Skip to content

Full changelog

Every CE release ever cut, newest first. Use this if you're on an older version and seeing an issue — scan the entries between your version and the current latest to see whether your bug is already fixed.

Format per entry: short headline → bullet list of user-visible fixes / features / known issues → link to the GitHub release page (binaries + SHA-256s).

For richer per-release pages (architecture deep-dives, validation summary, install tabs), see the Releases overview. New per-release pages are added going forward; pre-v1.8.5 lives entirely in this changelog.

v1.8 series

v1.8.5-ce — 2026-06-15

Agent-memory primitives. Full notes: v1.8.5-ce page.

  • New SQL: MEMORY_STORE(ns, content), MEMORY_RECALL(ns, query, top_k) (table-valued), MEMORY_FORGET(ns, id). Per-tenant scoped; per-namespace auto-create of _memory_<ns> backing table; embeds via the configured embedding model (default all-minilm:latest, dim 384).
  • Replaces ~150 LOC of hand-rolled CREATE TABLE + EMBED + HNSW + COSINE_SIMILARITY plumbing in every memory-having plugin.
  • Ships with @synapcores/sdk@0.5.0 (Node) + synapcores 0.5.0 (Python) + @synapcores/openclaw-memory@0.4.0 (OpenClaw plugin migrated to client.memory.*).
  • Live-validated on the i5-10400F canary (non-AVX-512, no GPU). 17/19 Python SDK live integration tests pass.

v1.8.4-ce — 2026-06-15

Mac chat_with_tools SIGSEGV fix. Carries all v1.8.3 security fixes.

  • Migrated grammar_lazy to the non-deprecated patterns API (llama_sampler_init_grammar_lazy_patterns) — deep-copies trigger-word strings on the C side, eliminating the CString-lifetime use-after-free that crashed AGENT_RUN on macOS.
  • Symptom on prior versions: Mac .ips crash with faulting address fragment of </tool_call>. v1.8.2's into_raw() leak workaround was not sufficient; this is the proper fix.
  • Public Rust API unchanged.

v1.8.3-ce — 2026-06-14

4 security P0s + WS agentic-mode default flip. Upgrade if you exposed the gateway to anything beyond 127.0.0.1.

  • #339 (P0): Hardcoded JWT fallback secret removed from the binary. Per-install random secret persisted at <data_dir>/.jwt-secret mode 0600. Tokens signed with the old fallback string now return HTTP 401.
  • #341 (P0): /admin/* routes now require super_admin role and cross-check URL tenant vs. JWT claim.
  • #342 (P0): TenantContext cross-checks JWT company_id claim against the stored user record. Forged claims no longer pick a tenant.
  • #343 (P0): tenant_id UUID-only canonicalization rejects .. path-traversal inputs.
  • WS parity: AIDB_CHAT_AGENTIC_MODE default flipped to false (was true on WS, false on REST — inconsistent).

v1.8.2-ce — 2026-06-12

Mac chat segfault hotfix + small-model AGENT_RUN viability.

  • First attempt at the Mac chat_with_tools segfault — CString into_raw() leak workaround in grammar_lazy. Superseded by v1.8.4's proper patterns API migration; if you're on v1.8.2 and still seeing Mac crashes, jump straight to v1.8.4 or later.
  • Tool-description compression in the agent prompt: smaller GGUFs (phi-3, qwen2.5:3b) now fit AGENT_RUN reliably.
  • Documented the persona-classifier env var for operators who want deterministic persona routing.

v1.8.1-ce — 2026-06-11

Critical fixes for the v1.8.0 cold-path.

  • CRITICAL panic: AI chat route used Lazy::force().unwrap() on a missing default model. Every AI route 500'd until a model was pulled. Fixed: graceful path that prompts for a model pull instead.
  • HIGH: UI menu still showed v1.5. Bumped version display.
  • HIGH: Login screen allowed credential creation in single-tenant CE mode (should be sign-in only).
  • MED: Image upload didn't generate vision text out-of-box — provider default improved + clearer error UX.
  • registry vs install path: unified model-pull source so install + registry CLI both go through the same path.

v1.8.0-ce — 2026-06-08

Native inference baked into the engine — no Ollama sidecar. Full announcement: blog post.

