ControlTheory's CTO credits Claude Code with a chat fix built on Dstl8's production logs

Authentication succeeded, so any check on the login path passed ^[4]. The failure showed up one call later. When the client asked to create a chat session, the request reached an upstream API that required user_id and returned HTTP 400 without it ^[4]. Anderson described it to Let's Data Science this way: "Authentication was working OK, but a new required field was now missing in some calls (but not all), which would cause sporadic and occasional problems for some users." ^[3]

"Some calls (but not all)" is the part a repository cannot show. The code shows where connection claims are built and where the session call goes. Which claims arrive empty in production, and what the upstream endpoint now demands, are runtime facts. LDS made the same point: repository access does not necessarily explain the running application, and the agent also needs evidence of the failed request, its conditions and the places worth investigating ^[15]. Dstl8 attached two log excerpts covering that pair. One records a successful authentication with no user ID in the claims. The other records a session-creation request that failed because the field was required ^[5]. I'd hand a new engineer the same two lines.

The rest of the incident record is sensibly built. It holds a summary, a longer explanation, code references, evidence and suggested actions, with related patterns, anomalies and deployment events added for context ^[7]. Its suggested checks covered the connection handling, the upstream endpoint's requirements and whether other organizations were affected ^[6]. That last one is the blast-radius question, and it is good to see a tool raise it unprompted. Engineers can read the record in the interface, paste a formatted copy into an agent, or expose it through the Model Context Protocol ^[8]. The incident hit ControlTheory's own production environment, three days after Dstl8's September 22 launch ^[2][11][2].

Anderson said: "So in this case, the human only needed to ask Claude to look, and the incident provided a clear enough picture for Claude to resolve the issue." ^[9] The interview package included incident details and screenshots. It did not include the patch, the deployment record or before-and-after test results, and LDS has not reproduced the repair ^[10].

Dstl8's closure rule widens the distance between that claim and proof. Dstl8 can close an incident when its symptoms disappear. Anderson listed three ways that happens: a human's code change, an external API that stops returning errors, or infrastructure that recovers without a new application patch ^[13]. An upstream API that starts behaving again closes the ticket as neatly as a merged fix.

The launch announcement cited 328 incidents across 13 Kubernetes clusters at an early fintech customer over two months ^[12]. Anderson said about 70% were closed from Dstl8 ^[14], roughly 230 incidents ^[1]. That figure describes one customer's recoveries. For it to transfer as a repair rate for agents, each closure would need a linked change and evidence that the change, and not an upstream recovery, ended the errors.

LDS treats finding evidence, making a change and establishing that the change repaired the service as separate steps, even when a product presents them as one workflow ^[16]. In my view the third step belongs to the engineer who merges the change. For the September 25 bug it is a short check. After the deploy, connection claims should carry user_id, session-creation requests should stop returning 400, and the drop should line up with the deploy time ^[4].