v0.9.23

[0.9.23] - 2026-03-12

Added

• Non-resolving confirmation shortcut: Bare "yes"/"ok"/"go ahead" replies to assistant questions that require a specific answer (which/what/or-choices) now get a direct follow-up requesting the actual choice instead of entering the full agent loop.
• Progress-based budget extension: Each successful tool execution extends the budget (LLM calls, tool calls, steps, wall-clock time), so productive multi-step runs are never artificially stopped by the initial budget ceiling.
• Successful tool calls count as concrete work: The stopping phase now considers any successfully completed tool call as concrete execution, preventing premature abandon when tools execute but their effects don't classify as observation/mutation.

Changed

• Pre-execution planning no longer charges budget (system-initiated quality check, not agent work)
• Budget tier limits raised: None tier LLM calls 8→14, validation rounds 1→3; Small tier steps 12→16, LLM/tool calls 10→14, validation rounds 2→3
• Clarification question detection expanded ("want me to", "shall i")

Fixed

• Productive runs stopped by tight budget even though every tool call succeeded

Full Changelog: v0.9.22...v0.9.23