Features¶

Interfaces¶

CLI — Rich-rendered REPL with streaming, history, command queue.
Textual TUI — fullscreen terminal UI with input bar, scrollable output, collapsible tool cards, modals, footer status.
Web UI — browser-served chat with tool cards and live streaming.
Daemon mode — long-running server with Unix-domain-socket transport so multiple clients share a session.

Local Ollama and any OpenAI-compatible endpoint.
Auto model selector — pick a tier (cheap / standard / heavy) from the prompt and active context.
Auto cost router with per-session token + dollar tracking and budgets.
Vision-capable model gating (warns and drops images on text-only models instead of 400-ing).
Mid-turn model escalation when the loop guard detects the small model is stuck.

Iterative, bounded tool-calling loop (XML and native tool-call formats both supported).
Auto-extending iteration cap — only warns on a real hard cap.
Loop guard — observes assistant text, tool calls, and file writes; stops on detected loops, repeat-write storms, or no-progress patterns.
Pause (Esc) keeps the conversation consistent: interrupt flag stops the loop at the next iteration boundary instead of mid-call.
Queued user submissions fold into the running turn at iteration boundaries — clarifications typed mid-task arrive on the next iteration as one merged user message.
Reflection memory on tool errors; deferred LSP diagnostics flushed per-iteration after writes.

File ops — read, write, multi-write, edit, apply diff, batch regex edit.
Search — grep, glob, symbol search, find-callers, list-dir, language-server queries (definition / references / hover / diagnostics), repomap (PageRank-ranked symbol map).
Refactoring — project-wide rename, extract-function, inline.
Shell — rlimits / sandbox backends, syntax-highlighted output.
Test runner — structured failure parsing.
Git — status / diff / log / blame / commit / pickaxe search.
Web — fetch, search, headless browser (text + screenshot).
Notebook — edit / run Jupyter cells.
Code analysis — dependency graph, dead-code, complexity, duplication, architecture map, type coverage, migration plan, diff impact, quality score.
Memory & doctrine — persistent project facts plus a JAAICODE.md doctrine file the agent maintains in cooperation with the user.
Sub-agents — general / explore / planner / test-runner / reviewer spawned as autonomous helpers.
Cron — schedule prompts to run on a cadence.

Named sessions with replay, transcript export, undo of file edits.
Skill library — auto-activated short instructions that match a prompt.
Plugin system for extending the tool surface.
Slash commands: /help, /queue, /model, /sessions, /review, /diff, /skills, /store, /workplan, /replay, /mcp, …

Pre-commit hook enforcing CHANGELOG updates for user-visible files via per-repo .changelog-watch patterns.
Confirmation modals for destructive ops; YOLO mode bypasses with a clear footer indicator.
Token budget guard, hard-cap iteration guard, loop guard — three independent kill switches.

chat (default) and code (full tool access).
plan mode — read-only tools only, designed for exploring a codebase and writing a plan before editing.
Mode personas: architect, ask, debug, review, etc.