@@ -44,3 +44,31 @@ We considered having the agent write to `~/.agents/AGENTS.md` for cross-session
4444 A dedicated `~/.config/mini-coder/memories.md` that's clearly agent-owned and separate from user config would be the right path — not overloading existing community standards.
4545
4646 **Summary:** For a coding agent that operates on a repo, the repo _is_ the memory. Users who want cross-session context write it in `AGENTS.md` themselves — that's an intentional act, not an LLM side effect.
47+
48+## Why no tool-call permissions?
49+
50+**Decision:** No approval prompts, no blacklists, no whitelists. Every tool call executes immediately.
51+
52+Our inspirations (Claude Code, OpenCode) require user approval for tool calls — shell commands, file writes, etc. We intentionally skip this.
53+
54+### Permission systems provide a false sense of security
55+
56+- **Shell bypasses everything.** An LLM with shell access can `curl`, `eval`, pipe through `bash`, encode payloads, or chain commands in ways no static blacklist can anticipate. Any permission scheme that allows shell but blocks specific patterns is playing whack-a-mole.
57+- **Blacklists and whitelists always have gaps.** Block `rm -rf /`? The model uses `find -delete`. Block `git push --force`? It uses `git push origin +main`. The surface area is unbounded.
58+- **Approval fatigue degrades security.** After the 20th "Allow shell command?" prompt, users auto-approve everything. The permission system trains the user to click "yes" reflexively — the opposite of its intent.
59+
60+### Permissions are cumbersome
61+
62+A coding agent runs dozens of shell commands per task. Requiring approval for each one destroys the flow that makes a CLI agent useful. The whole point of mini-coder is: small, fast, stays out of the way.
63+
64+### Isolation is a separate concern
65+
66+Sandboxing is a real need, but it belongs at the OS/container level — not inside the agent. Tools like [nono](https://nono.sh/) provide proper filesystem and network isolation that the LLM cannot circumvent. This is defense in depth done right: the agent runs unrestricted inside a sandbox that enforces actual boundaries.
67+
68+### Our approach
69+
70+- The system prompt includes safety rules (no secrets, confirm destructive actions, no unauthorized reverts).
71+- The user can interrupt at any time with ESC (preserve context) or Ctrl+C (hard exit).
72+- For real isolation, run mini-coder inside a sandboxed environment.
73+
74+**Summary:** Permission dialogs give the appearance of safety without the substance. Real security comes from sandboxing the environment, not gatekeeping individual tool calls. Mini-coder codes — isolating it is a job for the right tool.