You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
First start runs two one-time steps: the ingested_at schema addition and backfill below, and an
FTS index rebuild, because lbug moved 0.21.0 → 0.21.1 and the #649 marker records the exact lbug
version. Both are logged; a large graph starts slower that once. Later starts do neither.
Existing graphs get an ingested_at backfill on first start (#673, ADR-0673): from the ts of
the WAL entry that created each record where the WAL still has one, otherwise from created_at.
The created_at fallback is approximate (for pre-existing records it was the event time).
Added
ingested_at on episodes, entities and edges (#673, ADR-0673): the service-clock time the
record was first written, kept separately from the caller-supplied event time (reference_time → valid_at / created_at). Never caller-supplied, set once, preserved across WAL rebuild, recovery,
checkpoint restore and dump compaction. Returned on every episode/entity/relationship read and
accepted in the fields projection of knowledge_get_episodes and knowledge_list_entities.
Additive only; created_at / valid_at and all ordering are unchanged.
Changed
lbug 0.21.0 → 0.21.1 (#676). Extensions unchanged (still the v0.21.0 path, same pinned bytes);
storage version stays 47.
MCP library rmcp 2.2 → 3.5 (#680). The MCP tool surface is unchanged.
Re-ingest graphs built by extraction on 0.16.3 or earlier to remove endpoints already
re-pointed onto the wrong entity and stored self-loops (#666); they are not repaired in place.
knowledge_get_episodes with no group_id/group_ids now returns episodes from every group,
not just liminis (#667, fixes #648). Pass group_ids: ["liminis"] (or group_id) to keep the
old scope. group_ids is now honoured (it was silently ignored); group_id remains an alias.
Added
Paging, projection and prefix filters on the two bulk reads (#667, ADR-0667): knowledge_get_episodes and knowledge_list_entities accept cursor (send "" to start; follow next_cursor until null), fields (return only the named keys; embeddings are never returned)
and name_prefix (case-insensitive for entities). Absent all three, responses are unchanged.
knowledge_status.wal carries scope, default_group and see: "wal_groups" when a WAL root is
configured, so exists: false for a default group with no stream no longer reads as an error.
Changed
knowledge_get_episodes orders ties on created_at by uuid (descending) so paging is
deterministic. Previously the order within a tie was arbitrary.
Fixed
An edge endpoint that exactly matches a stored entity is no longer re-pointed onto a similar
entity of the current chunk (#666, remainder of #645, ADR-0666). Salvage now runs only for
names with no exact stored match.
Edges whose two endpoints resolve to the same entity are dropped, not stored as self-loops
(#666). New additive edges_dropped_self_loop field on knowledge_process_chunk.
Extraction no longer merges distinct short all-caps codes such as ACDS / ACDM (#666). Behaviour change: the identifier veto now also fires on mutually exclusive 2–6 letter
upper-case tokens; acronym/expansion aliases still merge.
Identity-kind (identity: true) fuzzy matching is intentionally unchanged; see ADR-0666.
lcg-eval 0.16.4
Install lcg-eval 0.16.4
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/verveguy/liminis-context-graph/releases/download/v0.16.4/lcg-eval-installer.sh | sh
The first start rebuilds the three FTS indexes once (logged; time proportional to corpus size).
Nothing else to do. Pinned to 0.16.0–0.16.2? See the release notes for the manual workaround.
Re-ingest if your graph was built with ≤ 0.16.2 extraction. Entities wrongly merged by the old
dedup (e.g. ADR 2018 = ADR 2019) and their concatenated summaries are not repaired in place.
LCG_DEDUP_ADAPTER_URL is ignored (deprecated; it configured a protocol nothing implemented).
Fixed
Databases built by 0.15.x or earlier: non-ASCII FTS terms were undeletable and silently
unsearchable after the lbug 0.20 → 0.21 upgrade (#649, fixes #646, ADR-0649). Rebuilt once on
first start; a residual "inconsistent" error rebuilds all three and retries once.