  • synapcores pull <model:tag> — OCI Distribution v2 protocol against registry.ollama.ai. Multi-layer manifests, sha256-verified blobs, resume on interrupt, idempotent re-pulls. Content-addressed model store at ${data_dir}/models/blobs/<sha256>.
  • New SQL: PULL_MODEL('name:tag'), LIST_MODELS() table-valued, DELETE_MODEL('name:tag'). All listed in the sql_manual MCP tool.
  • Beats apples-to-apples CPU Ollama: 20-recipe agent cert in 39:43 (20/20 PASS) vs. Ollama CPU 65:08 (0/20 PASS, timed out at internal 2-min limit). i5-10400F, no AVX-512, no GPU offload.
  • Model-authoritative tool calling: GGUF chat templates rendered in Rust via minijinja. Every GGUF with a tool-aware chat template just works (Qwen Hermes, Llama 3 <|python_tag|>, Mistral [TOOL_CALLS], generic JSON).
  • Persistent LlamaContext sessions per chat persona, KV cache reuse via prefix matching, LRU eviction at a configurable resident-memory cap. Measured prefix-reuse rate in the cert: 91%.
  • Streaming chunk channel via mpsc. Per-call tracing visible at RUST_LOG=info.
  • PK-equality cross-collection bleed fix (#315) rolled forward from v1.7.0.2.1-ce: WHERE pk = N no longer returns spurious [None] rows when N is ±1 from an existing integer PK.

v1.7 series

v1.7.0.2.1-ce — 2026-06-01

OLTP perf hotfix on top of v1.7.0.2. Upgrade if you saw a 24-31% point-select TPS drop after v1.7.0.2.

  • Removed a redundant tokio task-local scope wrap around the SQL execution future on /v1/query/execute. The path already received database_context as an explicit parameter; the task-local was overhead with no behavior.
  • #225 per-request database context (multi-DB in one CE session) preserved.
  • OLTP at or above the v1.7.0.1 baseline at every thread count.

v1.7.0.2-ce — 2026-06-01

Bug-fix bundle on top of v1.7.0.1. Recipe cert 182/193 PASS.

  • #225: Per-request database context via tokio task-local — multiple concurrent CE sessions can hit different databases without racing the process-global current-db. (Tuned again in v1.7.0.2.1 — see above.)
  • #181: OOB EMBED() restored — works without [query.ai_service] configured (regression introduced in v1.6.6.7).
  • #205: Chat-agent split provider/embedding URLs (in-DB agentic OOB without Ollama bridge gymnastics).
  • #180: storage_bridge entry logs + RocksDB reclaim on DROP TABLE.
  • #239/#240/#241: Dockerfile fixed for Debian bookworm + Rust 1.88 + cmake.
  • Bundle of ~15 SQL engine bug fixes (#206–#218, #222–#234). 10 new anomaly-detection recipes shipped.
  • Known deferred to v1.7.0.3: #224 alias-qualified columns, #232 AutoML inline-PREDICT planner rewrite, #220 RAG Document-not-found fallback (folded into later tags).

v1.7.0.1-ce — 2026-05-29

Provider + storage + HTTP row-cap fix bundle.

  • #174 (P0): OpenAI API key was being logged in plaintext in INFO logs. Redacted.
  • #175: SQL EMBED() was routing to the completion model when provider=openai split-mode was configured. Now uses the embedding model.
  • #176: Gateway hardcoded a 900s timeout on /query, ignoring [server].request_timeout. Now honors config.
  • #177: AGENT_RUN with OpenAI provider returned raw JSON tool-call as text instead of executing the tool. Fixed.
  • #178: Silent 1000-row truncation on PK BETWEEN when AIDB_EMIT_INDEXRANGESCAN was OFF. Default flipped to ON.
  • #169: RocksDB disk + memory reclaim on DROP TABLE — was leaking ~30 MB on-disk / ~55 MB resident per dropped 100K-row table.

v1.7.0-ce — 2026-05-28

Agentic SQL with AGENT_RUN (major). Headline feature. Supersedes the planned v1.6.6.10-ce — same code, larger version number to make the upgrade signal unmistakable.

