Friction
PR #12712 added timeouts + an orchestrator watchdog so a hung miniSummary backfill can no longer wedge maintenance — but that guards the symptom. The deeper friction (raised by @tobiu): why did a mini-summary model call hang for ~4h in the first place?
A mini-summary is cheap by design — e.g. a ~5k-char memory compressed to a tweet-size summary — and should complete in seconds. A multi-hour hang (no error, no return) points at a real root cause worth fixing, not just guarding against.
Hypotheses to investigate
- Model misconfiguration: is the backfill calling the right small/fast model, or a wrong / oversized / not-yet-resident one?
- Provider/endpoint health: the
openAiCompatible endpoint accepting the connection but never responding (vs erroring) — LM Studio model-not-resident, or saturation?
- Prompt/payload: an oversized or malformed prompt the endpoint chokes on silently?
- Concurrency/saturation: was the embedder / KB-sync saturation seen in the same window starving the model endpoint?
Acceptance Criteria
Refs #12711, #12712 (the symptom guard). Recommended for a peer with memory-core / model-provider context — author is staying on the grid lane.
Friction
PR #12712 added timeouts + an orchestrator watchdog so a hung miniSummary backfill can no longer wedge maintenance — but that guards the symptom. The deeper friction (raised by @tobiu): why did a mini-summary model call hang for ~4h in the first place?
A mini-summary is cheap by design — e.g. a ~5k-char memory compressed to a tweet-size summary — and should complete in seconds. A multi-hour hang (no error, no return) points at a real root cause worth fixing, not just guarding against.
Hypotheses to investigate
openAiCompatibleendpoint accepting the connection but never responding (vs erroring) — LM Studio model-not-resident, or saturation?Acceptance Criteria
Refs #12711, #12712 (the symptom guard). Recommended for a peer with memory-core / model-provider context — author is staying on the grid lane.