Home
covonaut v1.1.4 Released: Rolling Parallel Pool Speedup, Full Multi-Modal Reading Enabled, Deleted Files Can Also Be Recovered in the Recycle Bin

Covonaut, a production-grade Go agent framework, evolves into a dependable development partner. The latest v1.1.4 release introduces a rolling execution pool, multi-modal content handling, and robust security enhancements. Built entirely in Go under the MIT license, it covers agent loops, tool invocation, graph orchestration, and multi-protocol support. Developers can upgrade to v1.1.4 without modifying existing interface code, enabling immediate access to new features.
Scheduling efficiency drives this update. Previously, parallel execution acted like an uncoordinated workforce, launching all calls simultaneously and queuing based on concurrency limits. A slow tool could occupy a slot indefinitely, blocking subsequent tasks. v1.1.4 implements a single rolling pool scheduler: calls initiate strictly in the model-provided order, and empty slots are immediately filled. Slow calls no longer delay parallel tasks; for non-parallelizable operations, the system clears running tasks before executing them serially, creating a serial barrier. The model observes an execution order identical to the serial sequence. Each call independently captures panics, preventing a single tool failure from crashing other batch calls. Scheduling strictness depends on the tool's concurrency safety declaration and mode: parallel mode isolates only explicitly non-concurrency-safe tools, while the new hybrid mode executes all tasks without explicit concurrency declarations serially. Developers can select the mode based on their tool ecosystem's reliability. Additionally, a multi-modal content channel allows rich outputs, such as screenshots, to be transmitted alongside text to models supporting multi-part responses, with automatic fallback to plain text for unsupported services.
The reading tool now offers versatile processing capabilities. It automatically routes content to five distinct paths—text, images, PDFs, Jupyter Notebooks, and videos—based on file extensions and characteristics. Both local paths and network links are supported. Text and Notebooks support pagination by line or cell, with Notebooks rendered into readable text. Images allow up to eight targets for comparison, including optional questions for focused descriptions. PDFs support three methods: page range, maximum page count, and metadata extraction. Videos allow specification of extracted frames and time windows. Rich media processing logic is provided via plug-in callbacks, offering clear prompts rather than silent failures when not integrated. Truncated text includes a continuation prompt, enabling the model to resume reading without content loss. Real binary files are returned as format recognition results instead of being forced into the context.
Operational safety is significantly enhanced. Deletion operations now utilize a system recycle bin. The delete tool and shell's delete command move files to the platform-specific recycle bin rather than permanently destroying them: macOS uses the system recycle bin, falling back to the user's trash on error; Linux prioritizes the freedesktop recycle bin interface, falling back to a manual recycle directory; Windows moves files to the system recycle bin. Duplicate names automatically receive a timestamp suffix to prevent overwriting. Built-in safety barriers strictly reject deletion of the root directory, current working directory, and user's home directory, ensuring no permanent deletion occurs. On the command side, shell syntax tree parsing intercepts only top-level, single, path-literal delete commands. Composite commands, paths with variable concatenation, and statements with redirection proceed to the real shell. Irreversible commands, such as emptying or shredding, are blocked, prompting users to use recoverable methods instead.
Isolation and enhancements are comprehensive. Shell sub-processes and chat interfaces are completely separated, avoiding inheritance of control terminals, mouse focus, or job control. Running interactive programs like vim or less in the interface no longer disrupts rendering. The framework now validates structured output against JSON Schema, reusable for model JSON pattern returns and tool outputs. The passwordless backend for network search adapts automatically based on the operating region, prioritizing Bing in Chinese environments and DuckDuckGo elsewhere, with regions inferred from environment variables, system language, or time zones. Additionally, a new explicitly enabled persistent terminal tool family retains the working directory, exported variables, and background jobs across multiple calls using a pseudo-terminal, differing from creating a new process each time. By default, it limits concurrent sessions and maintains a scrolling output buffer for each session.
Upgrading requires only updating the module version to v1.1.4. There are no breaking interface changes. The only consideration is that the scheduling semantics of parallel mode have shifted from full concurrency to a rolling pool with a relaxed admission policy. To maintain conservative scheduling, switch to the newly added hybrid mode and ensure tools have completed their concurrency safety declarations.
Related article
Tencent Pledges $20M to Back AI Prodigy Lin Junyang’s New Lab
Significant activity continues to surge within the AI venture capital sector. Recent reports confirm that Lin Junyang, former head of Alibaba’s Qwen large language model team, has secured initial funding of $20 million from Tencent for his newly laun
CapCut Unveils Professional Edition and AI Assistant Xiaoying in Mobile App
CapCut recently hosted its "AI New Creation Launch Event," unveiling a suite of innovative AI tools, including CapCut Hub and the AI creation assistant "Xiaoying." This move marks a significant upgrade to its creator ecosystem, establishing a compreh
NVIDIA GTC Sparks AIPC Boom as Leishen Technology Rallies 30% to Lead Edge Computing Sector
Global attention in the artificial intelligence sector is pivoting back to edge hardware. Bolstered by NVIDIA’s upcoming GTC conference and strong endorsements from major tech firms, AIPC-related stocks surged on the morning of June 1st. Leishen Tech
Related Special Topic Recommendations
Comments (0)
0/500

Covonaut, a production-grade Go agent framework, evolves into a dependable development partner. The latest v1.1.4 release introduces a rolling execution pool, multi-modal content handling, and robust security enhancements. Built entirely in Go under the MIT license, it covers agent loops, tool invocation, graph orchestration, and multi-protocol support. Developers can upgrade to v1.1.4 without modifying existing interface code, enabling immediate access to new features.
