Hivemind — the open-source AI art ecosystem's collective intelligence

A drop-in skill that lets a coding agent (Claude Code, Codex, etc.) search the Banodoco Discord message feed — a public, read-only PostgREST mirror of the Banodoco server. Banodoco is where a lot of the practitioners working on generative video & image tooling (Wan, Wan Animate, VACE, LTX, ComfyUI, Kijai's nodes, SCAIL, InfiniteTalk, training, etc.) actually talk to each other. This skill teaches your agent how to query it.

The endpoint exposes the same data the Discord shows — public messages, attributed by display name. Be respectful when surfacing tips and credit authors.

---

v2: Unified Corpus Protocol

Hivemind v2 upgrades the raw message feed into a unified knowledge corpus with a read path (public PostgREST against unified_feed) and a locked write path (Supabase edge function). The corpus combines:

• Messages — the original Discord message feed (v1)
• Resources — articles, transcripts, ComfyUI workflows, blog posts
• Distillations — curated Q&A pairs with cited sources, submitted by agents and reviewed by humans

Everything is searchable: the pack's search executor queries the raw corpus tables (message_feed, external_resources, distillations) in parallel with per-token ILIKE predicates and ranks the merge client-side — the unified_feed UNION view is used only for kind-scoped single-row fetches (get_item), never for text search (its derived-view scan blows the anon role's 3s statement budget → HTTP 500 / SQLSTATE 57014). Distillations make the corpus self-improving: every researched answer becomes a permanent, findable entry for the next person.

---

Install targets

1. Claude Code skill (the classic)

git clone https://github.com/banodoco/hivemind /tmp/hivemind-skill-tmp \
  && mkdir -p ~/.claude/skills \
  && rm -rf ~/.claude/skills/hivemind \
  && cp -r /tmp/hivemind-skill-tmp/skill ~/.claude/skills/hivemind \
  && rm -rf /tmp/hivemind-skill-tmp \
  && echo "Installed. Restart Claude Code, then try: /hivemind"

Or run bash install.sh from a clone of this repo.

2. Astrid pack (executor-based)

python3 -m astrid packs install https://github.com/banodoco/hivemind.git

The pack exposes seven executors: hivemind.search, hivemind.get_item, hivemind.refresh_media, hivemind.contribute, hivemind.ingest_article, hivemind.ingest_workflow, hivemind.ingest_youtube. See AGENTS.md for the agent guide. (Requires Astrid with external Python-executor pack support, 2026-06-04+.)

3. Codex / any agent (instruction-file copy)

Copy skill/SKILL.md into your AGENTS.md (or equivalent instruction file) — the content is self-contained with endpoint, schema, query patterns, and the full contribute API.

4. Pip package (editable dev install)

The repo root is a Python package (hivemind), stdlib-only. Install editable so edits to the clone are live:

pip install -e .
pyenv rehash   # only if you use pyenv — exposes the hivemind-search shim

Then, from any directory:

python3 -m hivemind.executors.search.run --query "wan animate" --limit 10
hivemind-search --query "lora" --channel wan_chatter --limit 20   # console script
python3 -c "import hivemind"                                       # package import

The console script mirrors python3 executors/search/run.py exactly (same flags, same JSON/stdout contract, same stderr paging hints).

Repository layout

One repo, three products — the Astrid pack is contractually pinned to the repo root (astrid packs install <git-url> requires pack.yaml in a directory whose name equals the pack id, and the clone is named hivemind):

entry	belongs to	what it is
skill/SKILL.md	all agents	The canonical playbook — installed as the Claude skill, discovered by Astrid, copy-paste for anything else
pack.yaml, executors/, AGENTS.md, __init__.py	Astrid pack	Manifest, seven stdlib-only executors, agent guide, package marker for hivemind.executors.* imports
schema/, supabase/	backend	The corpus DDL and the contribute edge function (the only write path)
scripts/	ops	Contributor-key issuance
tests/	dev	309 Python unit tests (mocked HTTP) + deno tests under supabase/
install.sh, assets/	repo	Claude-skill installer, mascot
DESIGN.md	docs	Architecture: layers, flywheel, deferred decisions
.astridignore	Astrid pack	Keeps backend/assets/tests out of installed pack copies

---

What you get

Read path

• The endpoint URL + the public anon key (safe to commit — it's the publishable key, RLS makes it read-only).
• The unified_feed view: messages + resources + distillations in one table (use kind-scoped for single-row fetches — NOT for text search).
• Per-token ilike search over the raw tables (message_feed content, external_resources title/body, distillations question/answer/conditions), client-ranked and merged.
• Get single items by kind + id with full citation context.
• A taxonomy of which Discord channels are high-signal (daily_summaries, wan_chatter, wan_comfyui, ltx_chatter, comfyui, *_resources) vs. background noise.
• A short list of power-users to weight (Kijai, Ablejones, djbfilmz, 42hub, BNDC the summary bot).
• PostgREST query patterns the agent can use directly: ilike substring search, in.(…) channel filtering, repeated-key AND, or=(…) for spelling variants, author + time filtering.
• Gotchas — most importantly that fts (full-text search) times out on this table, so use ilike only.

Write path (contribute API)

• POST /functions/v1/contribute — the single write entrypoint.
• Auth via X-Contributor-Key: hm_<64 hex> header (SHA-256 checked against the contributors table, revoked keys rejected).
• Two actions:
  • add_resource — submit articles, transcripts, workflows.
  • submit_distillation — submit Q&A with ≥1 cited source. Duplicate detection via pg_trgm similarity (>0.6 threshold), supersedes_id for replacing outdated answers.
• Status forced to pending by the edge function; curators promote to approved.
• Responses: 201 (created), 400 (validation), 401 (unauthorized), 409 (duplicate), 500 (internal error).

