# gnoma

[![Release](https://img.shields.io/github/v/release/VikingOwl91/gnoma?style=for-the-badge&logo=go&logoColor=white&color=00ADD8)](https://github.com/VikingOwl91/gnoma/releases)
[![License](https://img.shields.io/badge/license-Apache%202.0-blue?style=for-the-badge)](LICENSE)
[![Go](https://img.shields.io/badge/go-1.26%2B-00ADD8?style=for-the-badge&logo=go&logoColor=white)](go.mod)
[![Container](https://img.shields.io/badge/ghcr.io-vikingowl91%2Fgnoma-2496ED?style=for-the-badge&logo=docker&logoColor=white)](https://github.com/VikingOwl91/gnoma/pkgs/container/gnoma)

**A provider-agnostic agentic coding assistant in Go.** gnoma routes each prompt
to the best available model — cloud or local — through a multi-armed bandit
router, executes tools on your behalf, and stays extensible through hooks,
skills, MCP servers, and plugins.

![gnoma TUI showing a routed turn](docs/img/gnoma-tui.png)

*Every turn shows which arm the router picked and why — here a local
`qwen3:14b` was selected for a `generation` task.*

## What makes gnoma different

- **Multi-armed bandit router.** Per-prompt arm selection based on
  capability gates, declared `Strengths`, latency, and cost. Visible in
  the TUI on every turn — no black box.
- **`[router].prefer = local | cloud | auto`.** Pin routing toward local
  models, cloud, or let the bandit decide. Offline-first workflows still
  reach for Claude when the local model would obviously flail.
- **Tier-0 SLM routing.** A tiny local model classifies each prompt and
  handles trivial tasks itself, keeping the heavy provider for real work.
- **Content boundary + secret scanner.** Every outgoing LLM message
  and incoming tool result is scanned for secrets (regex + Shannon
  entropy on long tokens), redacted or blocked at the content level.
  Paths are canonicalised (TOCTOU-safe), Unicode is sanitized
  (homoglyphs, BiDi tricks), and a `SafeProvider` boundary keeps
  incognito-mode data out of long-lived stores. *(Per-host network
  egress allowlist is on the roadmap, not in place today.)*
- **No phone-home.** gnoma itself sends nothing off-machine — zero
  analytics endpoint, zero metrics service, no remote logging.
  Prompts of course go to whatever provider you route them to:
  cloud arms ship data to that provider by design; pair
  Ollama/llama.cpp with `--incognito` if you want everything
  on-device.
- **Provider-agnostic from day one.** Anthropic, OpenAI, Google, Mistral,
  Ollama, llama.cpp, plus subprocess CLIs (`claude`, `codex`, `agy`,
  `vibe`). Mix cloud and local in the same session.
- **Vision end-to-end.** `[Image: /path]` markers in prompts, `Ctrl+V`
  paste in the TUI, capability-gated per arm.
- **Single static binary.** `CGO_ENABLED=0`, multi-arch container on
  ghcr.io. No daemon, no runtime deps.

## Status

Pre-1.0 (current: **v0.3.0**). Single maintainer, breaking changes
possible. The provider, router, and engine surfaces are settling;
config schema and TUI bindings may still shift between minor versions.
Apache 2.0.

## Table of contents

- [Install](#install)
- [Quickstart](#quickstart)
- [Vision / image input](#vision--image-input)
- [Providers](#providers)
- [Config](#config)
- [Routing defaults](#routing-defaults)
- [SLM routing](#slm-small-language-model-routing)
- [Session persistence](#session-persistence)
- [Extensibility](#extensibility)
- [Subcommands](#subcommands)
- [Security](#security)
- [Development](#development)
- [About](#about)
- [License](#license)

---

## Install

### Pre-built binary (no Go toolchain required)

Releases are built by [GoReleaser](.goreleaser.yml) for
`linux`, `darwin`, and `windows` × `amd64`/`arm64` as static (`CGO_ENABLED=0`)
archives. Grab the one matching your OS/arch from
<https://github.com/VikingOwl91/gnoma/releases>:

```sh
# Linux/macOS one-liner (substitute the asset URL):
curl -fsSL <ARCHIVE_URL> | tar -xz -C /tmp
sudo mv /tmp/gnoma /usr/local/bin/
gnoma --version
```

Windows: download the `_windows_*.zip`, extract `gnoma.exe`, and put it on
`%PATH%`.

### Docker

Multi-arch images (`linux/amd64`, `linux/arm64`) are published to GitHub
Container Registry on each tagged release:

```sh
docker pull ghcr.io/vikingowl91/gnoma:latest
docker run --rm -it -v "$PWD:/workspace" ghcr.io/vikingowl91/gnoma:latest --version
```

Mount your project as `/workspace` (the image's working directory) and pass
any provider keys via `-e VAR_NAME` — see the [Providers](#providers) table
for env-var names.

