Release Notes 0.5.35

This page summarizes the product and engineering changes included in 0.5.35, following v0.5.34.

Summary

Release 0.5.35 expands CoWork's durable runtime model in three visible directions: reusable managed agents, optional external-memory integration through Supermemory, and much better trace/debug visibility for task execution. It also simplifies the main-task turn-budget model, adds a new bundled novelist skill, improves task-event playback and daily briefing quality, and fixes the clean-install regression that had broken npm run release:smoke.

New Features

• Managed Agents and Managed Sessions: CoWork now includes a first-class managed-agent layer with versioned definitions, durable managed-session plumbing, control-plane methods/events, and Mission Control/runtime surfaces that observe the same underlying task graph. Learn more
• Optional Supermemory integration: Supermemory can now operate as an external memory lane with setup flows, tool handlers, runtime metadata, policy exposure, Memory Hub controls, and prompt-time profile context injection while CoWork's local memory system remains primary. Learn more
• Task Trace Debugger: the app now ships with a trace-debugger surface for inspecting task trace requests, projections, formatting, and renderer-side trace tabs. This makes runtime reasoning and failure triage much more visible during long or complex tasks.
• Bundled novelist skill: the default skill bundle now includes a novelist workflow and related starter-mission support for narrative and long-form drafting use cases.
• CoWork School: the docs now include a beginner-friendly CoWork School guide to help users ramp into the product's runtime model and core workflows more quickly.

Enhancements

• Explicit-only turn budgets: main interactive tasks no longer inherit implicit strategy-derived maxTurns windows. Tasks are uncapped by default unless a flow sets maxTurns or windowTurnCap explicitly, while lifetime safety fuses and existing recovery safeguards stay in place. Learn more
• Clearer runtime ownership and telemetry: runtime/session docs and telemetry now distinguish explicit capped runs from default-unbounded execution more clearly, which makes Mission Control, support, and debugging output more truthful. Learn more
• Renderer task-event playback: task-event handling now batches, throttles, derives replacement state more aggressively, and includes perf-oriented fixture coverage so larger task histories remain smoother to browse and replay.
• Mission Control and settings surfaces: Mission Control board actions, labels, and card summaries were refreshed, and Settings/Memory Hub now expose Supermemory controls, trace-debugger settings, and provider-specific failover controls more cleanly.
• Briefing quality: daily briefings now prioritize higher-signal context and filter low-signal automation noise more aggressively, so summary output stays closer to what operators actually need.
• Provider failover configuration: the LLM stack and settings surface now support provider-specific failover chains and clearer failover behavior.
• Release validation: packaging guidance and artifact checks now better enforce consistent release filenames and local installer validation.

Fixes

• Clean release smoke installs: the Electron runtime was restored to installed dependencies so npm run release:smoke no longer falls back into dependency bootstrap during a clean consumer install.
• Chat-session MCP discovery: MCP tool discovery now works correctly inside chat sessions.
• Workspace path recovery: stale absolute file paths can now be remapped into the active workspace more reliably during file reads.
• Executor/runtime edge cases: completion evidence handling, pending skill-parameter handling, iCloud execution routing, and permissive JSON Schema handling for tool definitions were tightened.
• Remote shadow-task cleanup: remote task pruning now deletes only rows that are actually covered by the fetched remote window.

Upgrade Notes

• Long-running main tasks now behave differently by default: unless a workflow explicitly sets maxTurns or windowTurnCap, the run is treated as uncapped and governed by lifetime and recovery safeguards instead of an implicit turn window.
• Supermemory remains optional and additive. Existing local memory, recall, and prompt behavior continue to work without it.
• Local macOS packaging on machines without Developer ID Application credentials or notarization environment variables still produces ad hoc signed validation builds rather than notarized distribution builds.

References

• Changelog - full version history
• Managed Agents - managed agent lifecycle and runtime model
• Supermemory - external memory integration
• Execution Runtime - task execution and turn-budget behavior
• Session Runtime - ownership and task-session state model

This page is the canonical high-level summary for the changes included in 0.5.35.