mateclaw/CHANGELOG.md
2026-10-26 22:04:10 -04:00

5.3 KiB

Changelog

3.0.0 — 2026-10-26

3.0.0 upgrades to Java 21, Spring Boot 4 and Spring AI 2. AgentScope 2 drives ReAct and plan-step execution loops; MateClaw retains approvals, permissions, Goals, SSE and persistence.

Release date: 2026-10-26. These notes cover the 3.0.0 stable release. Source and package versions are 3.0.0; download published packages from GitHub Releases.

  • MyBatis Plus 3.5.17 with the Boot 4 starter; MCP SDK 0.18.3.
  • Native DashScope chat/embedding adapters; native speech SDK retained.
  • Owned HTTP cancellation, guarded tool batches and durable Goal checkpoints prevent blind side-effect replay.
  • Existing frontend interaction contracts and conversation history remain supported.
  • Shared prompt budgets inherit model defaults and count effective tool schemas plus one output reserve; AgentScope remains on stable 2.0.4.
  • Latest budget regression: 342 focused tests and 6,509 full-suite tests (50 skipped) with no failures/errors. Real-model acceptance: 18/20 after manual review; tool argument and progress-count failures remain documented in the detailed notes.

English details · 中文说明

Finalization, team delivery and deployment

  • Long tasks check completion before returning a final answer and can continue unfinished work within runtime limits. Requests that need user input wait for a reply instead of repeatedly starting action turns (#668).
  • When reasoning consumes the response output budget, final-answer generation can retry within runtime limits. DeepSeek Flash thinking controls are applied consistently (#672).
  • Team worker progress retains full result evidence and runtime failure causes. A Team Run publishes completion after final synthesis, keeping worker output, progress and the delivered summary consistent.
  • DSH keeps its reviewed default SDK version; 0.2.1-rc.2 is an opt-in candidate. Docker runtime images pin Node.js 22 and include bubblewrap for Linux sandbox execution (#670). Candidate checks on macOS arm64 do not establish Linux or Windows acceptance; see DSH setup and limitations.
  • Every backend requires a non-empty JWT_SECRET of at least 32 UTF-8 bytes. Desktop local mode generates and persists its secret; source and server deployments must configure and retain theirs. See configuration.

Cumulative changes since 2.3.0

3.0.0 includes the former 2.4.0 development cycle; 2.4.0 was not published as a separate stable release. The groups below cover capabilities added before the architecture migration as well.

Workspaces and conversations

  • Member file isolation is an opt-in strict sandbox. With MATECLAW_WORKSPACE_SANDBOX_MEMBER_ISOLATION_ENABLED=true, authenticated user and workspace identities determine the file root; missing identity denies access without shared-directory fallback. Only six audited builtin file/code tools are available; DSH, MCP/plugin tools, shared skill scripts and unauthenticated channel identities are unavailable. This differs from personal memory-file fallback. (#644).
  • Persistent channel conversations expire stale query context. Uploaded images resolve from conversation storage; directory uploads retain directory metadata and validate containment.
  • Employee drafts accept JSON following reasoning blocks, and explicit tool opt-out is respected even for general capabilities (#648, #655).

Structured decisions, Goals, plans and teams

  • A bounded typed Choice / Boolean / Score decision pipeline adds durable audit, OFF / SHADOW / ACTIVE modes and scenario overrides. SHADOW is the default; suggestions cannot bypass domain constraints.
  • Goal continuation, plan employee routing and team worker-result judgment share the decision foundation. Domain transactions commit outcomes and state; delayed decisions validate identity and fences.
  • Richer routing context, worker contracts, delegated-step result checks and bounded Goal recovery retain existing completion gates and record applied follow-up suppression.

Models and DeepSeek Harness

  • DSH moves to the SDK runtime with pinned versions and controlled upgrades. Failed health checks invalidate cached health and broken settings remain editable (#650).
  • Model probes validate completed responses rather than accepting connection success or empty output (#658).

Read-aloud, chat and tool presentation

  • Edge and DashScope read-aloud are restored, with provider voice lists, selection and preview (#646).
  • Structured MCP results get dedicated presentation below the final answer. Ordinary tool results no longer automatically become final-answer text (#651, #656, #661).

Channels and database operations

  • Application-owned Feishu WebSocket readiness and lifecycle release owned resources on shutdown and reconnect (#659).
  • Unsupported webhook channels are rejected before configuration persistence (#654).
  • Flyway validation failures block startup by default; history repair requires explicit opt-in, with backup and recovery guidance (#657).
  • Commercial-edition overview and links are added; commercial modules are not included in the open-source edition.