Releases: mlhher/late-cli
Releases · mlhher/late-cli
Release list
v1.5.0
Feature:
- Add ephemeral researcher subagents for isolated codebase analysis and research
- Add per-agent model selection with
/model, including support for separate orchestrator, coder, and researcher models - Dynamically load subagent definitions while preserving existing shared subagent model configuration
- Add
--promptto start the agent immediately with an initial instruction - Show command descriptions in slash-command autocomplete
Fix:
- Use an embedded
cl100k_basetokenizer for more accurate context and token estimates by @giveen in #101 - Harden multi-line paste handling against binary input and placeholder collisions by @giveen in #101
- Improve
search_toolregex-to-glob conversion and.gitignorepath handling by @giveen in #101 - Improve
.gitignoretest setup for symlinked working directories by @jpia in #108 - Improve streaming history rendering, scrolling, and repaint performance
- Preserve runtime configuration when continuing after a cancelled agent run
- Return clearer subagent failures to the orchestrator and improve coder recovery instructions
- Use valid, collision-safe MCP tool names while retaining bare-name configuration compatibility
v1.4.2
Feat:
- Rewind inside chats with
/rewindcommand to backtrack to earlier messages - Support for drafting long or complex instructions in default external
$EDITORwith/composecommand - Collapse multi-line pastes into inline placeholders to keep history viewable
Fix:
- Resolve MCP spec compliance issues: replace manual stdio transport, reuse client instances, support dynamic tool re-discovery, handle more content types, add SSE/streamable HTTP support, and fix race conditions by @giveen in #98
- Restrict up/down history navigation to when the input bar is empty to prevent losing drafted edits
v1.4.1
Feature:
- Add
search_namesfast-path mode tosearch_toolfor instant filename globbing (bypassing content scanning) by @giveen
Fix:
- Auto-coerce regex patterns to valid globs for
search_namesmode to improve search reliability with smaller models by @giveen - Distinguish output truncation from context exhaustion
v1.4.0
Feature:
- TUI improvements: input history, slash commands, context bar, commit log, and more by @giveen in #90
- Add
.gitignoresupport tosearch_toolfor local LLM efficiency by @giveen in #88 - Add skill reference discovery by @schelmo in #85
- Add
search_toolas a native alternative to bash grep/find by @giveen in #83 - Add continue flag
Fix:
v1.3.1
Fix:
- Let the allowlist override soft signals (e.g. variable expansions, subshells) to reduce prompt fatigue, while keeping a mandatory prompt override for deletion commands (such as
rmorRemove-Item) to prevent accidental data loss
Chore:
v1.3.0
v1.2.7
Fix:
- Namespace
allowed_tools.jsonentries as{server}:{tool}to prevent name collision and incorrect permission reuse across servers by @HenrikSondergaard in #78 - Raise MCP truncation limit to 32,768 characters and fix UTF-8 slicing to prevent splitting multi-byte characters by @HenrikSondergaard in #79
v1.2.6
Feat:
- Sqz tool integration by @zacwolfe in #65
- Force-enable image support for unsupported servers via
--enable-images - Add universal installation script
Fix:
- Pass server.Env to NewStdioTransport in ConnectFromConfig by @HenrikSondergaard in #73
- Don't double-append /v1 when BaseURL already contains a path prefix by @HenrikSondergaard in #74
New Contributors
- @HenrikSondergaard made their first contribution in #74
- @zacwolfe made their first contribution in #65
v1.2.5
Fix:
- Queued messages: inject queued message on the next possible turn instead of waiting for the agent to return control
- Chat keybindings: disable component keybindings causing UI shifting during typing
- YAML frontmatter: introduce 1mb yaml frontmatter limit for skill parsing