Chronos/DBWaitlist

Changelog / Chronos DB

Chronos DB 0.4.0

Chronos Cloud from your terminal, and a server that's safe by default

Beta9 changesDownloads

Chronos Cloud from your terminal, and a server that's safe by default. chronos init connects a project folder to a hosted Chronos database at chronos.2090.tech in one command; chronos serve stops handing its server token the database's own powers.

Upgrading from 0.3.2: read the breaking change below if you run chronos serve with a token: clients that used it to write main need the new admin token.

Breaking, for 0.4: chronos serve is safe by default. The server token (--token, CHRONOS_TOKEN) now lets a client in as the agent guest: it reads, forks and changes only the guest's own worlds, and can't write main, merge, or manage agents. The database's own user over the network is the admin token (--admin-token, CHRONOS_ADMIN_TOKEN), a separate credential to keep away from agents. Before, the server token was the database's own user, so one copied into an agent's .env wrote main with no world and no review; 0.2 only warned at start. Unchanged: a server with no token at all (loopback only) takes every client as the database's own user, so the quickstart's first create table and insert still work; agents log in with their own tokens; --safe means what it did (every client that isn't an agent or the admin token is the guest, loopback without a token too), so --safe --admin-token A behaves as in 0.3; a role that logs in with its own password (CREATE ROLE ... LOGIN PASSWORD) is that role, with its privileges, in every mode (under --safe in 0.3 it was the guest). Upgrading: a DATABASE_URL or psql login with CHRONOS_TOKEN as the password is the guest now (42501 on writes to main, 403 over HTTP, and the server says the server token is the guest's as it starts): give people and the application that owns the data CHRONOS_ADMIN_TOKEN instead, or pass --unsafe to run exactly as before, with the old warning on an address other machines can reach. --safe --unsafe is refused. POST /v1/merge_outcomes takes the admin token (the server token, now the guest's, gets 403); GET /metrics takes either. The container's example sets both tokens. See upgrading from 0.3.

  • chronos init and chronos cloud: Chronos Cloud from the CLI. In a project folder, chronos init signs you in (a code you approve at chronos.2090.tech/device), finds or makes the project's hosted database (matched by the git remote, so a clone or a teammate gets the same one), and writes .chronos/project.json, the chronos MCP server in .mcp.json, DATABASE_URL in .env (and .env in .gitignore), and a block in CLAUDE.md / AGENTS.md that tells agents to work in a world, test, show the diff, then merge. chronos cloud login | logout | list | usage | production on|off manage the account and the folder's project. See Chronos Cloud.
  • Masked columns: four more ways to read the values are closed. Found by a code review of the masked columns feature. An agent without unmask could read a masked value as it is through a generated column made from it (exposed text generated always as (email) stored), which is now masked too, whether it was there first or added later; through INSERT ... RETURNING the table's whole row (row_to_json(customers), customers.*), now refused as RETURNING email is; and by dropping a policy (DROP POLICY) on a table it owns in a world, or a trigger (DROP TRIGGER, ALTER TRIGGER ... RENAME), now refused as creating them is (rules, which need admin, are checked too). And SELECT ... FOR UPDATE on a masked table quietly locked no rows (FOR UPDATE OF t said t wasn't a table), so two agents' read-modify-write transactions could race: it locks them now, by the table's key. ctid reads on a masked table too.
  • SHOW AGENT ... PRIVILEGES shows a column's own INSERT and UPDATE grants. With only GRANT UPDATE (status) ON orders, the agent may update status, but the table showed as not writable and the column wasn't listed. The column now shows update in its own worlds, and why.
  • Masked columns: agents read a mask, never the value. ALTER TABLE t ALTER COLUMN c SET MASKED [WITH HASH | NULL | PARTIAL]. An agent without the new unmask capability reads a keyed hash of each value (equal values hash alike, across tables, so joins, GROUP BY and DISTINCT still work), NULL, or the last four characters; people read the values. A read becomes a subquery that masks first, so filtering on the value (where email like 'a%') tests the mask, not the value, and row-level security still runs on the values. What can't be masked that way is refused for the agent: an UPDATE or DELETE whose WHERE, SET or RETURNING reads a masked column of its own table, MERGE into the table, changing its definition, search() over it, DIFF ... AS SQL, and the HTTP and MCP row API on it. DIFF and MERGE ... DRY RUN show masked values masked, and constraint errors don't show the key. The mask is the column's: it forks, merges, renames and drops with it. See masked columns. (Asked for on Reddit: "Claude never gets to see it".)
  • SHOW AGENT name PRIVILEGES [IN WORLD w]: what an agent may actually do. A row per table, and one per column that differs, with what the agent may read, write (and where) and how its merges go (on its own, waits for a person, refused), and why: its capabilities, the privileges of a role of its name, row-level security, masked columns and its merge policy put together. An agent may show its own. (Asked for on Reddit: "what matters is the effective permissions of the credential".)
  • The row API's get is checked as a read. In a world that checks its reads, a row read with get (HTTP, MCP or the shell) wasn't recorded at all, so a merge decided on it went through after the row changed, and CHECK READS didn't list it. And a table made without SQL (rows put as JSON) counted as read whole whenever it was named, so a change to any of its rows held the merge. Now get records its row, as a SQL lookup does, and telling that such a table exists reads none of its rows.
  • Cosine and L2 search on vectors of mixed lengths walk their graph at any size. The graph each of those distances gets (0.3.2) was used only when it measured 99% recall on its own vectors, where the inner-product graph is used past the scan limit whatever it measures. On 1M vectors of 768 dimensions of mixed lengths it measured just under, so every search compared all of them: 24 ms (cosine) and 22 ms (L2), against 0.9 ms for the same vectors at unit length. Both now follow the same rule.
  • Prisma works again on a database with a CHECK constraint. pg_constraint's one-letter columns (contype, confdeltype, confupdtype, confmatchtype) came over the wire as text, where Postgres has its one-byte "char" type (OID 18); Prisma's schema engine reads contype only as that, and prisma db push failed (Getting constraint_type from Resultrow ...) as soon as any table had a CHECK constraint. They are "char" now, and so is x::"char".
  • Benchmarks: Chronos vs pgvector side by side, at 1M vectors of 768 and 1,536 dimensions, every distance, filtered and from 8 clients (BENCHMARKS.md, bench/pgvector_vs.py). At the same recall: 4.1–5.8× pgvector's speed unfiltered, 30–100× under a 1% or 10% filter. The losses are listed first: filtered searches from 8 clients scale poorly. Issue #34.

The macOS builds need macOS 13 or later, as for 0.2.0.

Chronos DB 0.4.0 · Chronos