Bantam is a minimalist, dependency-free AI agent specification with reference implementations in **Go** (`main.go` + `term_*.go`, module `code.luxferre.top/luxferre/bantam`) and **Perl 5** as **MicroBantam** (`mb`, under 100 SLOC). It provides an agentic loop capable of autonomous tool execution, direct shell interaction, real-time response streaming, markdown terminal rendering with box-drawing tables, Fibonacci backoff network resilience, context window auto-discovery, token usage tracking with prompt cache breakdowns, prefix-cache-friendly conversation compaction, and subagent delegation using any OpenAI-compatible completions API.
The entire philosophy of Bantam is built upon two principles:
1. The structure must be as simple as possible for anyone to be able to reimplement the agent from a plain algorithm description.
2. The agent only needs to provide two tools: a tool to call shell commands and a tool to call itself. In theory, this should be sufficient to give LLMs the ability to handle tasks of any complexity.
Because of the second principle, Bantam itself was named after Victorinox Bantam Alox, a small and lightweight Swiss army knife with only two tools.
The Go port is a single `main.go` plus four platform files (`term_linux.go`, `term_bsd.go`, `term_windows.go`, `term_other.go`) for the built-in raw-terminal line editor — zero external dependencies.
In interactive mode, prompts can span multiple lines: press **Ctrl+J** to insert a real line break (the cursor moves to the next line), then **Enter** to submit the whole multi-line prompt. The Go port ships its own raw-mode line editor (arrow keys move the cursor, Up/Down browse history, Backspace edits, Ctrl+C clears line / interrupts in-flight run, Ctrl+D exits), working everywhere without third-party dependencies.
After every interaction, Bantam displays token usage (prompt tokens, cached/uncached breakdown when supported by the provider, completion tokens, and context window utilization):
- `/compact` — compact context down to the system message and a concise summary using the LLM; the compaction prompt is appended directly to the existing message prefix to guarantee a 100% prompt cache hit
When context window usage reaches **60% or higher**, Bantam automatically offers to compact the conversation:
```text
Context usage is at 62.4% (124800 / 200000 tokens). Compact conversation? [Y/n]:
```
Pressing **Ctrl+C** during an active inference run or long-running shell execution cleanly interrupts the turn without appending partial or malformed responses to the conversation context.
The current conversation is also **auto-saved** to `~/.bantam/sessions/autosave.json` after every turn, on `/clear`, `/load`, `/compact`, and on exit — so you can always `/load autosave` to resume where you left off.
Using these rules, everyone can build their own copy of Bantam from scratch in little time in any language that supports file access, shell and HTTP(S) calls.
1. **Initialization**: Read `system.txt` and `model.cfg`. Discover context window size from the `/models` endpoint (or fallback to `context_window` from `model.cfg` or 200,000 tokens). Prepare an array of messages starting with the system prompt `{"role": "system", "content": system_prompt}`.
2. **Input Processing**: Take user prompt (via command-line file parameter or interactive stdin). If prefixed with `!`, execute the command directly via `shell_exec` without appending to conversation context. Otherwise, append `{"role": "user", "content": prompt}`, and invoke `AL(cfg, messages)`.
- Send `messages` and tool definitions to the OpenAI-compatible `/chat/completions` API endpoint with custom `User-Agent` headers and `stream_options: {"include_usage": true}`.
- Support context cancellation (e.g. on `SIGINT` / Ctrl+C) to cleanly abort in-flight requests without appending incomplete messages.
- If `stream=true`, parse SSE data chunks (`data: {...}`) in real-time to stream reasoning content (`reasoning_content`) and response text directly to stdout, bracketing the reasoning block with `--- reasoning start ---` / `--- reasoning end ---` markers, rendering Markdown and tables constrained to terminal width.
- Reconstruct the assistant message and track usage tokens (`prompt_tokens`, `completion_tokens`, cached tokens). If `tool_calls` exist, trace the call (`[tool call: name(args)]`), validate JSON arguments, execute the requested tool (`shell_exec` or `run_subagent`), trace the result (`[tool result: name]`), append the tool response `{"role": "tool", "tool_call_id": id, "content": result}`, and repeat the loop.
- If no tool calls remain or `max_al_iterations` is reached, return the updated messages list and turn usage stats.
4. **Post-Turn Reporting & Compaction**:
- Display token usage and context window percentage.
- If context usage is >= 60%, prompt user to compact.
- Compaction appends `"You are now acting as a compaction engine. Summarize the preceding conversation concisely but completely..."` as a user message to the conversation, invokes the LLM (ensuring zero prompt cache misses), and resets the conversation to the system prompt and the resulting summary.
4. Read the first command-line parameter. If non-empty, read user prompt from the specified file. If prefixed with `!`, execute the shell command directly via `shell_exec` and exit. Otherwise, append to `messages` (`role: "user"`), run `AL(cfg, messages)`, display token usage, and exit.
