2026.10.24 We Upgraded Mastra, Fixed Maestro's Memory, and Locked Down RDS TLS
Mastra 1.79 went to production and broke Maestro's working memory with a TLS error. We fixed it forward the same day, then verified RDS TLS on every Postgres client in the repo.
Behind the scenes week. Forty PRs merged in the ChameleonLabs repo, and about thirty of them were Dependabot. The other ten are the interesting part, so that's where we'll focus.
Mastra 1.79, and the memory bug it shipped
We'd been holding the Mastra upgrade back for a while (it changes the production schema, so it couldn't ride along with a routine dependency bump). This week we did it on purpose: @mastra/core 1.4 to 1.79 and @mastra/memory 1.3 to 1.39, with init disabled in prod and a written DDL plan handed off to the admin side. It also cleared a Snyk critical on gray-matter, which is no longer in the lockfile at all.
Then the 0.13.0 build deployed at 17:49 CT and the first Maestro turn logged three failures: fetching working memory, fetching observations and saving messages, each complaining about a self-signed certificate in the chain. Chat kept answering (memory context is non-fatal by design), but Maestro had amnesia for that stretch. The new Mastra pools weren't trusting the RDS certificate chain. The fix-forward, 0.13.1, embedded the Amazon RDS us-east-1 CA bundle and turned full TLS verification on for those pools.
That sent us looking at every other Postgres client in the repo, and several were skipping verification too. So 0.13.2 finished the job. There's now one shared helper, rdsPoolConnection(url), every client goes through it, and rejectUnauthorized: false is gone from lib, app, scripts and infra. We added a Node tests job to CI so it stays that way.
Security plumbing
The October Node CVEs exposed two things we didn't love: the repo declared no Node version at all, and the scanning layer had quietly been switched off. Both are fixed — a patched Node floor in engines, free-first scanning with a delta-aware hook, Dependabot configured and CI actions pinned.
With scanning back on, we cleared the fixable high/critical backlog so the npm-audit and osv-scanner jobs can be made blocking. Direct dependencies got bumped to the lowest fixed version. Transitive ones whose parent range excluded the fix are pinned through overrides, and every pin is written down with its advisory in Docs/Developer/dependency-overrides.md. (Pins without a paper trail are how you end up stuck on an old version two years later with nobody remembering why.)
Then the sweep itself: 13 security-group updates across 7 directories, aws-cdk-lib 2.238 to 2.260, fast-xml-parser removed from the ai-stream Lambda entirely, and the httpx, pytest, boto3 and anthropic ranges bumped in every one of our Python agent directories. That's where most of the forty came from.
One candid note. Earlier this week we wrote in AGENTS.md and the Dependabot config header that our semver-major ignore also suppressed security updates. Then we checked, and it doesn't. Those ignores only gate version updates. We corrected both places. Checking Dependabot is now an explicit step in opening a PR instead of something that happens when somebody remembers.
Two smaller fixes we're happy about
Notification preferences no longer use per-send unsubscribe tokens. The page resolves you from your signed-in console session, so there's no capability secret sitting in email bodies or Discord messages, and forwarding a notification doesn't hand the recipient control of your settings.
And our fleet agents now refresh their GitHub tokens from SSM when they hit an auth failure. After the October 17 PAT rotation, two of them failed every git fetch with a 401 and a third silently stopped applying needs-triage labels, all because each one read its token once at startup and kept it forever. We restarted them by hand that time. Next rotation, they fix themselves.
Have you been bitten by a cached secret like that? We'd like to hear how you handled it.