Flywheel loop

1. Search distillations first on the user's question.
2. Hit → relay the answer with its cites.
3. Miss → research raw messages/resources, answer the human, then submit a cited distillation.
4. The next person who asks gets it immediately.

Worth-it criteria for distillations

Before submitting, check:

• The question is generalizable (not a one-off personal request).
• You did real research effort (surfaced sources, compared answers).
• You have at least one cite.
• If a similar question exists, supersede it rather than duplicating.

Contribute curl example

curl -s -X POST "$SUPABASE_URL/functions/v1/contribute" \
  -H "Content-Type: application/json" \
  -H "X-Contributor-Key: hm_$(cat ~/.hivemind/key)" \
  -d '{
    "action": "submit_distillation",
    "data": {
      "question": "What is the best upscale model for anime-style Wan output?",
      "answer": "4x-UltraSharp with 50% blend on original, then a second GFPGAN pass.",
      "confidence": "high",
      "conditions": "for anime-style video, 1080p target",
      "cites": [
        {"item_kind": "message", "item_id": 88123},
        {"item_kind": "resource", "item_id": 17}
      ]
    }
  }'

---

Executor inventory

Executor	CLI	Description
search	python3 executors/search/run.py --query "..."	Per-token ilike search over raw tables, client-ranked; never unified_feed

Filters: --kinds message|workflow|distillation (or a mix), --sources, --since, --channel <name>, --author <name> (messages only), and --thread <snowflake> (index-backed thread surface). Ordering: --sort relevance (default — score, then recency) or --sort recent (created_at desc, then score). The relevance score: +5 per distinctive token in title/question, +3 in body/answer/conditions, +4 approved distillation, +3 parseable workflow, +2 exact phrase. Page any query with --limit N --offset M. Ergonomics: every response carries count, total (ranked-pool extent), has_more, page, pages, and next_offset — when has_more is true, the next page is exactly --offset <next_offset> on the same command, so agents never reconstruct the offset. A human summary goes to stderr so stdout stays pure JSON: Showing 11-20 of 114 results (page 2 of 12) - next: --limit 10 --offset 20. total is the ranked pool bounded by the per-scope fetch (exact corpus counts would need the slow Prefer: count=exact, which the transport never sends). The ranking is deterministic while the corpus is unchanged, so page N is stable across calls. Examples:

python3 executors/search/run.py --query "wan animate" --limit 10                 # first page, all kinds
python3 executors/search/run.py --query "wan animate" --limit 10 --offset 10     # second page
python3 executors/search/run.py --query "lora" --channel wan_chatter --limit 20  # channel filter
python3 executors/search/run.py --query "lora" --author Kijai --limit 20         # author filter
python3 executors/search/run.py --query "context" --thread 1175229360781938718   # thread filter

| get_item | python3 executors/get_item/run.py --kind distillation --id 42 | Full untruncated row with citation context | | refresh_media | python3 executors/refresh_media/run.py --message-id 1512127379039060118 | Refresh expiring Discord CDN attachment URLs | | contribute | python3 executors/contribute/run.py --type resource ... | Submit resources or distillations via edge function | | ingest_article | python3 executors/ingest_article/run.py --url https://... | Extract HTML text → submit as resource | | ingest_workflow | python3 executors/ingest_workflow/run.py --path workflow.json | Parse ComfyUI JSON → extract models → submit | | ingest_youtube | python3 executors/ingest_youtube/run.py --url https://... | yt-dlp captions → submit transcript resource |

All executors are stdlib-only Python. See AGENTS.md for usage guidance and DESIGN.md for architecture details.

---

Full Dataset on Huggingface

Hivemind is for querying the live public message feed from an agent. If you want to train on the full archive or grab the whole dataset directly, use the Hugging Face dataset:

https://huggingface.co/datasets/Banodoco/discord-archive

That dataset contains the exported Discord archive with opted-out authors excluded.

---

Example prompts that will trigger the skill

• "What does Banodoco say about Wan Animate best practices?"
• "Search Banodoco for SCAIL vs Wan Animate"
• "What settings has Kijai recommended for the lightx2v LoRA?"
• "Find me workflows for long-video context windows in Wan"
• "What did people say about LTX 2.3 last week?"

---

Endpoint

GET https://ujlwuvkrxlvoswwkerdf.supabase.co/rest/v1/message_feed
Header: apikey: sb_publishable_O38oPBafrBoFrpi_rlWJvA_UJrulFsx

The unified_feed view is at the same base (kind-scoped single-row fetches only — see the search-executor note above):

GET https://ujlwuvkrxlvoswwkerdf.supabase.co/rest/v1/unified_feed

Smoke-test a raw search surface (per-token OR on message content — the fast shape; order=created_at.desc bounds the pool):

curl -s "https://ujlwuvkrxlvoswwkerdf.supabase.co/rest/v1/message_feed?select=message_id,content,author_name,channel_name,created_at&limit=5&order=created_at.desc&or=(content.ilike.*wan*,content.ilike.*animate*)" \
  -H "apikey: sb_publishable_O38oPBafrBoFrpi_rlWJvA_UJrulFsx" | python3 -m json.tool

---

Further reading

• hivemind/SKILL.md — full raw query playbook: channel map, power users, search snippets, trend questions, caveats.
• AGENTS.md — agent guide: when to use each executor, flywheel loop, key constraints.
• DESIGN.md — architecture, schema design, lifecycle, deferred work, design decisions.
• skill/SKILL.md — Astrid skill documentation: read/write paths, contribute API, curl examples.

---

License

MIT. The data the endpoint serves is public Discord content authored by Banodoco members — credit them when you surface their tips.