5. Read user prompt from standard input (with `readline` line editing and history in `~/.bantam_history`; **Ctrl+J** inserts a real newline into the line being edited). If equal to `/quit` or EOF, exit. If equal to `/clear`, reset `messages` to step 3 and return to step 5. If equal to `/save`, write the whole `messages` array to `~/.bantam/sessions/<id>.json` (with an auto-generated summary) and return to step 5. If equal to `/list`, print saved sessions and their summaries and return to step 5. If starting with `/load`, replace `messages` with the saved session's messages (by exact id or unique prefix) and return to step 5. If equal to `/compact`, ask the LLM to summarize the conversation by appending the compaction prompt to preserve KV cache, replace `messages` with `[system, summary-user-message]`, and return to step 5. If starting with `/cfg`, display the current value (`/cfg <param>`) or update `model.cfg` live (`/cfg <param> <val>`) and return to step 5. If starting with `!`, execute the command directly via `shell_exec` without adding the result to `messages` and return to step 5. If equal to `/help`, print the command list and return to step 5. After every user turn and on exit, auto-save `messages` to `~/.bantam/sessions/autosave.json`.
6. Append user prompt to `messages` (`role: "user"`), run `AL(cfg, messages)`, display token usage, check 60% context threshold for auto-compaction, and go to step 5.
- If `stream=true`, parse SSE stream (`data: {...}`) for real-time reasoning and text output, bracketing reasoning with `--- reasoning start ---` / `--- reasoning end ---` markers.
2. Append the assistant's response message object to `messages`. If non-streaming and response has reasoning tokens (`reasoning_content` or `reasoning`), output them wrapped in `--- reasoning start ---` / `--- reasoning end ---` markers.
3. If there are pending `tool_calls` in the assistant response:
- For each tool call, output a trace log (`[tool call: name(args)]`).
- Sanitize tool arguments to filter out non-printable and space-like Unicode characters (protecting against indirect prompt injection), and validate JSON arguments. If invalid, format a tool-error response so the LLM can self-correct.
- Sanitize the tool result output to strip any non-printable and space-like Unicode characters (leaving only ASCII space, tab, newline, and printable Unicode characters).
If the API rejects the request with an `Invalid assistant message: content or tool_calls must be set` error (usually caused by a previously cut-off stream that left an empty assistant message in the session), all implementations strip the last `assistant`-role message from the session and retry the call.
- `timeout` (HTTP timeout in seconds for LLM API calls, default 300; in the Go port it bounds connection setup and time-to-first-byte, so long streaming responses are not cut off mid-stream)
The Go port's built-in editor tracks the cursor with its own column math (terminal auto-wrap aware) and redraws from the first line of the buffer, so wrapped input stays clean at any terminal width.
When enabled, the interactive console uses a subtle ANSI palette: the pending-request status `...requesting...` is darkened bold, reasoning markers are cyan, reasoning text is dim, `[tool call: ...]` traces are yellow, `[tool result: ...]` headers are green, tool result bodies are dim (red for tool errors/unknown tools), and errors/network retries are red. Tool result payloads fed back to the LLM are never colored.
- Action: run `AL(cfg, [{"role": "system", "content": system_prompt + "\n\nImportant: this is a child agent"}, {"role": "user", "content": prompt}])` subject to recursion depth limit (`MAX_DEPTH = 5`) and return the text content of the last `assistant`-role message.
MicroBantam (`mb`) is a compressed Perl 5 reference implementation of the same agent in **under 100 SLOC**, written to stay readable while keeping the full agentic core. It reads the same `model.cfg` and `system.txt` from the current working directory.
- Full agentic loop: LLM calls, `shell_exec` / `run_subagent` tool execution with JSON argument validation (invalid args are fed back so the model can self-correct), and the 5-level subagent recursion depth limit
- Session management: `/save`, `/list`, `/load <id>` (exact id only, no prefix matching), `/cfg <param> [val]`, and auto-save to `~/.bantam/sessions/autosave.json` after every turn and on exit; session ids get `-1`, `-2`, ... suffixes on same-second collisions
The `extras/` directory contains small, dependency-light shell scripts that extend Bantam without changing its core. Because Bantam's only built-in tools are `shell_exec` and `run_subagent`, these helpers can be invoked directly by the agent through `shell_exec` to give it real-world capabilities (web search, live weather) that the base model alone does not have. They are plain `/bin/sh` scripts depending only on `curl` (and `jq` where noted), so the agent can discover and run them just like any other command.
### `extras/websearch`
A simple web search tool for Bantam built on [Exa's MCP server](https://mcp.exa.ai/mcp)
(Streamable HTTP). It speaks the JSON-RPC MCP protocol (`initialize` ->
`notifications/initialized` -> `tools/call` with `web_search_exa`) and prints
the formatted result text. Usage:
```sh
extras/websearch "your query" # default 5 results
extras/websearch "your query" 10 # custom number of results
The default system prompt instructs the agent to respect `AGENTS.md`/`GEMINI.md`/`CLAUDE.md` files.
### Is there any common config place for Bantam?
No, loading `model.cfg` and `system.txt` is deliberately only supported from the current working directory. This allows natural separation of configs and system prompts per project. In case there's no `system.txt` inside the project, the concise and sensible default system prompt will be loaded. In case there's no `model.cfg` inside the project, Bantam will use the free Big Pickle model from OpenCode Zen with the temperature 0.7. Big Pickle has been chosen as the default because it has no set expiration date, unlike other OpenCode's keyless tiers.
On Android, Bantam (Go) and MicroBantam (`mb`) are easily runnable within the Termux environment. The Go implementation is preferred for performance reasons.
On iOS/iPadOS, the easiest way to use Bantam is to run MicroBantam (`mb`) inside iSH. Some terminal features may not be available (run with `rlwrap` to bring them back), but the agent itself is fully functional.