  • SELECT AGENT_RUN(persona, task) — first-class SQL function running a complete ReAct loop (reason → tool → observe → repeat) and returning TEXT.
  • Agent has access to the same per-tenant DB as the calling session: execute_query, list_tables, describe_table, rag_search — all in one transaction.
  • Security hardening: 12 generic system-level tools removed from the agent's registry (shell, file, code, http, db, email, slack, calendar, browser, web_search, web_scraper, computer_vision). Tool surface 17 → 5. They cannot be re-enabled via agents.toml.
  • Default agent model: phi3 → qwen2.5-coder:7b (in-code default).
  • Result-cache contract changed: per-write invalidation now guaranteed. Old v1.6.6.x workarounds (AIDB_ENABLE_RESULT_CACHE=false, AIDB_RESULT_CACHE_TTL_SECONDS=1) are no longer needed — remove them when upgrading.

v1.6.6 series

v1.6.6.9-ce — 2026-05-27

Same code as v1.7.0 — v1.7.0 was retagged from v1.6.6.9 with a major bump to signal the AGENT_RUN feature add. If you're on v1.6.6.9, upgrade to v1.7.0+ for any later fix.

  • #165 canary catch: [query.ai_service].base_url was not threaded through GatewayCompletionProvider — silent fallback to localhost:11434 when operators configured Ollama via gateway.toml. Fixed.
  • 10 new AGENT_RUN-native recipes shipped to the homepage /recipes/agents repository (insurance claims triage, FinServ compliance, healthcare clinical Q&A, retail returns triage, SaaS support tier-1, sales briefing, devops incident triage, HR employee Q&A, legal contract review, logistics shipment exception).

v1.6.6.8-ce — 2026-05-26

6-branch bundle. Docker model-cache EACCES + warm-on-boot + path-aware AI timeout + AutoML PREDICT clarity + ml-worker module fix + Ollama error msg.

  • Recipe cert: 98/148 (+2 over v1.6.6.7).
  • macOS included.
  • ml-worker sqlx offline cache + RocksDB Send are pre-existing bugs deferred to v1.6.6.9.

v1.6.6.7-ce — 2026-05-25

Docker model-cache EACCES + recipe-import endpoint + frontmatter tolerance.

  • Fixed EACCES on the Docker model cache (write path was inside an immutable layer).
  • POST /v1/recipes/import endpoint added (was referenced by frontend + homepage but not registered — silent no-op).
  • Recipe-cert harness added to the release path (DEEP_SCAN).
  • macOS held off the tag-push (deferred-mac rule).

v1.6.6.6-ce — 2026-05-22

$N positional parameter binding on /v1/query/execute — server-side, injection-safe. Was passing raw strings through; now parameterizes via the parser.

v1.6.6.5-ce — 2026-05-22

Path-aware request timeout. The global 30s TimeoutLayer was 408'ing legitimate large imports — a 13K-row LeadDelta CSV exceeded 30s and aborted mid-flight. /import and /automl paths now get a 900s budget; everything else keeps the configured default.

v1.6.6.4-ce — 2026-05-22

AutoML sled model-store lock handling. Splits the error cases:

  • ENOLCK ("filesystem can't lock" — e.g. Docker Desktop / WSL2 volume) → clear error, no longer quarantines-as-corrupt + recreates (which looped and destroyed models on restart).
  • Transient EAGAIN → clear stale lock files + brief retry.

v1.6.6.3-ce — 2026-05-21

AVX2 baseline for x86_64 ggml — no AVX-512. Hard guarantee.

  • v1.6.6.2's GGML_NATIVE=OFF fix was correct but landed in the wrong CMake block. v1.6.6.3 explicitly pins GGML_AVX512=OFF (+VBMI/VNNI/BF16), GGML_AVX2/FMA/F16C=ON, -march=x86-64-v3 for the desktop x86_64-Linux block.
  • Verified on i5-10400F (AVX2 only): clean rebuild has zero AVX-512 instructions per objdump, 2895 AVX2; model eager-loads with no SIGILL.
  • This is now the permanent CE non-AVX-512 baseline. Every release after this is gated on an i5-10400F canary boot.

v1.6.6.2-ce — 2026-05-21

OOB Docker / install fixes + portable x86_64 ggml (SIGILL fix). First fix for the AVX-512 SIGILL on non-AVX-512 CPUs. Was incomplete — see v1.6.6.3.

v1.6.6.1-ce — 2026-05-21

AutoML predict + self-updating models + OpenClaw first-boot key.

