v2.3.0

Highlights

New capabilities

• Added deep_research for breadth-first web research.
• visit_webpage can now use LLM-powered web content transcription.
• Added HEARTBEAT.md-based periodic self-directed agent invocation.
• Added arc events (cron) for scheduled agent tasks.
• Added persistent agent memory via /workspace/MEMORY.md.
• Added agent-driven automatic skill creation.

Session/runtime controls & UX

• Replaced maxIterations with token/cost session limits.
• Added per-mode toolSummary control and support for inline in-session tool summaries.
• Intermediate agent responses are now sent back to callers.
• Persistence summaries now include write and edit tool calls.
• Steering got more robust: active-session highlight steering, explicit ! cross-mode warnings, and safer wrapped steered messages.

Reliability, logging, and cost tracking

• Cost reports now attribute usage more accurately and include safer backfill tooling.
• Response logs now show context-fill usage and model IDs more consistently.
• Per-message/event logging was tightened up, including proactive/session logging.
• System prompts now surface workspace skill diagnostics.
• Fixed several noisy/broken reply cases, including NULL-sentinel filtering and IRC newline handling that could break URLs.

Dependencies & defaults

• Upgraded pi dependencies to 0.56.2.
• Updated the default non-vision model from K2.5 to GLM-5.
• Included further prompt and anti-sycophancy tuning.