Scheduling efficiency drives this update. Previously, parallel execution acted like an uncoordinated workforce, launching all calls simultaneously and queuing based on concurrency limits. A slow tool could occupy a slot indefinitely, blocking subsequent tasks. v1.1.4 implements a single rolling pool scheduler: calls initiate strictly in the model-provided order, and empty slots are immediately filled. Slow calls no longer delay parallel tasks; for non-parallelizable operations, the system clears running tasks before executing them serially, creating a serial barrier. The model observes an execution order identical to the serial sequence. Each call independently captures panics, preventing a single tool failure from crashing other batch calls. Scheduling strictness depends on the tool's concurrency safety declaration and mode: parallel mode isolates only explicitly non-concurrency-safe tools, while the new hybrid mode executes all tasks without explicit concurrency declarations serially. Developers can select the mode based on their tool ecosystem's reliability. Additionally, a multi-modal content channel allows rich outputs, such as screenshots, to be transmitted alongside text to models supporting multi-part responses, with automatic fallback to plain text for unsupported services.
The reading tool now offers versatile processing capabilities. It automatically routes content to five distinct paths—text, images, PDFs, Jupyter Notebooks, and videos—based on file extensions and characteristics. Both local paths and network links are supported. Text and Notebooks support pagination by line or cell, with Notebooks rendered into readable text. Images allow up to eight targets for comparison, including optional questions for focused descriptions. PDFs support three methods: page range, maximum page count, and metadata extraction. Videos allow specification of extracted frames and time windows. Rich media processing logic is provided via plug-in callbacks, offering clear prompts rather than silent failures when not integrated. Truncated text includes a continuation prompt, enabling the model to resume reading without content loss. Real binary files are returned as format recognition results instead of being forced into the context.
Operational safety is significantly enhanced. Deletion operations now utilize a system recycle bin. The delete tool and shell's delete command move files to the platform-specific recycle bin rather than permanently destroying them: macOS uses the system recycle bin, falling back to the user's trash on error; Linux prioritizes the freedesktop recycle bin interface, falling back to a manual recycle directory; Windows moves files to the system recycle bin. Duplicate names automatically receive a timestamp suffix to prevent overwriting. Built-in safety barriers strictly reject deletion of the root directory, current working directory, and user's home directory, ensuring no permanent deletion occurs. On the command side, shell syntax tree parsing intercepts only top-level, single, path-literal delete commands. Composite commands, paths with variable concatenation, and statements with redirection proceed to the real shell. Irreversible commands, such as emptying or shredding, are blocked, prompting users to use recoverable methods instead.
Isolation and enhancements are comprehensive. Shell sub-processes and chat interfaces are completely separated, avoiding inheritance of control terminals, mouse focus, or job control. Running interactive programs like vim or less in the interface no longer disrupts rendering. The framework now validates structured output against JSON Schema, reusable for model JSON pattern returns and tool outputs. The passwordless backend for network search adapts automatically based on the operating region, prioritizing Bing in Chinese environments and DuckDuckGo elsewhere, with regions inferred from environment variables, system language, or time zones. Additionally, a new explicitly enabled persistent terminal tool family retains the working directory, exported variables, and background jobs across multiple calls using a pseudo-terminal, differing from creating a new process each time. By default, it limits concurrent sessions and maintains a scrolling output buffer for each session.
Upgrading requires only updating the module version to v1.1.4. There are no breaking interface changes. The only consideration is that the scheduling semantics of parallel mode have shifted from full concurrency to a rolling pool with a relaxed admission policy. To maintain conservative scheduling, switch to the newly added hybrid mode and ensure tools have completed their concurrency safety declarations.
Tencent Pledges $20M to Back AI Prodigy Lin Junyang’s New Lab
Significant activity continues to surge within the AI venture capital sector. Recent reports confirm that Lin Junyang, former head of Alibaba’s Qwen large language model team, has secured initial funding of $20 million from Tencent for his newly laun
NVIDIA GTC Sparks AIPC Boom as Leishen Technology Rallies 30% to Lead Edge Computing Sector
Global attention in the artificial intelligence sector is pivoting back to edge hardware. Bolstered by NVIDIA’s upcoming GTC conference and strong endorsements from major tech firms, AIPC-related stocks surged on the morning of June 1st. Leishen Tech











