Skip to main content
The typed repos are the storage protocol. MemoryStores and SqliteStores ship in core; SupabaseStores (behind the [supabase] extra) binds the same repos to a Supabase / PostgREST project. 
pip install "metalworks[supabase]"

​
Bootstrap the tables

SupabaseStores ships a schema you apply once. Run SCHEMA_SQL against your project (SQL editor or migration): 
from metalworks.stores.supabase import SCHEMA_SQL
print(SCHEMA_SQL)  # metalworks_briefs, metalworks_runs, metalworks_posts, ...

​
Use it

from metalworks.stores.supabase import SupabaseStores

store = SupabaseStores(url="https://xyz.supabase.co", key="<service-role-key>")
# or pass an existing supabase-py client: SupabaseStores(client=my_client)
SupabaseStores satisfies every repo protocol, so it drops into ResearchDeps or the Reddit OAuth account store exactly like MemoryStores.

​
Binding to pre-existing tables

If you already have columnar tables (this is how Clique migrates onto metalworks), pass a table_map to point the logical collections at your physical table names: 
store = SupabaseStores(
    url=..., key=...,
    table_map={
        "briefs": "research_briefs",
        "posts": "reddit_posts_research",
        "comments": "reddit_comments_research",
    },
)

​
The pagination guarantee

Every read paginates with .range() until exhaustion. PostgREST silently caps result sets at its max-rows setting (default 1000) and returns HTTP 200, which would otherwise truncate corpus reads and corrupt every downstream count. The conformance suite’s >1000 rows behind one filter case exists to catch any regression of this: 
from metalworks.testing import check_all_repos
check_all_repos(store, corpus_rows=1500)  # against a scratch project