feat(integrations): add Streamlit UI wrapper for the humanizer skill

[Rich627]

Summary

Adds a minimal Streamlit UI at integrations/streamlit/ that wraps the humanizer skill. Paste text → click Humanize → get the cleaned rewrite back with a native copy button, without the skill's default draft / audit / summary sections.

This is the first non-skill code in the repo. Everything is scoped under integrations/streamlit/ so the skill itself is untouched. If you'd prefer to keep this repo pure-skill, I'm happy to move the wrapper out to a separate repo and link from the root README instead — just say the word.

What the wrapper does

• Shells out to the host's claude CLI (claude -p --no-session-persistence …), so it reuses the user's existing Claude Code OAuth login — no API key required. ANTHROPIC_API_KEY is honored as a fallback.
• Wraps the user's text in a prompt that invokes the humanizer skill but adds strict output constraints so the returned text is just the rewritten prose — no "Draft rewrite" / "What makes this AI" / "Summary of changes" sections.
• Defensive post-process regex (_PREAMBLE_RE + _TRAILING_RE) strips any leftover preamble or skill-trailer block if the model ever disobeys. The trailing regex is anchored to the skill's actual bolded headers (e.g. **Summary of changes:**, **Changes made:**) so legitimate prose that merely starts with "Notes" or "Summary" is preserved.
• Renders the result via st.code(wrap_lines=True) so there's a native copy-to-clipboard icon.

Ways to run

• Local: cd integrations/streamlit && ./run.sh — creates a venv, installs streamlit, opens http://localhost:8501.
• Docker: cd integrations/streamlit && docker compose up --build. The compose file mounts ~/.claude/ into the container (RW, since macOS may need claude /login inside the container), runs Streamlit as a non-root user (uid 1000), and binds port 127.0.0.1:8501 so the unauthenticated UI isn't exposed to the LAN by default.

Files added

```
integrations/streamlit/
├── app.py # Streamlit UI + subprocess call to claude -p
├── run.sh # Local launcher (venv + streamlit)
├── Dockerfile # python + node + claude CLI + streamlit (non-root user)
├── docker-compose.yml # Mounts ~/.claude, loopback-bound :8501
├── .dockerignore
├── requirements.txt # streamlit>=1.56
└── README.md # Lifecycle / auth / stop / output / security docs
```

Notes for review

• Each click spawns a fresh claude -p --no-session-persistence subprocess — no state survives across clicks, nothing is written to the session store.
• Clicking Humanize again kills any still-running subprocess first.
• On Streamlit shutdown (Ctrl+C / docker compose down), an atexit hook (registered exactly once via st.cache_resource, not per click) kills any in-flight subprocess.
• README has a Security considerations section covering the ~/.claude trust boundary, the non-root container user, the loopback port bind, and the single-user scope.

Test plan

• Local: cd integrations/streamlit && ./run.sh, paste AI-sounding text, click Humanize, confirm cleaned output appears with a copy icon.
• Local: confirm Ctrl+C in the terminal cleanly stops Streamlit and kills any in-flight claude subprocess.
• Docker (Linux): docker compose up --build, same flow. (not tested locally — no Linux host available; Dockerfile is identical to the macOS path so this is expected to work, but flagging for reviewer verification.)
• Docker (macOS): after docker compose up -d --build, run docker compose exec humanizer claude /login once, then humanize.
• API-key path: ANTHROPIC_API_KEY=sk-ant-… ./run.sh, confirm no OAuth prompt. (not tested — no spare API key on hand; the code path is a single os.environ pass-through to the claude subprocess, which is the documented override.)
• Regression guard: paste prose beginning with words like "Notes" / "Summary" / "Changes" and confirm it is preserved intact by the cleanup regex.

🤖 Generated with Claude Code