Skip to content

Releases: openagent-uno/openagent-core

Core beta41: same-task MCP partial-entry rollback

Choose a tag to compare

@geroale geroale released this 31 Oct 01:03

Core beta41 fixes partial MCP connection entry: transport and ClientSession contexts unwind on the same supervisor task when entry fails or is cancelled. Session teardown failures no longer skip transport teardown. No uncertain tool mutation is retried.

Qualification: 669 tests and 97 subtests passed; focused MCP suites 87 passed; 15 installed-wheel regressions passed. Wheel RECORD/source equality, all397 packaged Python files under Python3.11 grammar, and fresh installed imports were verified.

Four new distributions: Core, SQLite storage, gateway and Python SDK1.1.0b41. Capability-host1.0.0b3 is reused byte-for-byte from public beta40 with source equality proof. Modules beta37, Tools beta6 and signed HostB7 bundles remain unchanged. Base Core dependencies remain empty.

The manifest binds the exact clean source, five wheel hashes and qualification. Product installed/device acceptance remains separate from package qualification.

Core beta40: deterministic SQLite connection lifetime

Choose a tag to compare

@geroale geroale released this 30 Oct 20:56

Capability-host1.0.0b3 closes all15 audit, idempotency and mutation-lease SQLite transaction connections deterministically, preserving commit/rollback and closing on setup errors. This fixes actual Windows Host release test failures caused by locked owned databases after shutdown, without GC workarounds or skipped coverage.

Core/SQLite/gateway/PythonSDK1.1.0b40 align Core pins. Core and SDK offer an optional capabilities extra requiring capability-host beta3; base Core dependencies remain empty. Independent modules retain their immutable beta37 versions. All five wheels come from clean commit b351396.

Qualification:663tests+97subtests passed,9installed-wheel lifetime regressions passed without source overlay,422wheelRECORDentries and packaged sources verified; isolated audio worker imports the installed Core40 wheel. Signed Host6/product qualification follows separately; previous candidates containing capability-host beta2 must be rebuilt.

Core beta39: exact MCP connection lifetime

Choose a tag to compare

@geroale geroale released this 30 Oct 17:55

Core beta39 binds capability references and MCP Apps views to an exact MCP connection generation, including reconnects with an identical tools/list catalog. Old views and cached replies fail closed; a fresh view works on the new connection. Credential, readiness, ping, resource I/O, durable admission and publication races cannot migrate or replay a mutation.

This release contains four new wheels (Core, SQLite, gateway and Python SDK 1.1.0b39). Capability-host 1.0.0b2 is reused byte for byte from beta38, with original provenance in manifest.json. No unchanged modules are rebuilt. The source commit, complete file hashes, qualification and SHA256SUMS are attached. Signed product installers and deployments remain separate release gates.

Core beta38: scoped capabilities and capability-host beta2

Choose a tag to compare

@geroale geroale released this 30 Oct 17:41

Core beta38 integrates opt-in instance-owned management, complete bounded MCP catalog refresh, scoped MCP Apps resources/model context and durable tool-display projection over current public main.

The integration preserves the published SQLite cancellation/draining, actionable run-error categories, canonical content envelopes and dead-session MCP reconnect fixes. The synthetic snapshot parent is excluded from release ancestry.

Changed distributions only:

  • openagent-core 1.1.0b38
  • openagent-storage-sqlite 1.1.0b38
  • openagent-gateway 1.1.0b38
  • openagent-sdk 1.1.0b38
  • openagent-capability-host 1.0.0b2, consuming independently published openagent-tool-protocol 1.0.0b2 from Tools beta5

Unchanged module distributions retain their exact beta37 versions, URLs and public bytes; they are not rebuilt or republished here.

Validation: 645 tests and 92 subtests passed in Python 3.12, including the real isolated audio worker from the same committed Core wheel against an owned local HTTP fixture. SDK/gateway, SQLite cancellation, MCP reconnect, management and capability-host regressions are included. No paid provider or customer audio was used.

