Skip to content

Netclaw 0.24.2

Choose a tag to compare

View all tags
@github-actions github-actions released this 25 Jun 20:19
0.24.2
This tag was signed with the committer’s verified signature.
Aaronontheweb Aaron Stannard
SSH Key Fingerprint: hcX7BX3nNQgJtL086k1oGNpa5IrSmU0ed/io3ryQL08
Verified
Learn about vigilant mode.
3ff1555
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

0.24.2 2026-06-25

Netclaw v0.24.2 — DwarfStar provider support, sub-agent operating rule inheritance, systemd update reliability, and reliability improvements

Features

  • DwarfStar (ds4) provider support — Added DwarfStar as a supported provider via the OpenAI-compatible backend strategy. (#1349)

  • Sub-agents inherit embedded AGENTS.md operating rules — Sub-agents spawned by a parent session now inherit the embedded AGENTS.md operating rules from the parent's identity context, ensuring consistent operating constraints across the agent hierarchy. (#1490)

  • 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. Long-running tools approved after session recovery now respect the original timeout. (#1470)

Bug Fixes

  • Subagents: fixed premature kill by parent watchdog during approval waits — Sub-agents waiting for human approval were incorrectly killed by the parent session's inactivity watchdog. Sub-agents now manage their own liveness end-to-end and are no longer subject to parent-side timeouts. (#1481)

  • 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)

  • 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)

  • 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)

  • 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)

  • netclaw update no longer crash-loops systemd on Linux — On Linux hosts with Netclaw installed as a systemd --user service, netclaw update was restarting the daemon as a detached process (bypassing systemd). This left the still-enabled systemd unit in a permanent crash-loop — every 5 seconds it tried to start, failed because the detached daemon held the exclusive lock file, and retried indefinitely (restart counters in the tens of thousands). netclaw update now delegates start/stop to systemctl --user when an enabled user unit exists. (#1469)

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 Termina from 0.14.0-beta3 to 0.14.0 — TUI framework promoted to stable; includes a fix for the MCP tool permissions page layout. (#1480, #1483)

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

Assets 10
Loading