Netclaw 0.24.2-beta.1

Netclaw v0.24.2-beta.1 — Schema-aware meta argument coercion, reliability fixes, and performance improvements

Features

• Schema-aware meta argument coercion — ChatGPT-trained models (Qwen3.6, etc.) were being hard-rejected for using non-canonical meta argument names (TimeoutSeconds vs _timeout_seconds). This adds schema-aware near-miss resolution that coercively maps model-supplied key variants to canonical names when no real tool parameter shadows them. Fixes "tool was NOT executed" errors that pushed models off tools and into sandbox-style output. (#1470)

• Preserve per-call meta hints across tool-batch re-drive — Tool calls awaiting approval that passivate and cold-recover were losing meta hints (_timeout_seconds, _rationale, _background), silently falling back to defaults. Fixes a pre-existing bug where the re-drive path never reinjected persisted MetaJson. Long-running tools approved after session recovery now respect the original timeout.

Bug Fixes

• Subagents: honor approval-pause in every tool-watchdog mode — In WallClock mode (applied to every opaque tool), the human-approval pause was ignored: the wall-clock budget kept ticking through a human approval, killing healthy tools mid-wait when humans took longer than the budget. Prevents premature tool termination during human-in-the-loop approvals. (#1473)

• Skills: block agent mutations to server-feed skill directories — Agents could freely edit/delete/write files into .server-feeds/ skill directories via both skill_manage and direct file-write tools. Changes were silently overwritten on the next sync cycle. Two enforcement layers added: skill_manage guard + ToolPathPolicy write-deny for server-feed paths. Prevents wasted agent effort and misleading session state from server-feed mutations. (#1466)

• Subagents: record spawn lifecycle in the session transcript — Sub-agent actor logs went through Akka's async logger bridge and were lost because SessionDiagnosticsContext.AsyncLocal was gone. Spawns that failed early (guard rejection, watchdog kill, child-actor failure) left the session transcript completely silent. Parent-side breadcrumbs now log spawn lifecycle under the parent session scope. (#1468)

• Providers: refresh inference OAuth tokens — OpenAI Copilot and other inference providers with OAuth tokens now refresh tokens at runtime rather than relying on stale tokens. Includes Copilot probe-time token refresh. Prevents authentication failures during long-running sessions when tokens expire. (#1465)

Performance

• Optimize LlmSessionActor storage — Only retain the most recent snapshot + last N messages in the journal, instead of storing full history. Reduces memory consumption for long sessions with many messages. (#1464)

Dependency Updates

• Bump MessagePack from 2.5.301 to 3.1.7 — Major version bump with serialization improvements. Please test your workflows after upgrading. (#1420)

• Bump Anthropic SDK — 12.29.1 → 12.30.0 (semver-minor). (#1460)

• Bump Aspire.Hosting.Testing — 13.4.4 → 13.4.6 (semver-patch). (#1462)