Skip to content

CHANGELOG.md

Latest commit

 

History

History
162 lines (129 loc) · 7.47 KB

CHANGELOG.md

File metadata and controls

162 lines (129 loc) · 7.47 KB

Changelog

All notable changes to Local CLI Agent are documented here. Format follows Keep a Changelog.

[2.4.0] — 2026-04-05

Added

  • Master-Orchestrator (orchestrator.py) — for complex tasks, a master LLM call produces a specialist plan (up to 10 steps, each assigned to the best profile). User can confirm, edit or abort before execution. Context summaries pass between steps. Available as /orchestrate <goal> and local-cli --orchestrate "goal".
  • Auto-suggest — when a prompt is detected as complex (heuristic score ≥ 0.55 based on word count + keywords), the system offers the orchestrator before sending to the agent. User can always decline with n.
  • Verification profile (verifikation) — automatically inserted as penultimate step in every orchestration that contains code steps. Runs tests, checks imports, checks syntax, outputs structured ✅/❌ report, fixes errors immediately.
  • Interactive plan editor — before orchestration starts, [e] opens an editor where the user can add, remove, or modify steps and specialist assignments.
  • Retry/skip/abort on step failure — if a step throws an exception, the user is asked [r]etry / [s]kip / [a]bort instead of crashing silently.

Changed

  • Profiles now include verifikation (15th profile total)
  • Profile selector updated with new entry

[2.3.0] — 2026-04-05

Added

  • Agent Profiles (profiles.py) — 14 built-in personas activated with /profile or /profile <id>. Each profile appends a focused system-prompt extension. Active profile is shown in the startup status line and injected into every prompt.

    ID Profil Zweck
    standard 🤖 Standard Ausgewogener Allzweck-Assistent
    vibe ⚡ Vibe-Coder Einfach machen, keine langen Erklärungen
    aufraumen 🏗️ Aufräumen & Strukturieren Monolith → saubere Module
    reviewer 🔍 Code-Reviewer Feedback geben, keinen Code schreiben
    debugger 🐛 Fehlersuche Bugs methodisch finden und beheben
    erklarer 📖 Erklärer Alles in einfacher Sprache ohne Jargon
    frontend 🎨 Frontend HTML, CSS, JS — schön und bedienbar
    backend ⚙️ Backend APIs, Datenbanken, Server
    tester 🧪 Tester Sinnvolle Tests schreiben
    security 🛡️ Sicherheit Schwachstellen methodisch prüfen
    docs 📝 Dokumentation README, Kommentare, API-Docs
    performance 🚀 Performance Engpässe finden und beheben
    architect 🏛️ Architekt Erst planen, dann bauen
    devops 🐳 DevOps Docker, CI/CD, Deployment

[2.2.0] — 2026-04-05

Added

  • Mission Mode (mission.py) — give the agent a high-level goal; it creates a numbered plan (up to 8 steps), asks for confirmation, executes each step via a separate agent_loop call, pauses between steps, and prints a completion summary. Available as /mission <goal> and local-cli --mission "goal".
  • Auto-Test-Loop (autotest.py) — after every write_file / edit_file, the configured test command runs automatically. Pass/fail output is appended to the tool result so the agent self-corrects without any user input. Available as /autotest <cmd> / /autotest off / /autotest status.

[2.1.0] — 2026-04-05

Added

  • Undo system (undo.py) — automatic file snapshots before every write_file / edit_file; /undo [n] restores the last n changes, /checkpoint [name] sets a manual restore point. Up to 30 checkpoints kept per session.
  • Project memory (project.py) — detects Python, Node.js, Rust, Go, Java, .NET, Flutter, PHP, Ruby projects automatically; reads metadata from pyproject.toml, package.json, Cargo.toml, go.mod; extracts git branch, last commit, README summary, test framework; injects a --- PROJECT CONTEXT --- block into every system prompt. Results cached via mtime fingerprint.
  • Watch mode (watcher.py) — polls a path every 2 s using os.stat() (no extra deps); triggers the agent automatically on any file change; /watch <path> <instruction> slash command and local-cli --watch <path> <instruction> CLI flag; /watch stop / /watch status.
  • diff tool — shows a unified diff of proposed changes before applying them
  • git tool — safe git operations (status, diff, log, add, commit, branch, stash); permanently blocks push --force, reset --hard, clean -f, branch -D
  • open tool — opens a file or URL in the OS default application (cross-platform)

Changed

  • Version bumped to 2.1.0 in constants.py, pyproject.toml, README badge

[2.0.0] — 2026-04-05

Breaking Changes

  • Backend replaced: llama-cpp-python removed. Now requires Ollama or LM Studio.
  • Python requirement raised: 3.8 → 3.10 (needed for str | None type syntax).
  • Model config format changed: file path → "ollama:model-id" / "lmstudio:model-id".

Added

  • backends.py — full Ollama and LM Studio HTTP client with SSE streaming
  • Animated spinner (⠋⠙⠹⠸⠼⠴⠦⠧⠇⠏ Generiere...) while waiting for model response
  • <|think|> / <|/think|> block filtering for reasoning models (QwQ, DeepSeek-R1, etc.)
  • Markdown tool extraction fallback — models that output code blocks instead of JSON tool-calls are handled automatically
  • Model compatibility check — warns if a loaded model is unsuitable for agent tasks
  • Model auto-select — when exactly one model is loaded (e.g. LM Studio), no dialog is shown
  • /save command — exports current conversation as a timestamped Markdown file
  • /model command — switch model mid-session without restarting
  • /compact command — summarizes the conversation to reduce token usage
  • self_improve: mandatory backup before any edit; syntax validation before any .py write
  • self_improve: file parameter to target any .py file in the package
  • /reload: validates syntax of all package files before restarting

Changed

  • Timeout split into connect (10 s) + stream (300 s) to prevent false disconnects on slow models
  • Automatic retry on ConnectionError (1 attempt) with user-visible message
  • System prompt: German language now enforced as first line
  • Banner replaced with ASCII-safe design (no Unicode box characters)
  • /tokens without argument now shows current value and usage hint
  • install.bat: Python 3.10 check, pip install --upgrade, Ollama model check, LM Studio start instructions, dynamic summary
  • uninstall.bat: checks if package is installed first, clarifies data location, documents what is NOT removed

Removed

  • /thinking command (was a no-op — parameter never reached the backend)
  • llama-cpp-python dependency
  • local_backend.py (replaced by backends.py)
  • NVIDIA NIM / Kimi K2.5 references

Fixed

  • Banner alignment broken by ANSI codes inside padded f-strings
  • Empty model list crash when no backend is running
  • self_improve backup failure was silently ignored

[1.3.0] — 2026-03-19

Changed

  • Complete rebranding from kimi-k-cli to local-cli-agent
  • Added slash command autocomplete via prompt_toolkit
  • Migrated from NVIDIA NIM API to local model backends

[1.2.0] — 2026-03-18

Removed

  • Plugin system (too fragile, low benefit)

Fixed

  • Startup errors on fresh install

[1.1.0] — 2026-03-17

Added

  • Initial release as Kimi K2.5 CLI agent
  • Tool use: bash, file ops, web search, memory, self-improve
  • Interactive and one-shot modes