Changelog

What's new in the stack.

Recent releases from all six Open Assistants projects — including the results we'd rather publish than hide.

assistant
v0.6.37Oct 7, 2026

Trigger runs are session runs.

  • Webhook, cron and file-change runs take the session claim and queue behind an active turn, bounded by a timeout that reports SessionBusyError instead of interleaving with the live loop.
  • Registration is refcounted — a finishing trigger run can no longer evict a loop another live run still owns.
  • TriggerRegistry.fire returns the agent's final answer, surfaced in the /webhooks/{trigger_id} response — the callback path the event-driven-agent spec described.
assistant
v0.6.37Oct 7, 2026

Breaking: the bundled mail, contacts and todos families are gone.

  • The built-in email, contacts and todos families, storage adapters, configuration, IMAP dependency and /emails, /contacts, /todos APIs are removed — plus the obsolete Gmail demo. Review dependent applications before upgrading.
  • Existing stores are retained unread: no automatic export, migration or deletion, and a fresh server creates none of the retired directories.
  • Use the existing app_* tools for structured data, or explicitly install the optional tasks/contacts/reading-list starter templates. Generic ConnectKit connectors, file sync, coding and browser capabilities remain — there is no first-party mail replacement.
assistant
v0.6.36Oct 6, 2026

Guardrail residuals closed.

  • Reasoning is checked, not just the answer: a clean answer can no longer flush blocked reasoning, and reasoning-only rounds are covered.
  • Unguarded deployments keep token-by-token streaming — the hold buffer now exists only when output guardrails are configured.
  • A deferred second user message now opens the next turn instead of being dropped.
assistant
v0.6.34Oct 6, 2026

Guardrails prevent, not alert.

  • Model content is held until the round's output guardrails pass — a tripped guard keeps blocked text off the client and out of the session log, instead of reporting it after the bytes were sent.
  • Streaming buffers text and reasoning per round and drops the buffer on a trip; tool-call narration is withheld while the guarded action still runs. Guardrail-enabled deployments trade per-token rendering for that guarantee.
  • The #74 residual is closed: the check runs before the message enters state, covering the ordinary answer, the post-nudge final answer, and tool-round narration. 8 regression cases, 5 of which fail on the previous code.
assistant
v0.6.32Oct 6, 2026

The 2026-10-02 audit batch ships.

  • Roughly 90 issues fixed across governance, boundaries, MCP, auth, transport, skills, storage and the loop — each reproduced failing before the change (suite: 3849 passed, 27 skipped).
  • Tool-authorization guards fail closed; approvals resolve through the live catalog and refuse a tool whose definition changed since the proposal; receipts keep UNCERTAIN/CANCELLED/REJECTED instead of reporting success.
  • Filesystem tools resolve only inside the caller's store plus operator-granted roots; webhook and schedule routes bind identity to the owner; OIDC claims are anchored on the subject with the session cookie set Secure.
assistant
v0.6.26Oct 2, 2026

Tracing consent defaults off.

  • config.yaml could force product tracing on over an explicit LANGFUSE_ENABLED=0, because the yaml bridge ran after construction. Env now beats yaml on the gate.
  • Consent default is OFF: a deployment opts in with LANGFUSE_ENABLED=1 or enabled: true, and an explicit =0 always closes the gate.
assistant
v0.6.24Sep 29, 2026

Scheduled subagents become reachable.

  • POST/GET/DELETE /subagents/schedules and GET /subagents/schedules/{id}/runs expose the durable scheduler; a subagent_schedule tool gives agents the same path under governed approval.
  • Schedules can only be created from a ready frozen launch plan, so unattended runs cannot be registered without the capabilities to execute.
  • Cross-user access returns 404, not 403 — the schedule ID never leaks whether it exists.
  • Trigger restoration keeps recurring schedules recurring; missed one-shots are reported as expired rather than run hours late.
ModelPanel
Gateway v0.7.0Sep 25, 2026

A failed command is no longer a successful receipt.

  • Tool and governance results now carry explicit outcomes: succeeded, refused, failed, timed_out, or killed — contradictory states are rejected.
  • Custom TOOL.md commands execute through the sandbox seam with real write, output, and timeout budgets; signal deaths are reported as failures.
  • Subagent launches fail closed when declared tools or skills are missing, disabled, denied, or approval-gated; terminal delivery is idempotent.
ModelPanel
Gateway v0.7.0Sep 25, 2026

The registry belongs to the user.

