v5.1.0

What's Changed

• Add steer-or-queue preference for mid-run follow-ups by @Quack6765 in #397
• Skip search planner call in auto mode by @Quack6765 in #399
• Send OpenAI Responses tools with strict: false by @Quack6765 in #403
• Store each assistant message's own token estimate by @Quack6765 in #400
• Keep provider connections alive for 60 seconds by @Quack6765 in #398
• Reuse the Exa/Tavily MCP connection across searches by @Quack6765 in #404
• Run image instruction preparation at low reasoning effort by @Quack6765 in #401
• Poll conversation only while the live connection is down by @Quack6765 in #402
• Send Anthropic effort under output_config by @Quack6765 in #405
• feat: run neighbouring read-only tool calls in parallel batches by @Quack6765 in #406
• perf(turn): faster turn setup and wrap-up by @Quack6765 in #408
• Keep the prompt prefix stable so provider prompt caches survive a turn by @Quack6765 in #407
• Add a secret vault the agent can use without seeing values by @Quack6765 in #409
• Return agent-browser screenshots to the model in the same turn by @Quack6765 in #410
• Add run_python tool and save long shell output to files by @Quack6765 in #411
• Style Take control button as primary accent action by @Quack6765 in #412
• Preserve streamed replies when turn restart races its response by @Quack6765 in #413
• Render one status line per tool run between messages by @Quack6765 in #414
• Fix context usage reporting for cleared, stopped, and failed turns by @Quack6765 in #415
• Add v5.1.0 release notes by @Quack6765 in #416
• Restyle take-over and return-control buttons as accent primary by @Quack6765 in #417
• v5.1.0 Release by @Quack6765 in #418

Full Changelog: v5.0.0...v5.1.0