Assistant v0.6.3 — governance that can't be bypassed, streams that can't hang
A middleware that vanished, a stream that hung forever when blocked — and the hotfix release that made governance terminal instead of optional.
Two bugs walk into a release. One is the kind you find in a security review, the other the kind you find because an agent froze mid-conversation and the only way out was killing the process.
v0.6.3 fixes both, and they’re the same bug wearing different coats: governance that could be bypassed.
The blocked stream that hung forever
The assistant streams responses from the model. Governance rules (budgets, tool-call limits) can block a call — but blocking a request that’s already streaming is a special kind of problem: if the block doesn’t terminate the stream, the user sees a response that never ends. The agent sits there, spinner spinning, budget already spent, no way forward. Fix: blocked streamed calls are now terminal (issue #19). When governance says no, the stream ends — cleanly, visibly, in a way the runtime can act on.
The middleware that wasn’t there
During the same pass, the test suite told us something worse: the governance middleware had lost its registration (issue #20). The rules still existed — they just weren’t enforced, because nothing stood in the request path. A governance layer that can silently unregister is a governance layer you can’t rely on; the fix is a regression test that proves the middleware is actually on the path, every release, forever.
The pattern matters beyond assistant: a control that can be dropped without anyone noticing is worse than no control — because you’ll build on it. v0.6.3 is the release where the controls prove they’re still there.
The series continues
Since v0.6.3 the repo’s changelog already carries:
- v0.6.4 — the determinism release: shell output spill reachable by sandbox truncation flag, and a full test-suite determinism pass (including a websocket disconnect hang class fixed by test, not by luck)
- v0.6.5/v0.6.6 — observability with a hard split: semantic telemetry to Langfuse, operational telemetry to a filtered OTLP/ClickStack endpoint — with export filtering that excludes prompts, tool arguments, SQL, command text, and request content. A real no-op without
OTEL_ENDPOINT. Full suite: 2,885 passed, 26 skipped, and the A/B telemetry benchmark shows the instrumentation costs p50 +3.1%, startup −0.020s. - v0.6.7 — native tools become a fail-closed allowlist (
all | selected | none) enforced across six code paths
Ship the controls, prove they ship (tests, terminal behavior), and only then add the telemetry that watches them. That ordering is the release.
Assistant is the flagship of the Open Assistants stack — the app that runs on CoreMem, HybridDB, ConnectKit, and ModelPanel — see the product page, or run it in one container: Docker deployment.