  • The gateway no longer ships a provider roster. Routes are added, previewed, updated, and deleted through the user's own registry.
  • An absent registry boots empty; invalid registry bytes still fail closed. Deleted routes carry tombstones so catalog refreshes do not resurrect them.
  • Provider differences stay in configuration while the gateway keeps one OpenAI-compatible adapter and the intrinsic fusion route.
CoreMem
v0.17.0Sep 25, 2026

Facts become governed, versioned memory.

  • Facts live in a versioned HybridDB table, so every write is on the audit chain and reversible with rollback_memory — no second source of truth.
  • add_fact, list_facts, get_fact, fact_history, supersede_fact, expire_fact, and merge_facts give agents deterministic CRUD and history.
  • Optional extract_facts ingest is off by default and opt-in only; it extracts candidates without changing the zero-LLM recall path.
HybridDB
v0.10.0Sep 30, 2026

Bundled MiniLM is the default engine.

  • The default vector engine ships a uint8-quantized all-MiniLM-L6-v2 inside the package, so no network or fp32 download is needed.
  • hash_embedding no longer serves as a silent fallback — it is exposed for accuracy harnesses only.
  • TEXT and INTEGER primary keys now resolve consistently so semantic search returns the rows it should.
HybridDB
v0.8.0Sep 3, 2026

One identity across SQLite, DuckDB, and Chroma.

  • Chroma vectors are keyed by the logical primary key rather than the physical rowid, removing the TEXT-PK delete/reinsert duplicate-key failure class.
  • migrate_vector_identity() copies legacy embeddings bit-identically, cleans orphan vectors, and is idempotent across longtext collections.
  • FTS ids, vector ids, and row fetches now resolve through the same primary-key identity.
ConnectKit
v0.1.6Aug 24, 2026

OAuth compatibility and tool descriptions that agents can use.

  • Google OAuth redirect_uri handling strips service query parameters for compatibility.
  • Connector tools can describe parameters with JSON Schema; parameter_descriptions is deprecated.
CoreMem
v0.16.1Sep 2026

Timestamp filters compare chronologically.

  • ts_after/ts_before now compare chronologically across mixed naive/aware formats — a naive 23:00 message is no longer excluded by an earlier UTC filter (the documented known limitation since 0.13.2).
  • Boundary semantics unchanged; unparseable values fall back to the legacy lexical comparison.
AgentProfile
v0.2.1Aug 26, 2026

Portable profiles, ready to publish.

  • README, AGENTS.md, refreshed strategy facts, and PyPI metadata make the package and its PROFILE.md contract easier to discover and use.
CoreMem
v0.16.0Sep 2026

Filtered recall can no longer starve.

  • Equality filters (role, session, user, agent) push into the Chroma scan before the vector search — adopting HybridDB 0.7.0's search(where=).
  • Demonstrated: with 150 same-phrase noise rows, the old over-fetch path returned empty for a session-filtered recall; pushdown returns the target as the top hit.
CoreMem
v0.15.0 / v0.14.0Sep 2026

Versioned memory — opt-in, then default.

  • Git-like versioned tables (HybridDB 0.6.0): recall against memory as of any point in time.
  • Versioned=True became the default in 0.15.0; HybridDB 0.7.0 is the floor.
HybridDB
v0.7.0Sep 2026

Metadata pre-filtering, lazy analytics.

  • search(where={...}) pushes scalar-column filters into Chroma before the vector scan — multi-tenant scoping with equality and $gte-style operators, in all modes.
  • DuckDB mirrors are created on first OLAP query instead of at open — tables you never query with olap cost nothing to maintain.
  • Long-document chunking helper: deterministic paragraph/sentence splitter with optional overlap, plus the chunks-as-rows ingestion pattern.
HybridDB
v0.6.0Sep 2026

Versioned, git-like tables.

  • Versioned tables bring time-travel reads and restores; update-heavy rollbacks are up to 41× faster via batched restores (measured in CoreMem's 0.14.0 perf gate).
ModelPanel
Gateway M2Sep 2026

Streaming, Anthropic, notices, fusion sidecar.

