[arcaven]

Fleet seats read the shared bd store over the network on every look at their queue, so a read depends on the global server being up and near. This records an idea for a per-cluster local read cache with writes still going to the one global server, so the next probe starts from a stated shape and three named open questions instead of a conversation.

One new file, _kos/ideas/local-read-cache-of-the-shared-bd-store.md. It is an idea only: no probe, no design, no commitment. It cites Dolt's read-only server and read-replica options and bd's --readonly flag as context, links the two existing callbook ideas it touches, and keeps three points open: lag tolerance, what a write does when the global server is unreachable, and how a seat knows its cache is stale.

Additive only; nothing else in the repo changes.

Refs: I-147

[arcavenai]

Approve at f3c2398. Every cited point checks out against its source. Both links resolve. The three hard questions are kept open, and the file stays an idea.

Checks

1. Cited context, each against its source.
   • Dolt read-only. The Dolt docs (dolthub/docs-2 at 725642b8, the source of the dolthub.com docs) confirm all three:
     • dolt sql-server documents -r, --readonly: "Disable modification of the database." (reference/cli/cli.md).
     • behavior.read_only is documented as turning the server "into read only mode, preventing any write queries from succeeding" (reference/sql/server/configuration.md).
   • Dolt replication. reference/sql/server/replication.md, the page the file cites, covers both forms:
     • remote-based read replicas that pull on read, through @@dolt_read_replica_remote;
     • direct-to-standby replication, where a standby "will be accept read requests and writes replicated from the primary".
   • bd 1.1.2 --readonly. The beads release notes record feat: add --readonly flag for worker sandboxes in v0.29.0 (2025-12-04). That is before v1.1.2 (2026-07-26), and no later release note removes the flag, so 1.1.2 had it, and the "worker sandboxes" wording at :37-38 matches.
2. Links. conflict-free-versioned-db-substrate-landscape.md and claim-liveness-and-agent-reachability.md both exist in _kos/ideas/ at the head, beside this file.
3. Open questions. Each of these stays a question, none answered by assertion:
   • lag tolerance (:59-61);
   • a write while the global server is unreachable (:62-64), with three options listed and none chosen;
   • staleness (:65-67).
4. Idea language.
   • The PR adds only this file, under _kos/ideas/.
   • :3-4 says "No probe and no design", and :57 says "None is decided here".
   • "Will", "must" and "shall" do not appear.
5. Public text.
   • No figures from records, no host names and no private repo names.
   • The disclosure scan from aae-orc main is clean on the file and the PR body. On a control line it flagged an em dash and a banned word, so the scan was working.
6. Text.
   • The commit is signed and GitHub-verified, with one parent.
   • The file and the body are ASCII, with no em dash, no banned words and no attribution.
   • The body opens with the why and has one cost-of-merge line.
   • The trailer is Refs only.

Note for the probe

bd's own read-only mode stops short of a server. The v1.3.0 release notes say "bd --readonly serve is refused". So a cache built on what bd offers today would sit at the Dolt layer (a read-only sql-server or a read replica), not at a read-only bd server. This is a note, not a defect; the file does not claim otherwise.

Agent review: host mokuzai, team reviewer, seat reviewer-reviewer-claude-g17-2, agent reviewer-claude.