[blove]

Summary

Follow-up to #880 (client-provided AG-UI tools): the retained tool-call record now covers every model tool call on an AG-UI run where the store is resolved, not only client-provided ones. Spec: docs/superpowers/specs/2026-10-02-tool-call-record-generalization-design.md; plan: docs/superpowers/plans/2026-10-02-tool-call-record-generalization.md.

• Server rows, identity only. Every server tool call is recorded (kind: "server": thread, route, run, tool name, issued and settled times; no result text) by the LangChain tool converter around the tool body: issue before it runs, settle in finally. A permission park (GraphInterrupt) does not settle. The client stub stays the only writer of client rows, recognised by a new clientTool: true marker on B4ToolDefinition. The parent task call is not recorded.
• Scope. Recording happens only when a store resolves: a route in server.agui.clientTools, server.agui.clientToolStore set, or the default .b4/client-tool-calls.sqlite still present from an earlier opt-in. Apps with no store and non-AG-UI transports are unchanged.
• The forgery guarantee, stated in server state. A role: "tool" message is consumed only when it names an open client row this server issued on this thread. Messages naming a server row, a closed row, or nothing are dropped as history, as before. The message filters are not relaxed.
• pendingToolCallIds from the record. Read from the store when RUN_FINISHED is built, scoped to this run's parks (live), the snapshot's parks (partial), or none (no-op path). The stream normalizer's park callback and the in-memory set are gone. @b4run/ag-ui's pendingToolCallIds option may return a Promise; a rejection ends the run as RUN_ERROR.
• Pruning: reconciled with feat: prune settled client tool call records #898. feat: prune settled client tool call records #898 landed its own prune while this was in flight (global prune({ before }), server.agui.clientToolRetentionMs, hourly sweep on turn settle, b4 client-tools prune, TTL floor on the cutoff). This PR adopts that design unchanged and only extends its predicate: a server row is swept once settledAt is behind the window; open rows of either kind are never deleted. The per-thread prune and toolCallRetentionMs from the original spec were dropped at merge (spec §4 note).
• Store contract. ClientToolCallRecord gains kind and settledAt; ClientToolCallStore gains settle (plus feat: prune settled client tool call records #898's prune); ClientToolRecorder gains issue and settle. Memory, SQLite and Postgres stores implement them; SQLite and Postgres append migration 2 (kind text NOT NULL DEFAULT 'client' CHECK (...), settled_at text); existing rows backfill as client. An operator store missing settle/prune fails the boot naming them.

Behavior changes and risks

• Blast radius. On an app with a store, a store write failure in issue fails that server tool call, on every AG-UI route. With the store unavailable, server tool calls on AG-UI runs fail until it is back. Previously only client-tool runs depended on the store.
• Load. Two extra store writes per server tool call, one read on runs that parked.
• Rolling upgrades on a shared Postgres store. A replica on the previous version has no kind filter and reads new server rows as open client rows (it may void them). Nothing becomes answerable; finish the rollout before mixing traffic.
• Migration 2 adds two columns; on Postgres it is a metadata-only ALTER.

Test plan

• After merging main (feat: prune settled client tool call records #898/feat(ag-ui)!: serve the AG-UI HTTP+protobuf binding and advertise httpBinary, reasoning and state #906/feat: prune settled approval grant records #907): pnpm lint && pnpm build && pnpm typecheck && pnpm test green — 9621 passed, 287 skipped, 0 failed.
• Postgres store suite run locally against Docker (B4_TEST_PGSTORAGE=1), including the v1→v2 backfill through runMigrations; the gated postgres-storage-docker lane covers it in CI.
• node scripts/check-docs.mjs, node scripts/check-changesets.mjs, pnpm --dir apps/web seo:lastmod:check pass.
• Each task reviewed for spec compliance and code quality by independent subagents; the prune-vs-resumed-replay race and the stray-row leak into pendingToolCallIds were caught in review and fixed with regression tests. The merge resolution was reviewed separately.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
b4-run	Ready	Preview	Oct 3, 2026 2:19pm UTC