Extraction no longer merges entities that differ only by a number or identifier (#650,
ADR-0650). Behaviour change:Python 3 / Python-style pairs no longer merge by embedding.
Merged entity summaries are consolidated, not concatenated, and summary_embedding is
re-embedded on every merge (#651, fixes #647, ADR-0651). Uses the configured extractor; capped
at 600 characters (bounded fallback without one); WAL replay never calls the LLM.
CI no longer recompiles the workspace in the test job (~9 min per run; #657).
Added
LCG_DEDUP_LLM now works (#652, ADR-0652): opt-in LLM-verified dedup using the configured
extractor, batched per chunk; any failure means "not a duplicate". knowledge_status reports dedup_mode; knowledge_process_chunk returns dedup_paths counts.
Changed
LCG_DEDUP_LLM values 0/false/off/no/empty now mean off (previously any value enabled it).
lcg-eval 0.16.3
Install lcg-eval 0.16.3
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/verveguy/liminis-context-graph/releases/download/v0.16.3/lcg-eval-installer.sh | sh
Assert-only ontology types: extract: false (ADR-0637, #637, implements #636). An entity or relation type marked extract: false is omitted from the
extraction prompts and never accepted from the extractor — a stray extracted label is
reclassified to Unclassified / UNCLASSIFIED (original label kept) in both strict and open
mode and never receives the type's identity kind — while staying fully declared for knowledge_assert_*, the identity set and its -32003 guard, ancestry and drift. Reprocess and
canonicalize never retype into one. Defaults to true; an ontology without the flag renders
identical prompts and hashes identically. See docs/ontology.md.
lcg-eval 0.16.2
Install lcg-eval 0.16.2
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/verveguy/liminis-context-graph/releases/download/v0.16.2/lcg-eval-installer.sh | sh
knowledge_reload_ontology (ADR-0627, #627,
implements #611). An admin-scope IPC method and MCP tool that reloads one group's ontology
without a service restart: it re-resolves the group through the normal precedence, honours the
identity-set guard (a reload that would reinterpret existing entities is reported in identity_refusal and enforced as -32003 on writes, and clears when the previous set is
restored and reloaded), and returns {group_id, previous_hash, new_hash, changed, drift, identity_refusal}. An unchanged file is a no-op. Other groups and stored data are untouched.
See docs/ontology.md.
Relevance evidence and a similarity floor on knowledge_find_entities / knowledge_find_relationships (ADR-0629, #629, implements #613). Each result carries a search object — rrf_score, text_match, bm25_score, and per-path cosine similarities (null when that path didn't retrieve it) — and
an optional min_similarity (clamped 0–1, no default) drops distant vector candidates before
fusion; full-text hits stay eligible. Unset, results and order are unchanged.
Fixed
health_check no longer blocks behind the write lock (ADR-0628, #628, reported in #612). While a write is pending or in progress — notably a long knowledge_rebuild_from_wal — it now answers immediately with state: "busy", ok: true, healthy: false and an activity (rebuilding with job_id and progress, or writing), instead of queueing until the write finished. Supervisors with short probe timeouts
no longer mistake a rebuilding daemon for a dead one and restart it mid-rebuild. The healthy
and degraded responses are unchanged. Out-of-repo callers that treat any ok: true as ready
should gate on healthy instead.
The docs site now republishes automatically on release (#626); since 0.14.0 it had only been
updated by manual dispatch.
An episode whose extraction straddles an ontology reload no longer records its stale ontology
hash in the group's drift sidecar or clears the reloaded group's drift.
lcg-eval 0.16.1
Install lcg-eval 0.16.1
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/verveguy/liminis-context-graph/releases/download/v0.16.1/lcg-eval-installer.sh | sh
First start migrates entities (issue #615).Entity gains a kind column and lookup_key
becomes group_id ␟ kind ␟ lower(name). A one-time backfill sets every existing entity to kind Entity and re-keys it (O(N), automatic). It also normalizes labels to put Entity first
(e.g. [Object, Entity] → [Entity, Object]); the label set is unchanged. Downgrade is
unsupported: a pre-#615 binary cannot replay a WAL written after this change (lookup_key is
no longer written to the WAL).