  • AutoML predict now discriminates by feature values. Was returning a constant 0.85 in v1.6.6. Cause: numeric temp-table typing + Float64 scaler cast + input-order preservation + model resolution. State-asserting AutoML section added to feature_validator.py so this class of bug can't ship again.
  • AUTOML.TRAIN: async self-updating models (SQL + AFTER INSERT trigger) with single-flight coalescing and retrain-cooldown debounce to prevent storms.
  • First-boot mints a FullAccess OpenClaw memory API key + macOS launchd service docs.

v1.6.6-ce — 2026-05-20

The SQL/Cypher parity push. Three planned releases collapsed into one. feature_validator 79/0/0 on a bundled-UI build.

  • Unified Cypher: runs on /v1/query/execute AND the MCP execute tool (was /v1/graph/match only). SQL UNION no longer misrouted into the Cypher planner.
  • SQL parity: INFORMATION_SCHEMA.{tables, columns, key_column_usage}. WITH RECURSIVE (recursive CTEs with declared column lists). CREATE VIEW + CREATE MATERIALIZED VIEW + DROP VIEW. (Already in the line: INSERT … RETURNING, FULL OUTER JOIN, window frames, non-recursive CTEs, DISTINCT ON.)
  • Triggers — full feature set: BEFORE/AFTER × INSERT/UPDATE/DELETE, FOR EACH ROW, WHEN conditions, multi-event (OR), recursion depth cap, persistence across restart, in-place NEW.col mutation in BEFORE triggers.
  • Stored procedures — true procedural language: IF/ELSIF/CASE/WHILE/LOOP/REPEAT with LEAVE/ITERATE, DECLARE + SET variables, OUT/INOUT params, RETURN, RAISE + exception HANDLER.
  • Companion: SDK Bug D (graph.edges.delete) fixed in @synapcores/nodejs-sdk@0.4.1.

v1.6.5 series

v1.6.5.6-ce — 2026-05-19

24x7 stability hotfixes.

  • SQL-executor panics caught at the entry point: any panic during query execution is logged server-side (with payload + SQL excerpt + incident id) and returned to the client as generic 500. Tokio runtime worker survives.
  • substr() / SUBSTRING() no longer byte-indexed — uses character iteration. Em-dashes, smart quotes, emoji, CJK no longer trigger a byte index N is not a char boundary panic.
  • Global panic hook at gateway boot (defense in depth).
  • EMBED() correctness fix: [query.ai_service].embedding_model in gateway.toml is now actually used by EMBED(). Prior behavior silently fell back to the chat model's hidden state (e.g. 3584-dim Qwen2.5-Coder) — produced vectors that did not separate natural-language meaning. If you ran EMBED() on v1.6.5.5 or earlier with split-mode Ollama, your embeddings are wrong; re-embed after upgrading.
  • 404 on unmatched /v1/* routes no longer blames the CE — says Route not found: <path> and points at /v1/api-docs + /v1/openapi.json.

v1.6.5.4-ce — 2026-05-19

UI cleanup + carries v1.6.5.3 hot-fixes. Removed the per-key "View Statistics" icon from API Key Manager that 404'd on a never-implemented endpoint and surfaced a misleading "Enterprise-only" error.

v1.6.5.3-ce — 2026-05-19

Install-day blockers + PDF preview.

  • RocksDB per-tenant lock race fixed (the install-day blocker that surfaced in fresh-Mac smoke testing of v1.6.5.2).
  • install-ce.sh drops the MCP bridge at ~/.synapcores/.
  • New GET /v1/mcp/bridge endpoint serves the bridge offline.
  • PDF preview in image gallery: <iframe blob:> → <object blob:> bypasses the Chrome PDF-plugin blob-URL restriction.

v1.6.5.2-ce — 2026-05-18

OLTP overhaul + gateway wire-fixes. Massive perf jumps.

