v0.99.6

What's Changed

MR: #494
Full Changelog: Full Changelog: v0.99.5...v0.99.6

Includes:

• Update conversation structure with randomized calls to action to encourage models to keep going
• Make tree sitter versions explicit
• Add /load-mcp * and /remove-mcp * with mcp tools
• Disable connection keep alive so models providers with implicit caching can actually write to their KV and prompt caches

v0.99.5

What's Changed

MR: #492
Full Changelog: v0.99.4...v0.99.5

Includes:

• CLI-8: Make ls tool OS-agnostic with improved path handling by @szmania in #490
• Cymbal wrapper that works cross-platform and doesn't do weird directory shenanigans

v0.99.4

What's Changed

MR: #489
Full Changelog: v0.99.3...v0.99.4

Includes:

• Update py-cymbal version to fix it changing the root directory
• Updates to conversation order (reminders come before context blocks and current todo list)
• Allow mcp, skills, and tools commands to accept multiple space separated arguments

v0.99.3

What's Changed

MR: #488
Full Changelog: v0.99.2...v0.99.3

Includes:

• Updated token and cost reporting
• Integrate cymbal library with the ExploreSymbols tool
• Remove ViewFiles* tools in favor or the above
• Make terminal output for tools more consistent with custom format_output() methods
• Add automatic tool call corrections for array-like tools when LLMs do not wrap the arguments appropriately themselves
• Incremental compaction summaries with observational memory pattern and background thread summarization

v0.99.2

What's Changed

MR: #487
Full Changelog: v0.99.1...v0.99.2

Addresses:

• #484

Includes:

• CLI-5: Restore user input required notifications by @szmania in #483
• CLI-6: Fix Keyboard Interruption Handling by @szmania in #485
• CLI-7: Add evals.json files to skills for quality assurance and testing by @szmania in #486
• Fix for git repo initialization switching paths inside of workspaces
• Add ability to use custom ignore files with individual workspace projects

v0.99.1

What's Changed

MR: #482
Full Changelog: v0.99.0...v0.99.1

Addresses:

• #480
• #481

Includes:

• Modify Command tool to support sending input to background process stdin stream, a step towards stronger terminal automation
• Less aggressive alternation between search phase and edit phase in agent mode (only switch on edit error)
• Allow ShowContext to search multi-line entries since some models keep attempting to do that

v0.99.0

What's Changed

MR: #479
Full Changelog: v0.98.2...v0.99.0

Addresses:

• #478

Includes:

• Multi-project support with new --workspaces flag and key in config file
• Truncate TUI footer path with home folder tilde

v0.98.2

What's Changed

MR: #477
Full Changelog: v0.98.1...v0.98.2

Includes:

• Fix: Tools Paths Standardization and Tool Registry Improvements by @szmania in #476
• Updates to tool invocation tracking

v0.98.1

What's Changed

MR: #475
Full Changelog: v0.98.0...v0.98.1

Addresses:
#474

Includes:

• Leave temperature out when model doesn't support it (looking at you GPT-5+ series)
• Prefix MCP server tool names to disambiguate between tools from different servers
• Modifications to ShowContext to inform the model when it has complete contextual information already
• Track tool invocations to aggressively prevent a model from calling a tool with the same tool parameters between edits

v0.98.0

What's Changed

MR: #472
Full Changelog: v0.97.5...v0.98.0

Addresses:
#469

Includes:

• Add support for multiple --mcp-servers-files arguments by @szmania in #473
• Add agent model display and configuration improvements by @szmania in #470
• Modify ShowNumberedContext into ShowContext for content based targeting for stable hashline/hashpos based edits
• Add "rules"-like file support to allow dynamically loadable files similar to what AGENTS.md or CLAUDE.md are used for
• Add --rules argument and /rules command to support the above
• Agent mode lints files as it edits them to detect syntax errors and ameliorate them during operation
• Modify similarity detection so it is less intrusive
• Change hashline/hashpos seperator since most LLM tokenizers see double colons as a single token (nice on the eyes, nice for the machine)
• Conversation system decomposition into an instance-able set of base classes that will support multiple context streams for sub-agent development
• Memory usage improvements by using GitPython's GitCmdObjectDB and general caching to search through the underlying git repository for generation of the repository map
• Agent mode prompt updates to make the steps more easily accomplishable to the models and use slightly fewer tokens
• Make available commands easier to search through by matching sub strings after start strings
• Fix quoted file names being incorrectly spliced in the /add-like commands
• Allow sending/injecting messages while a generation task is running in the TUI to increase general real-time steerability