Source: 54ea736, based on public main c77cf31. The manifest records clean-source provenance and exact wheel bytes/SHA-256 values. Product deployment, authenticated live-provider acceptance and signed desktop installers remain separate gates.

Core beta37 — safe SQLite search cancellation

Choose a tag to compare

@geroale geroale released this 26 Oct 21:26

Interrupt and drain SQLite index workers before connection close. Includes repeated cancellation and cancelled-open cleanup. Source and installed public suite: 506 tests and 90 subtests, with four cancellation regressions. Canonical schema unchanged.

Core 1.1.0-beta.36: actionable AI failures

Choose a tag to compare

@geroale geroale released this 26 Oct 20:50

Preserve safe AI failure categories through engine streaming and durable failed runs. Distinguish missing accounts/models, invalid or expired auth, credit/quota exhaustion, temporary rate limits, budgets and context size. Unknown failures keep no provider diagnostics. Empty model catalogs fail before publishing a fake assistant answer.

Validation: 502 public tests and 90 subtests passed. SQLite schema is unchanged.

Core 1.1.0 beta35: reliable explicit file delivery

Choose a tag to compare

@geroale geroale released this 26 Oct 19:15

A successful authenticated send-file request omitted from the terminal reply now asks the same model for a corrected final answer, at most twice and within the configured iteration limit. The model still authors the exact markers. Preview media are never promoted into final output; normal session, artifact and recipient authorization remain required. Persistent omission fails clearly, cancellation is retained, and nested/concurrent runs remain isolated.

Verified from the source and a clean normal wheel installation: 132 tests, 26 subtests, and 721 first-party RECORD file hashes. Real provider/channel and native product qualification remain separate.

OpenAgent Core 1.1.0 beta34

Pre-release

Choose a tag to compare

@geroale geroale released this 26 Oct 11:43

Interactive document and application work no longer silently stops after sixty tool calls when no host or operator ceiling is configured. Explicit execution policies, operator ceilings, lean background limits, deadlines, cancellation and authorization remain authoritative.

Built locally from clean source f2df941. Installed Core34 and channel bridges14 integration: 377 tests and 98 subtests passed, all 24 package RECORDs verified, pip check passed. Live provider/channel acceptance is separate and is not claimed by these deterministic tests.

Core 1.1.0 beta33 — canonical stream attachments

Choose a tag to compare

@geroale geroale released this 25 Oct 16:36

Canonical runtime media now survives the streaming gateway: clean text is no longer reparsed for consumed file markers. The authored final selection is revalidated against current session artifacts without promoting previews. Stream generators close in the originating task after terminal publication, retaining privacy checks.

Fresh normal installation and focused runtime, privacy, storage, SDK and media regression tests passed. Real Telegram acceptance is performed by the standalone product release.

OpenAgent Core 1.1.0 beta32 — reliable model catalog writes

Choose a tag to compare

@geroale geroale released this 25 Oct 10:08

Catalog writes alongside runtime persistence

Model-manager keeps a shared read connection. A still-open catalog reader can retain a WAL snapshot while another runtime connection commits. Promoting that snapshot into a writer raises SQLITE_BUSY_SNAPSHOT immediately; increasing busy_timeout or replaying the same tool call cannot repair it.

Provider/model upserts, enable/default flags and removals now use short dedicated BEGIN IMMEDIATE transactions with the configured SQLite lock timeout. Validation, mutation and ID lookup share that transaction, which commits or rolls back before closing. Existing atomic model/provider edits retain their dedicated transactions. Other MemoryDB writers are unchanged. Concurrent first MCP calls wait until the singleton database has finished initialization; failed initialization is closed and never published.

Regression tests retain a live reader, advance the WAL on an independent connection, then exercise additions and catalog mutations. Parallel MCP additions and a rejected provider reference verify subsequent writes remain usable. Initial MCP connection publication is tested separately.