which sources to use and read only the chosen ones: sent mail only, the past four weeks of calendar, and the owner's own posts.
  • It never called the send or post tools.
  • It wrote the profile with a source per fact, kept the existing CLAUDE.md line and made a backup.
  • A new session then used the profile correctly.

Not covered: real Gmail, Calendar or X connectors, and cloud agents such as Grok Bot or OpenAI Dots. These use the same prompt but were not run here.

Checklist

• I read CONTRIBUTING.md
• Docs updated (SKILL.md, owner prompts, example)
• No secrets, tokens, or personal data committed (test data is fictional)
• No new dependencies
• Tests added

🤖 Generated with Claude Code

https://claude.ai/code/session_01Y32jGbUW3JKCyeeySUBm27

---

Generated by Claude Code