reliquary_tc_acl_denied_total | reason, scope_class, vault | ACL/scope denials by vault, scope class, and reason. |
reliquary_tc_acl_walk_pruned_total | vault | Paths the graph-walk ACL filter excluded from a walk that an unfiltered walk over the same seed frontier would have reached, by vault. Zero for an unrestricted caller by construction (nothing to prune); a non-zero, rising count for a restricted one is the filter’s recall cost made visible, not an error. |
reliquary_tc_activation_recompute_chunks_total | vault | Chunks whose cached_activation_score was recomputed by the periodic ACT-R activation job, by job name. Cumulative. A flat line while the scheduler reports the job running means chunk_retrievals has stopped growing, not that the job stalled. |
reliquary_tc_audit_write_failed_total | tool, vault | Security-audit event writes that failed, by vault and tool. Audit is fail-open by design (a failed write must never break dispatch), so this counter is the only signal that the audit trail has gone lossy. |
reliquary_tc_auth_family_revocations_total | reason | OAuth grant-family revocations by the bundled authorization server, by reason. authorization_code_reuse, refresh_token_reuse and refresh_token_foreign mean a presented credential was a replay (or minted under a replaced server secret) and its whole family was revoked on sight (RFC 9700 section 4.14): any growth is worth investigating, since one leaked credential is a one-request kill switch on the client’s session. revoked_by_client (RFC 7009) and authorization_code_failed are the benign counterparts. |
reliquary_tc_auth_rejections_total | reason | Refused tokens at the HTTP edge, by rejection reason. |
reliquary_tc_embed_batch_rejections_total | vault | Embed requests the provider rejected for exceeding its context (HTTP 400/413), then bisected + retried, by vault. Each is an extra round-trip; a persistent count means embeddings.maxBatchTokens is too high. |
reliquary_tc_embedder_host_recycles_total | reason | Embedder child processes the burst-end recycle policy drained, by reason (reconcile = a finished index pass that embedded chunks, budget = parent + child RSS over embeddings.memoryBudgetMb at a write-queue drain while the parent alone is under it). Cumulative. A crash-restart is not counted here. |
reliquary_tc_governor_truncations_total | tool, vault | Response-byte governor truncations/refusals, by vault and tool. |
reliquary_tc_hitl_elicited_total | tool, vault | HITL elicit confirmations required, by vault and tool. |
reliquary_tc_idempotency_cache_skipped_total | tool, vault | Idempotency results skipped over the byte cap, by vault and tool. |
reliquary_tc_idempotency_hits_total | tool, vault | Idempotency cache hits, by vault and tool. |
reliquary_tc_idempotency_release_failed_total | gate, tool, vault | Idempotency claims left ORPHANED because the final release attempt on a pre-handler failure failed, by vault, tool and the rejection gate that led there (throttle | hitl | other). Release is best-effort so it never masks the error the caller must see, which also makes this counter its only signal. Counted at the LAST attempt, not at the gate: the gates get a retry in the outer catch, so a gate-site failure that retry then cleans up leaves no orphan and is deliberately not counted. When this does fire the claim survives, and a matching retry inside the reclaim window (idempotencyReclaimSeconds, default 60s) returns idempotency_in_flight; at or after that window claimOrReplay reclaims the row. The 24h idempotencyTtlSeconds is the later alternate expiry, not the blocking bound. |
reliquary_tc_index_frontmatter_failures_total | vault | Notes skipped in a pass because their frontmatter failed to parse as YAML, by vault. Not swept as stale (the note was still walked) and retried automatically on the next index_vault pass once the note’s YAML is fixed — see doctor’s index.coverage check for which paths. |
reliquary_tc_index_stale_skipped_total | vault | Notes an indexVault batch skipped because a concurrent write_note/watcher commit changed the path’s chunks after the plan was computed, by vault. Not a failure — the skipped note is re-planned against current content on the next index_vault pass. |
reliquary_tc_index_write_failures_total | vault | Notes skipped in a pass because the embed provider rejected them even at single-text size, by vault. Unlike the batch rejections above these are NOT retried within the pass, so the note is absent from the index until the next reconcile. |
reliquary_tc_ingest_dedup_skipped_total | vault | Chunks whose embedding was reused from an identical-body sibling instead of recomputed, by vault. This is work AVOIDED, so a rise is good ONLY for the chunks it actually resolved — see reliquary_tc_ingest_dedup_unresolved_total for the subset that copied nothing; a fall in this counter means the dedup path stopped matching. |
reliquary_tc_ingest_dedup_unresolved_total | vault | Chunks skipped for embedding by cross-path dedup whose source had no stored vector to copy, by vault. A rise is BAD — these chunks are FTS-only (no dense/sparse/colbert) until the owner note re-embeds successfully; it is the loss side of reliquary_tc_ingest_dedup_skipped_total, not work avoided. |
reliquary_tc_ingest_secrets_skipped_total | vault | Chunks the secret gate refused to index, by vault. Non-zero is expected on a vault containing credentials; a SUDDEN rise means content that used to index no longer does. |