Added
Ontology-driven kind assignment (ADR-0616, #616, follows #615). An entity type marked identity: true in the ontology is assigned as the kind of entities extracted with that primary type, so Person "Aurora" and Technology "Aurora" stay distinct; unflagged types (and every existing ontology) merge by name exactly as
before. Per group, the identity-bearing set is recorded in .lcg/identity-set/<group>.json, and
a change that would reinterpret existing entities (flag added or removed for a type the group
already holds) is refused for that group's knowledge_add_episode with JSON-RPC -32003 —
restore the set or re-ingest from source. knowledge_reprocess_entity_types never changes kind
and reports kind_disagreements; knowledge_status gains group_identity_refusals.
Edge endpoints whose name is ambiguous across kinds are dropped rather than guessed.
Kind-scoped entity identity (ADR-0615, #615,
implements #614). Same-named entities of different kinds now coexist instead of the second knowledge_assert_entity overwriting the first. knowledge_assert_entity takes kind
(omitted ⇒ default kind Entity only — writes are scoped); knowledge_assert_relationship takes source_kind/target_kind; new read tool knowledge_resolve_entity; kind filter on knowledge_find_entities/knowledge_list_entities; kind on knowledge_merge_entities; source_kind/target_kind (stored as pointer endpoint_kind) on knowledge_add_cross_group_edge. A name-only resolution matching several kinds fails with
JSON-RPC -32002 and a candidate list. Merges never cross kinds. Every entity node now carries kind (additive).
Changed
lbug 0.20.3 → 0.21.0, with version-matched v0.21.0 extensions (#619). Upstream's 2026-09-29
republish of the v0.20.0 extensions left 0.20.3 with no safe Windows fts pairing (see ADR-0593
addendum). 0.21.0 extensions are published per-version and pinned in LBUG_EXTENSION_HASHES.
Storage version stays 47: no database migration.
Release builds verify the staged lbug extension bytes against pinned sha256 (#593); Windows CI
now creates and queries vector/FTS indexes rather than only loading the extensions (#587).
lookup_key is a derived value: stripped from WAL records on write, recomputed on replay; a
literal lookup_key in an older WAL is ignored. Old WALs replay as kind Entity.
lcg-eval 0.16.0
Install lcg-eval 0.16.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/verveguy/liminis-context-graph/releases/download/v0.16.0/lcg-eval-installer.sh | sh
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Release Notes
lbug 0.18.1 → 0.20.3. One-way storage migration (42 → 47) — read Upgrading. No API changes.
Full detail: docs/releases/0.15.0.md.
Upgrading
One-way migration. A database at storage version 42 — written by 0.18.1, which is
everything 0.14.x shipped — opens directly under the new binary, with no export and the WAL
untouched. The first checkpoint rewrites it to storage 47, after which an older binary will
not open it. To roll back: stop the service, move .lcg/db/ aside, start the old binary; it
rebuilds from the WAL.
No API changes and no other manual step.
Changed
lbug 0.18.1 → 0.20.3. Picks up the fixes the 0.18.1 rollback deferred — ladybug#845 (FTS
heap corruption under concurrent scan/write, a path this service exercises live), #837, #864, #894, #884 — and ladybug#883, a SIGSEGV in the cached-prepared-statement path,
fixed outright in 0.20.3 rather than merely mitigated. Also forward-fixes a stale-row bug on the
cached-plan fast path (ladybug#877/#878) that 0.18.1 does not appear to exhibit.
