comet-code 0.1.4

Agentic CLI coding assistant

CometCode

An agentic coding assistant for the terminal. Give it a task — it reads your codebase, reasons about what to change, proposes edits, and applies them only after you approve.

Built from scratch on LangGraph and OpenRouter, with a streaming Rich terminal UI.

---

Demo

mode → implement
╭────────────────────────────────────────────────────────────╮
│ add a /clear command that resets conversation history       │
╰────────────────────────────────────────────────────────────╯

  tool
    ● Read src/cli/commands.py done
    └ def handle_command(text, console, state, orchestrator):
    ● Edit src/cli/commands.py done
    └ [ok] replaced 1 occurrence(s)

╭──────────────────────────── response ──────────────────────╮
│ Added `/clear` — it calls `orchestrator.reset_history()`   │
│ and clears `state.last_tool_history`.                       │
╰────────────────────────────────────────────────────────────╯

↓ ~3.2k tokens  Cooked for 8s

---

Architecture

CometCode is structured as a LangGraph state machine where each node receives the full AgentState and returns state updates. The graph has two nodes — call_llm and execute_tools — connected by conditional edges that route based on the model's response type.

START → call_llm ──→ execute_tools ──┐
            ↑                        │
            └────────────────────────┘
            (loops until final answer or budget exhausted)

Key design decisions:

• AgentState as the single source of truth — conversation history, token budgets, tool-call counters, evidence notes, and attempt metadata all live in one typed dict that flows through the graph. Nodes are pure: they receive state, return a patch.

• Dual invocation paths — models that support native function calling stream through _invoke_native; models that don't fall back to _invoke_json_fallback with a structured Pydantic output schema. The rest of the system is unaware of the difference.

• Explicit budget enforcement — the graph tracks tool_calls_used, consecutive_no_signal, and repeat_call_streak. Hitting any limit triggers a soft-stop: the model is forced into a final answer using whatever evidence it collected, rather than erroring.

• Human-in-the-loop approval — mutating tools (write_file, replace_text) are gated behind a request_approval callback. After approval and execution, the loop returns to the LLM so it sees the updated file state before proposing the next edit — preventing duplicate proposals.

• Mode-scoped prompting — explain, debug, refactor, implement, and plan modes each carry different system instructions and tool permissions. Read-only modes (explain, plan) never receive mutating tools in their schema.

---

Features

• Five task modes — explain, debug, refactor, implement, plan
• Multi-model support — route to any OpenRouter model; native tool calling used when available, JSON schema fallback otherwise
• Streaming terminal UI — live spinner with real-time token count and elapsed timer; tool history with colored unified diffs for proposed changes
• Persistent conversation history — context carries across turns within a session; mode switches are surfaced to the model explicitly
• Slash command interface — /mode, /model, /tools, /clear, /help with tab-completion
• Self-limiting agent — budget caps on tool calls, no-signal streaks, and repeated identical calls; graceful degradation to best-effort answers

---

Stack

Layer	Technology
Agent orchestration	LangGraph
LLM routing	OpenRouter via LangChain ChatOpenAI
Terminal UI	Rich + prompt_toolkit
Schema validation	Pydantic v2
Python	3.12+

---

Getting Started

pip install comet-code
# The first run will prompt for your OpenRouter API key.
# You can also set it later via `/key set <key>` or the `OPENROUTER_API_KEY` environment variable.
comet

Slash commands

Command	Description
/mode <name>	Switch task mode (explain, debug, refactor, implement, plan)
/model <name>	Switch model (by alias, slug, or label)
/tools	Show last run tool history
/clear	Reset conversation history
/help	List all commands

---

Project Layout

src/
  cli/          # Terminal UI, input handling, rendering, slash commands
  core/         # LangGraph graph, nodes, orchestrator, state schema
  llm/          # OpenRouter client, model catalog, prompt builder
  schemas/      # Pydantic models (events, tasks, state, tools)
  tools/        # Tool implementations (read_file, search_text, replace_text, …)

---

License

MIT