llama.cpp constrains the sampler to ONLY emit tokens matching a GBNF grammar. For small models this kills format drift at the token level — `CMD: <cmd>` is enforced by the sampler rather than hoped for via prompt discipline. Probe finding (this commit's pre-implementation): cloud (Anthropic via Bedrock) silently IGNORES the `grammar` field — returns normally via standard sampling. Default passthrough is safe for all routes; no per-model opt-in/opt-out needed in v1. Changes: - broker.lua build_request: `if opts.grammar then req.grammar = opts.grammar end`. Misformed grammar surfaces at request time via the existing transport-error path. - repl.lua ask_ai: `grammar_override = config.routing.grammars [req_class]` (same gating shape as #86's system_prompts override). Passed via opts.grammar in the call_broker invocation. - safety.lua is_destructive threads cfg.safety.probe_grammar through opts.grammar so llm_probe constrains the YES/NO output. Skips the regex-match dance entirely when the model can't drift. Caller-provided opts.grammar takes precedence over cfg. - config.lua gains two commented examples: * routing.grammars per class * safety.probe_grammar for the destructive probe 6 unit cases verified (stubbed curl.post_sse / broker.chat): - default: no grammar in body - opts.grammar -> body contains grammar JSON-encoded - safety probe_grammar reaches llm_probe via opts - no probe_grammar configured -> opts.grammar nil - caller opts.grammar takes precedence over cfg.safety.probe_grammar E2E against live local broker: - `routing.grammars.default = "root ::= \\"ACK\\""` configured; prompted "tell me a long story about a fox" -> model output EXACTLY "ACK" (sampler forced; would normally produce paragraphs). Grammar passthrough end-to-end confirmed. Regression: test_safety 87/87, test_router_model 31/31, repl loads. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
aish
aish — AI-augmented conversational shell.
A single REPL that interleaves shell command execution and language-model conversation, backed by a llama.cpp HTTP broker. Implementation is LuaJIT 2.x with FFI bindings to libcurl, GNU readline, and libc — no C extensions, no build step, one source tree.
Why
Three flows that currently live in three windows fold into one:
- "Run this command and show me the output" — fast feedback loop, no copy-paste between terminal and chat.
- "Explain or write code based on the output we just looked at" — exec output is automatically injected into the model's context.
- "Plan and execute a multi-step task with confirmation gates" — landing in Phase 3 as Chuck Norris autonomous mode.
aish is not a wrapper around bash. It's a first-class interactive environment where the shell is one of several execution channels.
Status
| Component | State |
|---|---|
| Repository skeleton | ✅ in this commit |
| Phase 0 manifest | ✅ docs/PHASE0.md — locked |
| Phase 0 implementation | 🔜 next session |
| Phase 1+ | 📋 enumerated in PHASE0.md §11 |
Every module file currently raises not implemented (Phase 0 pending)
when called. luajit main.lua fails loudly at the first un-implemented
function, never silently.
Quick orientation
| Read this | If you want to know |
|---|---|
docs/PHASE0.md §1–2 |
What aish is and what Phase 0 ships |
docs/PHASE0.md §3 |
Technology decisions (LuaJIT, FFI, readline, libcurl, llama.cpp) |
docs/PHASE0.md §4 |
Directory layout — these file names are stable across all phases |
docs/PHASE0.md §5 |
How input is dispatched (meta / shell / AI) |
docs/PHASE0.md §6 |
Broker contract: /v1/chat/completions, CMD: extraction |
docs/PHASE0.md §10 |
Config schema and resolution order |
docs/PHASE0.md §11 |
Phase sequence (what lands when) |
docs/PHASE0.md §13 |
Open questions, tracked per phase |
CLAUDE.md |
Project conventions for AI-assisted contributors |
Directory layout
aish/
├── main.lua # entry point
├── repl.lua # readline loop, dispatch, prompt
├── broker.lua # llama.cpp HTTP client
├── router.lua # input classifier (meta/shell/AI)
├── executor.lua # command exec + CMD: extraction
├── context.lua # in-memory turn history
├── history.lua # disk persistence (Phase 1+)
├── safety.lua # destructive-op gate (Phase 3+)
├── renderer.lua # output formatting
├── config.lua # default model registry + preferences
├── ffi/
│ ├── curl.lua # libcurl easy interface
│ ├── readline.lua # GNU readline
│ ├── pty.lua # forkpty (Phase 1+)
│ └── libc.lua # chdir, errno, strerror
└── docs/
└── PHASE0.md # locked substrate
Build / runtime dependencies
System packages (Debian / ALARM / Arch names):
luajit(>= 2.0)libcurl4/libcurl-openssl-3runtimelibreadline8runtimelibc6runtime (always present)
No compilation, no luarocks, no make. Just luajit main.lua.
Running
Once Phase 0 ships:
luajit main.lua # uses ~/.config/aish/config.lua
luajit main.lua --config ./config.lua # explicit config path
AISH_CONFIG=/path/to/config.lua luajit main.lua
Config resolution order is documented in docs/PHASE0.md §10.
Configuration
config.lua is a Lua file returning a single table. The committed
config.lua in this repo is both the canonical example and the
development-fallback config (lowest precedence). Copy it to
~/.config/aish/config.lua and edit endpoints to your local llama.cpp
servers, or point AISH_CONFIG at your own.
The default endpoints assume mfritsche's home network:
fast→dirac.fritz.box:8081(Qwen2.5-Coder-7B q4 8k ctx)deep→dirac.fritz.box:8080(Qwen2.5-Coder-7B q4 32k ctx)cloud→hossenfelder.fritz.box:8082(forwards to OpenRouter)
Replace these with your own llama.cpp endpoints if you're not on that LAN.
License
Not yet selected. Default-private until decided.
Project conventions
See CLAUDE.md for contribution conventions, commit style,
and the phase-loop discipline this project follows.