  • All three dialects stream: chat-completions with fallback-before-first-token, Responses typed events, and Anthropic Messages — E2E-verified against Claude Code.
  • modelpanel/fusion sidecar: task-routed panel recipes with streaming judge synthesis (Python stdlib, OpenAI-compatible on the wire).
  • Opportunity notices: when the served model also lives somewhere healthier, one header + body field says so — never an auto-switch.
  • Adapters: Groq, NIM, Cerebras, OpenRouter, Mistral, Gemini, LLM7; store:false respected end to end.
ModelPanel
Gateway M1Aug 2026

The core gateway, shipped.

  • Canonical request model, Responses + chat-completions ingress, OpenAI-compatible egress, tier-aware router with rate budgets, registry hot-reload.
  • Key vault with tier flags (free-permanent, free-credits, discounted, paid); CLI: serve, add-key, models, status.
  • Evidence loop: benchmark runs feed panel selection — panel synthesize 95.0% vs 90.0% best single on matched GSM1K.
CoreMem
v0.13.2Aug 2026

Filter plumbing, fixed end to end.

  • Every recall filter is now honored by every strategy — preference queries, session_cap selection, and fusion no longer leak messages from other sessions.
  • ingest() returns the turn_id actually stored; MCP ingest surfaces invalid metadata JSON instead of dropping it silently.
  • Use-after-close now raises; recall() rejects empty queries instead of returning arbitrary rows.
  • 12 regression tests from a full E2E bug hunt — suite at 175 passing.
CoreMem
v0.13.1Aug 2026

Memory hygiene, lifecycle, wider MCP surface.

  • New API: list_sessions(), delete_messages(), stats(), close() with context manager — pooled handles no longer wedge the vector store.
  • MCP server widened from 5 to 8 tools — delete, fetch_session, stats; recall accepts role, ts_after, ts_before, session_cap.
  • CLI gains stats and delete subcommands; recall output now includes message ids.
CoreMem
v0.13.0Aug 2026

Defaults validated by answer evals.

  • Bundle budget dropped to 4k chars with evidence-first ordering — 0.678 answer accuracy vs 0.608 at 16k, with ~60% less context.
  • Preference union and temporal query decomposition folded into the default episodic path (+0.033 / +0.037 session recall on S).
  • ingest_many() — batch ingestion 4.3× faster (550 messages: 49.9s → 11.5s) with identical retrieval.
  • session_cap opt-in: +0.124 message recall@5 at the cost of session recall — documented tradeoff.
HybridDB
v0.5.6Aug 20, 2026

The reliability release.

  • FTS5 indexes backfill after every rebuild — keyword search no longer silently returns zero results after reindex, drop_column, rename_column, or import_sql.
  • Journal processing is chronological with last-op-wins — fixes a permanent journal wedge on delete-then-reinsert.
  • DuckDB analytics supports custom-PK tables; stale broken mirrors are detected and rebuilt automatically at init.
  • read_query is enforced read-only at the SQLite level — the WITH-clause write bypass is closed.
  • GraphAPI exposes search_graph_ppr and sync_graph_nodes; graph sync rules are validated at registration.
  • insert/update honor explicit primary keys; drop_column refuses to drop the primary key.
CoreMem
v0.12.0Aug 2026

MCP server, CLI, hooks.

  • coremem mcp — stdio server; hook handlers for Claude Code and Codex; integration configs for Claude Code, Codex, and OpenCode.
  • CLI: recall, ingest, compile, rebuild, sessions, hook, mcp.
  • Zero-LLM recall stays the default — retrieval remains deterministic and LLM-free.
HybridDB
v0.5.5Aug 16, 2026

Graph fixes.

  • Synced graph node IDs are namespaced by table — two tables with the same keys no longer overwrite each other.
  • Personalized PageRank runs on the undirected subgraph — graph-aware retrieval actually works.
  • Edge rules work with custom primary key names; ghost nodes are removed on re-sync.
HybridDB
v0.5.4Aug 15, 2026

search_graph fix.

  • search_graph and search_graph_ppr returned [] for default integer-PK tables — the rowid alias collision is fixed.
HybridDB
v0.5.3Jul 31, 2026

PPR graph retrieval.

  • search_graph_ppr(k_seeds) — separate control over seed count vs. final results.
  • sync_graph_nodes() — public API for syncing table rows into graph nodes.

Full changelogs live in the repos.

Every release is documented in detail, including the bugs that took the longest to find.

HybridDBCoreMemModelPanelConnectKitAgentProfileassistant