Michael NguyenandClaude Opus 5.5 f3d9f141ba fix(tool-gateway): read policy context once per tool listing ...

A tools/list request decided access for each tool in the company with a
separate policy decision. Each decision read about 13 tables with
select *, including the full heartbeat run row and its large context
snapshot. With 900 tools, one listing made about 11,800 queries and
used about 2 GB of heap. Several listings at run start crashed the
server with out-of-memory errors.

- loadContext selects only the columns that a decision uses. It reads
  the run issue, project and routine ids from context_snapshot in SQL.
- decide accepts an optional cache for one batch of decisions. A
  listing reads the shared inputs one time: agent, run, profiles,
  profile entries, policies, grants, connections, applications and the
  company catalog. Rate-limit counters are not cached, and a decision
  that consumes a rate limit does not use the cache.
- The gateway makes one cache for each listing and runs a maximum of
  16 decisions at the same time.
- connectedMcpToolsForCompany reads each connection row one time, not
  one time for each tool. Its two reads share one read-only snapshot.

Access decisions do not change.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

2026-10-01 09:57:24 -07:00

..

scripts

fix: attribute Sentry errors to the loaded source release (#13719)

2026-09-20 08:08:33 -07:00

src

fix(tool-gateway): read policy context once per tool listing

2026-10-01 09:57:24 -07:00

test-baselines

fix: use persisted state for lifecycle continuation and retry budgets (#13888)

2026-09-25 15:28:11 -07:00

CHANGELOG.md

fix: bound workspace Git scans (#11572)

2026-08-17 22:11:30 -05:00

package.json

build(deps): bump multer from 2.2.0 to 2.4.0 (#14493)

2026-10-01 15:00:48 +00:00

tsconfig.json

Fix root TypeScript solution config

2026-03-09 14:09:30 -05:00

vitest.config.ts

fix(server): bundle the vendored paperclip-runner instead of hand-mirroring its deps (#13121)

2026-09-09 16:38:17 -07:00