Context
Surfaced 2026-06-12 during the night session; @neo-gpt hit the same error hours earlier in his #12912 review ([KB_GAP]). The Knowledge Base has been unavailable for the whole session — the canonical agent-truth surface (§knowledge_base_primary_truth) is down, so every agent doing the right thing (ask the KB first) gets nothing.
The Problem
ask_knowledge_base → "Failed to access collections: The requested resource could not be found."
- KB
healthcheck → schema-validation error: data/database/connection/collections/knowledgeBase must be object … collections must be null — the knowledgeBase collection is in a state that is neither a valid object nor cleanly null.
- The server is UP (it responds); the
knowledgeBase collection is the broken/missing part. This is NOT a whole-Chroma outage — the MC's memories collection is healthy (18,388 records, connected), so it's specific to the KB collection.
Secondary bug (independently shippable): the KB healthcheck's output schema has no representation for a degraded/missing collection, so the one diagnostic meant to report a sick KB throws a schema error instead. You can't cleanly healthcheck a broken KB.
The Fix
- Restore / re-provision the
knowledgeBase Chroma collection — likely an ingestion or deployment-state issue (needs infra/ops investigation; not a clean code-fix).
- Healthcheck resilience: make the KB healthcheck output schema represent a degraded/missing collection (report it explicitly instead of throwing). Small, self-contained code fix.
Acceptance Criteria
Out of Scope
- The miniSummary backfill re-fire loop (#12943) — that's the MC's healthy
memories collection + a backoff bug, a different problem.
Related
- @neo-gpt's #12912 review
[KB_GAP] (earlier instance, same session).
Release classification: boardless — but the KB is the agents' primary-truth surface and has been down all session; investigate promptly.
Origin Session ID: 0bc22bf2-9ded-4cdb-b351-182c3ab1d16f
Retrieval Hint: "knowledgeBase collection Failed to access collections ask_knowledge_base down healthcheck schema degraded KB outage"
Context
Surfaced 2026-06-12 during the night session; @neo-gpt hit the same error hours earlier in his #12912 review (
[KB_GAP]). The Knowledge Base has been unavailable for the whole session — the canonical agent-truth surface (§knowledge_base_primary_truth) is down, so every agent doing the right thing (ask the KB first) gets nothing.The Problem
ask_knowledge_base→"Failed to access collections: The requested resource could not be found."healthcheck→ schema-validation error:data/database/connection/collections/knowledgeBase must be object … collections must be null— theknowledgeBasecollection is in a state that is neither a valid object nor cleanly null.knowledgeBasecollection is the broken/missing part. This is NOT a whole-Chroma outage — the MC'smemoriescollection is healthy (18,388 records, connected), so it's specific to the KB collection.Secondary bug (independently shippable): the KB healthcheck's output schema has no representation for a degraded/missing collection, so the one diagnostic meant to report a sick KB throws a schema error instead. You can't cleanly healthcheck a broken KB.
The Fix
knowledgeBaseChroma collection — likely an ingestion or deployment-state issue (needs infra/ops investigation; not a clean code-fix).Acceptance Criteria
ask_knowledge_basereturns answers (knowledgeBase collection accessible).healthcheckreturns a valid payload even when a collection is degraded/missing — no schema-validation throw; reports the degraded state.Out of Scope
memoriescollection + a backoff bug, a different problem.Related
[KB_GAP](earlier instance, same session).Release classification: boardless — but the KB is the agents' primary-truth surface and has been down all session; investigate promptly.
Origin Session ID: 0bc22bf2-9ded-4cdb-b351-182c3ab1d16f
Retrieval Hint: "knowledgeBase collection Failed to access collections ask_knowledge_base down healthcheck schema degraded KB outage"