v0.100.10

What's Changed

MR: #589
Full Changelog: v0.100.9...v0.100.10

Addresses:

• #584
• #587

Includes:

• Updated Command tool implementation so models can use stdin for interactive terminal commands
• Remove instances of mocking LLM assistant messages about files since it does more harm than good to their reasoning
• Cache hit ratio updates, namely debouncing ephemeral messages so pre-populated LLM prefix caches remain available
• Represent files as JSON-encoded in agent mode since the expected outputs are JSON
• Model override configuration updates LiteLLMs internal cache which allows specifying models de-novo entirely in the main config file

v0.100.9

What's Changed

MR: #582
Full Changelog: v0.100.8...v0.100.9

Addresses:

• #537
• #581
• #583

Includes:

• Miscellaneous updates to HashPos content addressing system
• Defer tree-sitter loading to help start up speed/size
• ContextManager add on missing path upgrades to create by @JessicaMulein in #579
• Allow agents MCP management + HashPos Base64 + TUI reliability improvements by @szmania in #585

v0.100.8

What's Changed

MR: #578
Full Changelog: v0.100.7...v0.100.8

Addresses:

• #554
• #576

Includes:

• Add /hot-reload command that can live reload configuration and preserve current conversation
• Fix adjacent replace operations merging onto one line
• Small TUI updates, ctrl+c when text is selected, copies it, and when not selected cancels the current prompt
• Allow agents and subagents to have distinct sets of MCP servers active
• Remove insert from EditText tool since it can be accomplished with replace and it confuses models which to use when
• ReadRange tool to count tokens in snippet not lines for decision logic on what to show LLM
• Let LiteLLM manage github copilot provider headers

v0.100.7

What's Changed

MR: #575
Full Changelog: v0.100.6...v0.100.7

Includes:

• Automatic tree-sitter based edit remediation for LLMs being slightly off in range specification
• Fix SwitchCoderSignal propagation so sub-agent parallel sessions can do things like changing the active model
• Fix output routing from sub agents to their dedicated panes instead of to the primary agent

v0.100.6

What's Changed

MR: #563
Full Changelog: v0.100.5...v0.100.6

Addresses:

• #564
• #569

Includes:

• Update tool validation to handle ReadRange params
• Simplify Grep tool to default to case insensitive regex searches
• Add allowed_commands under agent-config to whitelist some commands regardless of confirmation settings
• Make file messages more precise for the LLM
• Debounce reminder messages to just once every 5 turns to improve cache efficiency
• Updates to system prompts to include instructions included in debounced reminders
• Terminal aesthetic updates (updated announcements, thinking block display, TUI tool result printing)
• Fewer emojis, more unicode
• Many updates to tool and tool parameter descriptions and fall back behaviors to help keep LLMs on track
• Add retry logic for empty LLM responses (CLI-49) by @szmania in #565
• Transfer abs_fnames back after /agent temp-coder turn by @JessicaMulein in #568
• Sanitize ExploreCode queries for Cymbal FTS5 safety by @JessicaMulein in #570
• Initialize reflected_message in init to prevent AttributeError by @JessicaMulein in #571

v0.100.5

What's Changed

MR: #557
Full Changelog: v0.100.4...v0.100.5

Addresses:

• #555

Includes:

• Update model configuration to include api, llm, and agent configurations to granularly configure models directly in .cecli.conf.yml or in the file specified at --model-overrides-file
• Tool call validation pipeline that can address many JSON formatting mistakes smaller LLMs make from time to time
• Fuzzy-matching model names on model commands
• Local path: projects and repo-local .cecli.workspaces.yml by @JessicaMulein in #547
• Fix gitignore pathspec checks to resolve paths against repo root by @JessicaMulein in #556
• Steer ReadRange away from empty-file read loops by @JessicaMulein in #560

v0.100.4

What's Changed

MR: #551
Full Changelog: v0.100.3...v0.100.4

Includes:

• Fix for confirmations in TUI causing unpredictable hanging
• Add fuzzy matching for file paths for /add-like commands (and also more optimal storage of files as a prefix trie)
• Allow ReadRange to accept line numbers
• Avail --max-reflections parameter for all program modes (was previously locked to 3 for agent mode)
• Change show_completion_notification default to False and remove redundant settings by @szmania in #552

v0.100.3

What's Changed

MR: #545
Full Changelog: v0.100.2...v0.100.3

Addresses:

• #543
• #544

Includes:

• Updates to agent prompts to ideally encourage LLMs to be a bit less flippant
• Resolve AttributeError in TUI.show_error by @szmania in #546

v0.100.2

What's Changed

MR: #541
Full Changelog: v0.100.1...v0.100.2

Includes:

• Make Delegate tool non-blocking
• Remove /invoke-agent in favor of /spawn-agent being able to accept a prompt e.g. /spawn-agent <type> <...prompt text>
• Calling /spawn-agent switches to that agent in the TUI
• Rename Finished tool to Yield tool so it is intuitive for both finishing responses and waiting for sub agents
• Allow automatic reaping of agents spawned by Delegate by default unless auto_reap: false is specified in sub agent configuration
• Allow passing a sub agent identifier to /reap-agent so you don't have to be on the agent in the TUI to reap it
• Make agent skills loadable by sub agents
• Updates to tool call parsing inspired by PR #536 from @JessicaMulein
• Optional AES-256-GCM encryption for saved sessions by @JessicaMulein in #533
• Fix guard missing verbose on headless args by @JessicaMulein in #535
• Ollama-friendly empty LLM tool warning by @JessicaMulein in #538
• Add UUID disambiguation for duplicate agent names by @szmania in #539
• Catch BaseException in worker thread to prevent tracebacks on interruption by @szmania in #540

v0.100.1

What's Changed

MR: #531
Full Changelog: v0.100.0...v0.100.1

Includes:

• Add !! prefix to run commands with output into context suppressed and !!! to background the tui for commands that might need input like git rebasing or sudo passwords, also without putting text into model context
• Improve /add ignore feedback and Ollama keep_alive defaults by @JessicaMulein in #530
• fix(tui): Handle KeyboardInterrupt during worker shutdown (CLI-38) by @szmania in #528

New Contributors

• @JessicaMulein made their first contribution in #530