  • Insert throughput 42 TPS → 18,300 TPS (430×) at 16 threads. Root cause: per-insert reqwest::Client::new() was consuming ~96% of CPU on libcrypto X.509 root-CA bundle parsing. Fix: license-aware HTTP client pool (CE singleton, EE per-tenant).
  • Read-only range queries 1.9 TPS → 443 TPS (230×) at 16 threads. Root cause: IndexRangeScan optimizer rule wasn't recognizing Sort/Aggregate/Filter-wrapped plans. Fix: recursive plan walker.
  • Gateway wire-format: explicit 400 with remediation hint replaces silent-field-drop bugs for MatchRequest.params, named graphs, AutoML config.inline_rows.

v1.6.5.1-ce — 2026-05-17

MCP+14, AUTOML.PREDICT planner fixes, SortExec fix, recipe-builder. Manual + prompt + tier + UX gate.

v1.6.5-ce — 2026-05-16

AutoML out-of-the-box.

  • Default Auto-mode AutoML returns calibrated, polarity-correct probabilities end-to-end on realistic data with no operator hand-tuning.
  • 100K rows train in 11s on CPU; predictions span [0.0001, 0.9989] with 444/500 distinct values on a Gold cohort.
  • MCP-in-gateway, Gemini provider, Cypher fixes. 20 commits across AutoML perf + correctness, infrastructure, docs.

v1.6.0–v1.6.4

v1.6.4-ce — 2026-05-12

Vector / RAG path correctness + install portability. Closes 9 vector-path bugs.

  • VECTOR columns no longer read back as NULL — four sibling JSON-to-Value converters all coerce numeric JSON arrays to Value::Vector.
  • CREATE TABLE … VECTOR(N) preserves the declared dimension for every N — the Custom-type token parser previously fell back to 384 for any shape other than the strict three-token form.
  • INSERT and UPDATE accept literal vector arrays; bulk UPDATE with EMBED(col) evaluates per-row and persists.
  • Dimension validation at write time catches embedding-provider swaps before they silently degrade COSINE_SIMILARITY to zero.
  • COSINE_SIMILARITY returns a real float in SELECT projections; IS NULL/IS NOT NULL agree across WHERE and projection contexts.
  • SELECT … WHERE pk = lit no longer errors with Index '<t>_pkey' not found when the PK index isn't materialized — falls back to TableScan.
  • provider = "native" config actually registers the bundled MiniLM as an embedding provider (was silently skipped).
  • data_dir sweep: every active storage path routed through a published data_dir module — gateway is now drop-in portable on any install layout. Schema cache files, schemas RocksDB, columnar storage, partition catalog, statistics DB, procedure/trigger registry, users repo, AutoML store, ai_chat schema lookup all consume it.

v1.6.3.1-ce — 2026-05-11

Docker/config hotfix on top of v1.6.3.

v1.6.3-ce — 2026-05-11

Cypher dispatch on /v1/query/execute dual-parser path + /v1/graph/match/profile fixes + /v2/graph/graphs + SHOW PROPERTY GRAPHS dispatch + COSINE_SIMILARITY null-in-projections + gateway defensive 404 JSON for /multimedia/* without /v1/ + CALL db.labels() / db.relationshipTypes() syntactic sugar.

v1.6.2.2-ce — 2026-05-11

Chat agent fix #2.

v1.6.2.1-ce — 2026-05-11

Chat agent stop-token hotfix — agent responses were running into delimiter tokens.

v1.6.2-ce — 2026-05-10

Frontend fixes + Graph Schema panel.

  • Image gallery preview silently blank: getMultimediaUrl emitted /multimedia/... (no /v1/ prefix), which the gateway's SPA fallback returned as 200 OK + index.html. PreviewModal wrapped the HTML in a Blob; the <img> rendered nothing. Both preview AND download URL builders fixed.
  • Cypher block with scalar projection rendered "no graph data": the Cypher branch now runs the same SQL→graph transformer the SQL branch uses. Graph-traversal Cypher renders as cytoscape graph; aggregation/projection Cypher renders as table.
  • New: Graph Schema panel on the Graph Explorer page. Auto-fetches distinct labels + relationship types, renders them as clickable pills.
  • Build: linux-x86_64-cuda removed from the release matrix until the CUDA inference path is stable end-to-end.

v1.6.1-ce — 2026-05-08

Hardens against the v1.6.0 silent breakage. Every bug here passed the v1.6.0 recipe smoke harness because the harness only checked HTTP 200, not data state.

