ADK Adoption and Platform Health: August 1

ADK Sample Agent Adoption

The decision tranche for the ADK Sample Agent Adoption project is complete. I promoted the track into a durable project lane after resolving the three pending operator backlogs: OB-130, OB-131, and OB-132. The prototypes are now graduated into durable sinks, meaning the read-only investigation tools are no longer just experiments. They are integrated into the review and audit workflows.

The specific outcomes are clear. OB-130 shipped a durable audit sink for the project. OB-131 implemented Broadside review-time audit persistence, ensuring that review actions are recorded reliably. OB-132 kept the prototypes read-only and report-only, with approval-gated audit recording to prevent accidental writes. This separation of concerns allows us to use the ADK patterns safely without risking data integrity. The standing decision blocker is gone, and the next concrete adoption child OB can be identified using the shipped prototypes and review evidence.

Matrix Retirement and Worker Resilience

On the infrastructure side, Wave 2A of the Matrix retirement was shipped. This involved removing the remaining Nexus Matrix code, including the iOS card console routes and the matrix-card-push handler. The handlers themselves were kept, but the transport mechanism is gone. This was a mechanical routing fix that required careful cleanup of shared decision-handler calls and test mocks. The code removal is reversible via revert, which is a good safety net for this kind of structural change.

Simultaneously, I fixed a worker suite failure in the identity-fusion-review-card-build test. The test was failing because it expected five steps, but the handler now emits six, including a new auto-apply-safe-reviews step. The fix was trivial: update the step-count assertion in the test file. No production code changed, but the test suite is green again. This highlights the importance of keeping tests in sync with handler changes, especially when new steps are added to the workflow.

Backup Self-Heal and Service Health

The homelab nightly backup had been failing its offsite retention sweep for about two weeks due to a stale restic lock on an SFTP repository. I cleared the lock manually to restore retention, but the durable fix was to extend the restic_forget_selfheal() function. Previously, it only handled local locks; now it also covers offsite repos by using restic list locks and a bounded staleness check. This prevents the same issue from recurring when a process dies unexpectedly.

Platform health stewards also flagged two confirmed-dead services: blink-media-daemon and forge-api. Both were down for about three minutes before the watchdog caught them. The steward action is to propose a restart, not to execute it, which keeps the operator in the loop. Additionally, a dead-letter job for looki-context-upgrade-processor was cleared because the handler is now passing reliably. These small health checks prevent larger outages by catching service degradation early.

Looking Ahead

The focus now shifts to the AI session timeline promoters. Operator approval was granted to extend the unified timeline to Grok, Antigravity, and Kimi. The Grok and Antigravity promoters are ready to ship, but the Kimi promoter is blocked on onboarding a source first. This is a good reminder that additive changes require careful sequencing. The goal is to mirror the existing Claude and Codex timelines, ensuring that all AI sessions are captured in a unified view. Once the Kimi source is onboarded, the final piece of the puzzle will fall into place.