in browser inspection, selection and commenting functionality

• lets have option to put it as claude comment
• then we can open the comment icon to view the message
• also we can basically batch and send it
• we can send it in batch but as queue

• markdowns, files etc to have option to comment like claude kidna

• ability to revert from sent chat like git revert the changed things

• highlevel map scrolling option https://x.com/darrenjr/status/2089755857382744166?s=20

• hover to view changed files next to fork when chat agent is done https://x.com/darrenjr/status/2089755857382744166?s=20

• search view analyze, AI sessions https://jazzyalex.github.io/agent-sessions/

• login from inside the vscode, also install agent interminal from inside out editor

• port forwarding on agent window to view the build in other device

• start once reset is done option to continue task once reset is back scheduling

• schedule a task option

• voice insert option (optional)

• / inside the editor text area

• Questions, Plan mode recommendation, Work progress

• SubAgent rendering view

• show sub agents as nested tabs as such

Progress animation for tasks https://x.com/wherescz/status/2101478969170092113?s=20

Landing Page: https://wrkr.dev/

• Doom-loop detection
• Context compression
• Context management
• Tool calling
• Tool selection
• Tool routing
• Intent classification
• Task decomposition
• Planning and replanning
• Goal tracking
• Progress tracking
• State management
• Sub-agent orchestration
• Agent swarm coordination
• Parallel task execution
• Agent delegation
• Agent handoffs
• Agent specialization
• Tool-result interpretation
• Error recovery
• Retry strategies
• Strategy switching
• Self-correction
• Self-verification
• Output validation
• Test-driven iteration
• Codebase exploration
• Repository understanding
• Dependency analysis
• File relevance ranking
• Symbol-aware navigation
• Semantic search
• Code search
• Terminal execution
• Background processes
• Long-running task management
• Checkpointing
• Session persistence
• Memory management
• Working-memory management
• Context prioritization
• Token budgeting
• History summarization
• Context pruning
• Relevant-context retrieval
• Knowledge retrieval
• RAG orchestration
• Structured tool outputs
• Tool permission management
• Tool failure handling
• Timeout handling
• Cancellation handling
• Concurrency control
• Resource management
• Cost management
• Latency optimization
• Streaming responses
• Incremental rendering
• Event-driven architecture
• Agent lifecycle management
• Task lifecycle management
• Human-in-the-loop
• Approval gates
• Safety guardrails
• Sandbox isolation
• Secret protection
• Command risk assessment
• Change impact analysis
• Git integration
• Checkpoint commits
• Diff inspection
• Rollback support
• Automatic testing
• Build verification
• Lint verification
• Type verification
• Runtime verification
• Browser automation
• UI verification
• Screenshot analysis
• Visual regression testing
• Observability
• Tracing
• Agent telemetry
• Tool-call telemetry
• Token usage tracking
• Performance metrics
• Failure analytics
• Structured logging
• Decision tracing
• Explainable execution
• Adaptive model selection
• Model routing
• Fallback models
• Fast-model escalation
• Reasoning-model escalation
• Prompt optimization
• Dynamic system prompts
• Skill loading
• Capability discovery
• MCP integration
• ACP integration
• Plugin orchestration
• Tool discovery
• Environment discovery
• Workspace awareness
• Multi-project awareness
• Cross-file reasoning
• Cross-agent communication
• Shared agent state
• Conflict resolution
• Result aggregation
• Evidence tracking
• Confidence estimation
• Uncertainty detection
• Ambiguity handling
• Completion detection
• Goal satisfaction detection
• Stuck-state detection
• Infinite-loop prevention
• Novelty detection
• Progress heuristics
• Strategy evaluation
• Outcome-based iteration
• Automatic task recovery
• Graceful degradation
• User interruption
• Resume capability
• Background agent execution
• Parallel exploration
• Speculative execution
• Caching
• Semantic caching
• Tool-result caching
• Context caching
• Prompt caching
• Deterministic replay
• Execution replay
• Audit trail
• Session branching
• Task branching
• Agent checkpoints
• Final-result synthesis
• Result presentation
• Rich UI rendering
• Artifact generation
• Interactive progress UI
• Live execution timeline
• Human-readable agent traces

Nearly every feature from the last two weeks is untested at runtime: durable resume, the self-healing orchestrator, keep-working, harness hardening, channel switching, the Terminals tab, the design canvas, and the agent server. The repo has no benchmark or replay suite, only unit tests, and some of those already fail at HEAD. So we can't tell whether a change helps long runs or quietly breaks them.

Build a small headless set of real tasks (10–20 repo tasks across short, multi-hour, crash-in-the-middle and approval cases). It runs against the agent server and scores pass/fail, steps, tokens and recoveries. Durability features can't be built on without being measured, and "agents lose hope after 2 hours" is exactly the kind of bug only this would catch.