  • DEFAULT clauses dropped on INSERT (still fixed; gated here): CREATE TABLE … created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP then INSERT stored NULL. Root cause: transactional_executors.rs + indexed_transactional_executors.rs skipped default-fill; only legacy executor.rs called it.
  • WHERE col = literal returns 0 rows (still fixed; gated here): top-level equality on non-PK columns was misrouted by a Wave-8 IndexScan planner rule to a non-existent index. NOT(…), <>, IN, LIKE, AND id > 0 AND name='Alice' all worked.
  • CUDA tarball aborts at boot (still fixed; gated here): ggml-cuda's CUDA_CHECK calls abort() (real SIGABRT) on driver/toolkit mismatch. Fix: child-process canary; on probe failure the gateway boots and serves SQL with only native LLM endpoints disabled.
  • CSV import silently no-ops: POST /v1/data/import/csv returned {success: true, rows_imported: 0} then SELECT errored "table does not exist". Tenant-context mismatch between import handler (no tenant scope) and SELECT path (JWT-scoped). Fixed.
  • Image gallery preview blocked by CSP: default() + development() CSP presets were missing blob: in img-src/media-src/worker-src and Google Fonts hosts in style-src/font-src. The strict() preset already had them — wrong preset selected at runtime.
  • Cypher MERGE returns nodes only, no relationships: executor populated columns/rows only with named identifiers (nodes); auto-created relationships, while stored, were absent from the response. Graph Explorer couldn't draw edges it didn't have.
  • SQL editor renders nothing for Cypher MERGE responses: frontend now detects node-shaped cells in a SQL response and transforms them to {nodes, edges}.
  • New release gate: scripts/qa/feature_validator.py — comprehensive state-asserting test runner. Replaces HTTP-200 smoke as the release ship gate. No release tag until 0 failures.

v1.6.0.1-ce — 2026-05-08

Three P0 engine hotfixes. Same three bugs gated into v1.6.1 — see above. Both v1.6.0.1 and v1.6.0 are superseded; upgrade to v1.6.1+ directly.

v1.6.0-ce — 2026-05-08

Recipe gallery 95.9% pass + license-aware CE. Shipped with three P0 regressions (DEFAULT clause silently dropped, WHERE col = literal returns 0 rows, CUDA tarball aborts at boot). Do not deploy v1.6.0 — upgrade to v1.6.1+ immediately. v1.6.0 was the trigger for adding the state-asserting feature_validator.py gate.

v1.5 series

v1.5.3-ce — 2026-05-07

License acceptance + SDK 0.2.0 merge.

  • LICENSE.txt embedded in the binary (include_str!). synapcores --license prints full text and exits.
  • First-run click-through prompts on TTY; non-interactive (Docker, systemd) requires --accept-license or AIDB_ACCEPT_LICENSE=1.
  • <data_dir>/license-accepted.json mode 0600 with sha256 anchor; re-prompts only when license text changes.
  • @synapcores/sdk@0.2.0 (Node, scoped public) + synapcores 0.2.0 (Python, first PyPI publish) source in tree.

v1.5.2-ce — 2026-05-07

Cypher variable-length relationship binding fix. Single fix: var-length relationship variable now binds to a List<Edge> instead of a single Edge, so Cypher quantifiers (ALL/ANY/NONE/SINGLE) and list functions (relationships, reduce) work over var-length patterns. fraud_ring_detection 5/5 (was failing block 2).

v1.5.1-ce — 2026-05-06

Hotfix.

  • execute_saved_recipe stub returned fake success without running anything.
  • CSP img-src/media-src missing blob: blocked upload preview thumbnails.
  • Multimedia GET responses missing Cross-Origin-Resource-Policy blocked gallery render under COEP: require-corp.
  • execute_cypher_for_tenant only accepted MATCH; MERGE/UNWIND/UNION recipes failed.
  • Multi-clause MERGE in one block lost variable scope (parser + executor + splitter chain).
  • Default --log-level changed from info to error. 17 misclassified tracing::error! calls downgraded to debug!.
  • tracing-subscriber switched to EnvFilter so RUST_LOG works for per-target overrides.

v1.5.0-ce — 2026-05-06

Perf defaults + multimodal vision + UI capability disclosure.

