agent.sh

Repo: github.com/surya-prakash-susarla/agent_sh

Single bash script LLM REPL with tool use (command execution). Dependencies: curl, jq. No daemons, no runtimes, no install — download and run.

Design

All commands require explicit per-invocation approval before execution
Multi-turn conversation: full history sent each turn as JSON lines (.agent_conversation)
Each history entry: {role, content} — roles: system / user / assistant / tool
Built with bashly: bashly.yml + build.sh → dist/agent binary

File	Role
`src/root_command.sh`	Entry point — resolves config, enters REPL
`src/lib/instructions.sh`	System prompt: tool JSON format, web safety rules
`src/lib/repl.sh`	REPL loop — input, approval gating, history, display
`src/lib/network.sh`	POSTs conversation to endpoint, parses tool calls vs. replies
`src/lib/logging.sh`	Debug output, gated by `--debug` flag

get_response() returns typed JSON:

{ "type": "tool", "assistant_message": { ... }, "command": "git log -n 5 --oneline" }

type: tool → print [Tool Request], prompt approval, run command, append tool message to history, loop
type: assistant → print reply, exit tool loop

Without this typed return, the loop would re-run the same command on multi-step flows.

Scripts in tests/, driven by env vars:

export AGENT_ENDPOINT="http://your-llm:11434"
export AGENT_MODEL="your-model"
bash run_tests.sh

test_conversation.sh — chat continuity and memory
test_tool_use.sh — command execution and approval flow
test_simple_denial.sh, test_multi_denial.sh, test_recursive_approval.sh — denial/approval edge cases

Logs: outputs/*.log

dist/agent binary versioned in git. MIT licensed. v1.0.0 available as a downloadable asset.