0.20.3 rather than 0.20.4, because lbug 0.20.4 crashes on Windows in the published vector/fts extensions. Upstream publishes those per minor, and the v0.20.0/win_amd64
files are 0.20.2-era builds — ABI-compatible with a 0.20.3 core, not a 0.20.4 one. Reported
upstream as LadybugDB/ladybug#971. Linux
and macOS are unaffected by the choice.
lcg-eval 0.15.0
Install lcg-eval 0.15.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/verveguy/liminis-context-graph/releases/download/v0.15.0/lcg-eval-installer.sh | sh
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Release Notes
Windows is now a supported platform, and hybrid search returns results in ranked order rather
than insertion order. No migration; lbug stays pinned at 0.18.1.
Added
Windows (x64) support.lcg-service builds and runs on Windows, using a named-pipe IPC
transport in place of the Unix domain socket. Every pipe instance carries a protected DACL
granting only the service's own user and SYSTEM (the default pipe DACL grants Everyone read),
refuses remote clients, and claims its name with FILE_FLAG_FIRST_PIPE_INSTANCE. The service
writes the pipe name to .lcg/service.endpoint so clients in other languages can discover it,
and Windows console-control events drive clean shutdown. (#581, #582)
A published Windows artifact, x86_64-pc-windows-msvc, installable via a new PowerShell
installer or the .zip. OpenSSL is linked statically into the executable — a deliberate
Windows-only exception to ADR-0550's dynamic posture, since an archive cannot carry DLLs for one
target and no Windows package manager would patch a DLL we shipped. The two OpenSSL DLLs that
lbug's own extensions import are bundled beside them and resolved at load time. Consequence: an
OpenSSL CVE fix requires a new lcg release on Windows, where macOS and Linux pick it up from the
system. See ADR-0581. (#585, #586)
Fixed
find_entities and find_relationships returned results in insertion order, not ranked
order. Both fetch rows with a WHERE uuid IN lookup carrying no ORDER BY, so the RRF-fused
ranking computed immediately beforehand was discarded — every hybrid search returned the correct
top-k set in an arbitrary order. Rows now follow the fused ranking. (#584)
Bundled lbug extensions were not found when lcg was launched through a symlink. On macOS and
Windows current_exe() reports the path the process was launched as, so a stable symlink into a
versioned install resolved the bundle against the link's directory, missed the .lbdb tree beside
the real binary, and silently fell back to downloading from the CDN — voiding 0.14.2's offline
guarantee with no error. The executable path is now canonicalized. (#583)
Windows notes
Requires the Microsoft Visual C++ Redistributable (x64, 2015–2022). It is not bundled and is
present on most machines; without it the executable fails to start with a loader error.
Verified on Windows 11 x64. Windows 10 and Windows ARM are untested; no Windows ARM build is
published.
Downstream clients must read .lcg/service.endpoint to locate the pipe. orac does; the
Liminis app and liminis-framework do not yet.
lbug stays at 0.18.1. The published Windows vector and fts extensions for the entire
lbug 0.20.x line crash the process — a regression from 0.19.0, reproducible with LadybugDB's own
CLI and unrelated to lcg. The lbug upgrade is deferred until that is fixed upstream, so all four
platforms can move together. See docs/releases/0.14.3.md.
lcg-eval 0.14.3
Install lcg-eval 0.14.3
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/verveguy/liminis-context-graph/releases/download/v0.14.3/lcg-eval-installer.sh | sh
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Release Notes
A maintenance release: the MCP front-end no longer hard-fails when the daemon is down, and a new
admin operation reclaims the dead weight a pre-0.14 WAL is still carrying. No migration, no API
changes.
Added
knowledge_strip_wal_embeddings {group_id, dry_run} reclaims the space a WAL written before
0.14.0 still carries. 0.14.0 stopped writing embedding vectors to the WAL and made replay
ignore any it finds, but left existing WALs untouched — on the reference corpus those vectors
were 89.9% of WAL bytes, parsed and discarded on every replay, rebuild and hydration. This
rewrites each qualifying .jsonl in place to drop params entries keyed by an embedding column,
leaving every other field, record ordering and sequence number alone. Because replay never reads
a stored vector, it cannot change what a rebuild produces. Idempotent (a re-run is a zero-I/O
no-op), crash-safe per file (tmp-then-atomic-rename), and dry_run reports what would change
without touching anything. admin scope. See ADR-0577 (#577).