reliquary_tc_memory_defense_hits_total | pattern | memoryDefense (GH #994) redact/block matches, by pattern id only — no vault, tool, or content — so an operator sees the policy firing without reading what was matched. |
reliquary_tc_memory_orphans_swept_total | class | Memory rows deleted by the periodic orphan sweep, by class (dangling_relations, dangling_intervals, retired_entities, removed_vault_entities, removed_vault_relations, removed_vault_intervals). Cumulative. Only the two dangling_* classes are on by default; the other four move only when the operator sets their retention window. A dry run does not increment it. |
reliquary_tc_morgiana_emit_dropped_total | reason, vault | MORGIANA events dropped, by vault and reason. |
reliquary_tc_morgiana_spool_files_pruned_total | reason | Morgiana event spool day files deleted by the retention sweep, by reason (age = older than observability.retention.spoolRetentionDays, size = oldest files dropped to fit spoolMaxBytes). Cumulative. The current day’s file is never deleted. |
reliquary_tc_output_schema_drift_total | tool, vault | Handler payloads that did not match their advertised outputSchema, by vault and tool. In production this is WARN-only — the payload still ships — so a non-zero value is the only signal that a tool’s declared contract has drifted from what it returns. In dev/CI the same condition is a hard internal_error. Any non-zero count names a tool whose schema or handler is wrong; there is no benign case. |
reliquary_tc_provenance_faults_total | kind, tool, vault | Write-provenance recording faults, by vault, tool and kind. omitted: a committed write left NO record (recording is fail-open, so the write still succeeded). head_untrusted: the record was written but the chain head failed validation and was not re-signed. Any non-zero value means the provenance trail is incomplete or was tampered with; doctor reports the same events from event_log. Labels are vault id, tool name and the two-value kind, never a path or an error message. |
reliquary_tc_rate_limit_backend_outages_total | backend | Shared rate-limit backend outages (sqlite or redis unreachable), by backend. Counted once per outage, not per request; the failure policy decides what calls do meanwhile. |
reliquary_tc_rate_limit_hits_total | scope_class, vault | Rate-limit refusals, by vault and scope class. |
reliquary_tc_rerank_outcome_total | outcome, vault | rerankWithScores decisions, by vault and outcome. executed is the only outcome where the reported ranking actually came from the reranker; every other value is the synthetic-descending-score fallback for a different reason — not_configured (no reranker injected), skipped_by_policy (gatedRerank’s hardness gate did not fire), timed_out, malformed_response (the call returned but produced no usable hit), provider_error (the call rejected for any other reason), fallback_used (no more specific reason — e.g. an empty candidate set). |
reliquary_tc_retrieval_content_bytes_in_total | stage, vault | Content bytes materialized entering a stage boundary, by vault and stage. Populated only at candidateAssembly (pre-dedup, across streams) and diversity/gatedRerank (pre-top-K-cut). Compare to the out counter for the same stage: the gap is hydrated content that was never used. |
reliquary_tc_retrieval_content_bytes_out_total | stage, vault | Content bytes surviving a stage boundary, by vault and stage. See the in counter. |
reliquary_tc_retrieval_stage_candidates_in_total | stage, vault | Candidates entering each graph-search stage, by vault and stage. Divide the out counter by this for the stage’s pass-through ratio: a stage sitting at 1.0 is no longer filtering anything while still costing its latency. |
reliquary_tc_retrieval_stage_candidates_out_total | stage, vault | Candidates leaving each graph-search stage, by vault and stage. See the in counter. |
reliquary_tc_sql_busy_total | reason, txn, vault | Write transactions that failed on a busy database, by vault, transaction, and reason. reason=busy means contention outlived busy_timeout (5s) — the writers genuinely overlap that long. reason=snapshot is a BUG REPORT, not tuning: it can only be produced by a deferred BEGIN that read and then tried to write, a failure busy_timeout cannot retry, so any non-zero count names a write path still using BEGIN where it should use inWriteTransaction. |
reliquary_tc_tool_calls_total | status, tool, vault | Tool calls by vault, tool, and terminal status. |
reliquary_tc_vec_fallback_total | reason, vault | Searches that abandoned the vec0 KNN index for the exhaustive brute-force scan, by vault and reason. Results stay correct; the cost profile does not. reason=error is usually a dimension mismatch after an embedding-model change (the index no longer matches the query) and a persistent count means a real misconfiguration; reason=underfill means ACL-invisible chunks may be crowding out visible ones, so the over-fetch could not fill k visible hits. |
reliquary_tc_vec_rebuild_total | reason | vec_chunks DROP+rebuild events, by reason. legacy_shape is a one-time pre-partition upgrade; fingerprint_changed means the embedding provider/model/dimensions, distance metric, or chunk/enrichment representation changed since the index was built; table_missing means vec_chunks was absent over active embeddings and was rebuilt from them. Every vault’s dense index is cold until it re-embeds — any non-zero count outside a deliberate model migration is worth investigating. |