  • Performance defaults flipped: AIDB_ENABLE_RESULT_CACHE=true (was false), AIDB_OPTIMIZER_VERSION="unified" (was "legacy"). 102× oltp_point_select gain available zero-config.
  • Multimodal vision (NEW): CompletionProvider::complete_multimodal() with OpenAI / Anthropic / Ollama adapters. ImageProcessor::vision_analyze dispatched by AIDB_VISION_PROVIDER env var. /v1/system/vision admin routes for config; API key encrypted at rest with AES-256-GCM (HKDF-SHA256 from AIDB_JWT_SECRET).
  • Licensing: Tenant-isolation license-locked in CE binaries via #[cfg(not(feature = "multi-tenant"))]. Config-load forces enabled=false with a tracing warn — EE feature can't leak into a CE install regardless of TOML contents.
  • Routing: V2 API routes mounted by default (decoupled from tenant_isolation.enabled). V2 gains /multimodal + /vector-algebra parity backports. Removed duplicate /query/prepare handler that caused gateway panic on every startup since perf-wave-8 P2.
  • Frontend: PlatformCapabilitiesBanner on Dashboard + Filesystem Collections pages — 7 capability rows showing what's ready vs. what needs setup. Settings → Vision Provider tab.

v1.3 series

v1.3.1-ce — 2026-04-30

Debian 12 install fix.

  • install-ce.sh was mapping ubuntu:22.04 and debian:12 to the same FFmpeg 4 tarball variant. They are different ABIs: Ubuntu 22.04 → FFmpeg 4.4 (libavutil.so.56), Debian 12 → FFmpeg 5.1 (libavutil.so.57), Ubuntu 24.04 → FFmpeg 6.x (libavutil.so.58). CE has builds for first + third only.
  • check_distro now splits these. Debian 12 gets a clear "use Docker" message with the exact docker run command and exits early.
  • Bootstrap's edition-check (synapcores --version | grep Community) was running BEFORE deps were installed → always emitted "binary did not report 'Community' on --version" on a fresh box. Removed; check moved post-deps with an ldd hint.

v1.3.0-ce — 2026-04-30

Intel Mac dropped from release matrix. macos-13 runner allocation on personal-account repos was unreliable (3+ hour waits). darwin-aarch64 (Apple Silicon) is ~80%+ of the 2026 Mac install base; Intel Mac users have a working Docker fallback documented at Run on macOS. Bootstrap's detect_platform() now exits early on darwin/x86_64 with a clear Docker hint instead of 404'ing on a missing tarball.

v1.2 series

v1.2.1-ce — 2026-04-29

Docker tag-vs-binary version grep dropped. The Dockerfile's pre-ship verification was comparing SYNAPCORES_VERSION (release tag, e.g. v1.2.1-ce) against the binary's --version output (Cargo workspace version, e.g. 0.1.0). These are independent by design — every release tag was failing the docker build with binary version mismatch: expected v1.2.1-ce even though the binary was correct. The Community edition grep is kept.

v1.2.0-ce — 2026-04-29

FileSystemRAG GA. See the Filesystem Collections page.

v1.0

v1.0.0-ce — 2026-04-28

Community Edition GA release. First public CE tag.

How to upgrade

For any version older than two releases back, the safest path is:

  1. Stop the running engine.
  2. Back up <data_dir> (the RocksDB volume).
  3. Run the installer one-liner — it always grabs the current :latest.
  4. Restart with your existing config. The new binary will read the old data_dir in place; no migration step exists between any two CE tags.

If you're on v1.6.0 specifically, jump straight to v1.6.1+ — three P0 regressions shipped in v1.6.0 (DEFAULT dropped, equality returns 0 rows, CUDA SIGABRT).

If you're on anything before v1.6.6.3 and running on a non-AVX-512 x86_64 CPU (i5-10400F, older Xeons), upgrade — v1.6.6.3 is the first release with the AVX2-baseline guarantee (no SIGILL on boot).

If you upgraded from v1.6.5.5 or earlier with split-mode Ollama embedding configured, re-embed your vector collections after upgrading to v1.6.5.6+ — earlier versions silently fell back to the chat model's hidden state for EMBED().