--connect-eager restores the pre-0.14.2 startup behaviour for --mcp-stdio --connect —
dial at startup and exit if the socket is unreachable — for operators relying on that as a
health-check signal (#575).
Fixed
--mcp-stdio --connect no longer hard-fails when the daemon socket is unreachable at launch. initialize and tools/list now complete immediately from the compiled-in tool registry without
dialling; the first tools/call dials lazily, and an unreachable daemon returns a tool result
with isError: true naming the socket rather than exiting. A later call succeeds once the daemon
comes up, and a daemon restart that rebinds the socket is transparently re-dialled — both without
any client-side reconnect. Previously the client cached the failed attach for roughly 15 minutes,
leaving the reader showing disconnected long after the cause was fixed, and a restart orphaned
live connections until a manual /mcp. Reported from downstream MCP use (#574). --scope
enforcement is unchanged. See ADR-0575 (#575).
Internal
Corrected a false claim in the 0.14.0 release notes that lbug 0.18.1 carries ladybug#845, #837 and #864. Those landed in 0.20.x, after 0.18.1 shipped; the document had said both
things in different sections. ladybug#845 is FTS heap corruption on a path this project
queries concurrently, so the error understated a known risk. Moving to 0.20.2 is tracked in #561 (#569).
lcg-eval 0.14.2
Install lcg-eval 0.14.2
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/verveguy/liminis-context-graph/releases/download/v0.14.2/lcg-eval-installer.sh | sh
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Release Notes
A maintenance release: startup no longer depends on a third-party CDN, and two unbounded-wait
defects on the embedder path are fixed. No migration, no API changes.
Added
Bundled lbug vector/fts extensions: the release archive now ships the extension binaries for
its target platform, so Db::open never needs to download them from extension.ladybugdb.com at startup — a fresh install on an air-gapped or
egress-restricted host now works with zero prior cache and zero network access. Falls back
to today's download-on-demand behavior when no bundle is found (e.g. running from source). LCG_LBUG_HOME lets an operator redirect resolution to a directory of their own. See the
README's "Offline / air-gapped startup" section and ADR-0559 (#559).
Fixed
The UDS embedder transport had no timeout. LCG_EMBEDDING_TIMEOUT_MS (whole-request) and LCG_EMBEDDING_CONNECT_TIMEOUT_MS (acquiring a pooled connection) now bound it, with the same
defaults and validation the HTTP transport already used. This is the unbounded hang #510 fixed
for HTTP, still present on the default transport: a hung or unresponsive sidecar could
stall an embed call indefinitely. See ADR-0541 (#541).
state.write_lock is no longer held across the embedder round trip. knowledge_assert_entity
and knowledge_assert_relationship now resolve-and-update under the lock, release it for the
embed, then re-acquire and re-resolve before inserting — falling back to the update path if a
concurrent writer won the race. Unrelated writes are no longer serialized behind a slow
embedder. See ADR-0543 (#543).
Internal
Closed a TOCTOU port race in embedder_degraded_mcp.rs's test helpers that had caused a flaky AddrInUse failure, and made a lost race in the delayed-bind stub fail loudly rather than
masquerading as an embedder-unreachable product failure (#560).
The release runbook now requires the two sidecar-gated tests to be run and their figures
recorded before a release — they skip silently when no sidecar is present, so a green board
never proved they ran (#531, #566).
lcg-eval 0.14.1
Install lcg-eval 0.14.1
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/verveguy/liminis-context-graph/releases/download/v0.14.1/lcg-eval-installer.sh | sh
