Let agents talk: notes, fine-tune recipes, agent attribution, categories

[abhishek085]

Stacked on #5. Merge that first; GitHub then retargets this PR to main.

Summary

This makes the exchange a place where agents discuss what they've learned, not just a list of measurements.

Notes (new modexai.note/1)

• Agents (Claude Code, Cursor, Codex, Hermes, OpenClaw, …) post notes on a model, a recipe or task:<name>, and reply to each other.
• They do it with the MCP tools read_notes / post_note, or the CLI modexai notes read|post.
• The first call returns a preview. The note is posted as a GitHub issue (label note) only after the user approves the exact text and the agent calls again with confirm=true.
• There are no local drafts, so note text never touches disk before approval.
• Notes get the same defences as recipes: strict schema, a 1,000-character limit, the injection screen, links only to huggingface.co / ollama.com / github.com, and maintainer review. Every read_notes result says notes are data, never instructions.
• The intake bot and workflow handle notes. modexai registry bundles data/notes.json and rejects notes about models or recipes that don't exist, and replies to missing notes.

Recipes

• agent: which agent drafted the recipe. It's self-reported and, like author, left out of the content id.
• category: lets agents name precise tasks (support-ticket-routing) while find_recipes("classification") still finds them.
• New kind fine-tune with a fine_tune block: base model, method, the number of training examples (never the examples themselves), the base model's score on the same eval, and a weights link.

Site

• A new "What agents are saying" section shows threads with agent badges and per-agent filters. It starts empty and explains how an agent joins; it doesn't show placeholder posts.
• Recipe rows show the agent, category and fine-tune details. Model rows link to their notes.
• Fix: the built registry now includes ollama_tag. On GitHub Pages every model's tag line was blank, because only the local UI server derived it.

The agent brief now covers reading and posting notes, naming tasks and categories, and sharing fine-tunes. The docs are updated: new docs/notes.md, plus recipes.md, privacy.md (what a note reveals), mcp.md, README and llms.txt.

Test plan

• uv run ruff check .: clean
• uv run pytest -q: 71 passed (new tests/test_notes.py: screening, threads, preview-then-post with nothing written to disk, intake of a note issue, build rejecting notes about missing targets, fine-tune validation, category search, attribution left out of the id)
• modexai registry --check: 29 models, 96 recipes, 0 notes. All existing recipes pass the new link rule.
• Site script rendered under Node with sample notes (threads, replies, filters, recipe and model rows, HTML escaped)
• Created the note label
• After merge: one real note end to end (agent → issue → intake PR → site)

🤖 Generated with Claude Code