### Go users

```sh
go install somegit.dev/Owlibou/gnoma/cmd/gnoma@latest   # latest tagged
go install somegit.dev/Owlibou/gnoma/cmd/gnoma@main     # bleeding edge
```

### Build from source

```sh
git clone https://somegit.dev/Owlibou/gnoma && cd gnoma
make build       # → ./bin/gnoma
make install     # → $GOPATH/bin/gnoma
```

Requires Go 1.26+.

---

## Quickstart

Set at least one provider key (env var names are listed in the
[Providers](#providers) table below) — or run a local model and skip
the keys entirely.

```sh
gnoma                              # interactive TUI
echo "list files" | gnoma          # pipe / one-shot mode
gnoma --provider ollama            # use a local model (no API key needed)
gnoma --version
```

Inside the TUI, `Ctrl+X` toggles **incognito** (no session saved, no router
learning); `/help` lists slash commands; `Esc` cancels an in-flight turn.

---

## Vision / image input

`Ctrl+V` in the TUI pastes a screenshot from the system clipboard:
gnoma writes the bytes to your user cache and inserts a
`[Pasted image #imgN]` placeholder, which expands to `[Image: /path]`
when the turn is sent. You can also type a literal `[Image: /path]`
marker anywhere in a prompt to reference an existing file:

```
explain this error [Image: /tmp/screen.png] — what's the root cause?
```

Image markers are parsed by the engine, files larger than 10 MiB are
skipped (the marker stays as plain text), and the router only routes
vision-tagged turns to arms that declare the `Vision` capability
(Anthropic, OpenAI, Google, and Ollama models that advertise
multimodal support). Image paste is disabled under `--incognito` to
honour the no-persistence contract.

---

## Providers

| Provider | Env var | Default model | Also available |
|---|---|---|---|
| Anthropic | `ANTHROPIC_API_KEY` | `claude-sonnet-4-6` | `claude-opus-4-7`, `claude-haiku-4-5-20251001` |
| OpenAI | `OPENAI_API_KEY` | `gpt-5.5` | `gpt-5.5-pro`, `gpt-5.2`, `gpt-5.2-chat-latest` |
| Google (Gemini) | `GEMINI_API_KEY` (alt: `GOOGLE_API_KEY`) | `gemini-3.5-flash` | `gemini-3.1-pro-preview`, `gemini-3.1-flash-lite` |
| Mistral | `MISTRAL_API_KEY` | `mistral-large-latest` (Mistral Large 3) | `mistral-medium-3.5`, `magistral-medium-2509` |
| Ollama (local) | — | `qwen3:8b` (override with `--model`) | any model on your Ollama instance |
| llama.cpp (local) | — | reported by `/v1/models` | n/a |
| Subprocess (`claude`, `gemini`, `agy`, `codex`, `vibe` CLIs) | provider-specific | binary name | configurable via `[cli_agents]` |

Override per-invocation:

```sh
gnoma --provider anthropic --model claude-opus-4-7
gnoma --provider openai    --model gpt-5.5-pro     # GPT-5.5 is the default; pro is the higher-accuracy tier
gnoma --provider google    --model gemini-3.1-pro-preview
gnoma --provider ollama    --model qwen2.5-coder:3b
gnoma --provider llamacpp                          # model picked from server
```

`gnoma providers` prints every discovered provider, model, and CLI agent.

**Subprocess sandbox bypass.** The `agy` and `codex` CLIs each run with
their respective sandboxes enabled by default. Two env vars exist for the
rare case where a sandbox blocks legitimate work (e.g., reading files
outside the project root):

| Env var | Effect |
|---|---|
| `GNOMA_AGY_BYPASS_PERMISSIONS=1` | Skip agy's permission prompts |
| `GNOMA_CODEX_BYPASS_SANDBOX=1` | Disable codex's filesystem sandbox |

These are footguns — set them deliberately, per-invocation. They do not
disable gnoma's own permission system, hooks, or firewall.

### Local models

Start your local server, then point gnoma at it:

```sh
# Ollama (default http://localhost:11434/v1)
ollama pull qwen2.5-coder:3b
gnoma --provider ollama --model qwen2.5-coder:3b

# llama.cpp (default http://localhost:8080/v1)
llama-server --model /path/to/model.gguf --port 8080 --ctx-size 8192
gnoma --provider llamacpp
```

Override the endpoint in `.gnoma/config.toml`:

```toml
[provider.endpoints]
ollama   = "http://myhost:11434/v1"
llamacpp = "http://localhost:9090/v1"
```

---

## Config

Configuration merges (lowest → highest priority):

1. Built-in defaults
2. `~/.config/gnoma/config.toml` — global base
3. `~/.config/gnoma/profiles/<name>.toml` — active profile (when profile mode is enabled)
4. `<projectRoot>/.gnoma/config.toml` — project override
5. Environment variables (`GNOMA_PROVIDER`, `GNOMA_MODEL`, `*_API_KEY`)

Example global config:

```toml
[provider]
default = "anthropic"
model   = "claude-sonnet-4-6"

[provider.api_keys]
anthropic = "${ANTHROPIC_API_KEY}"

[provider.endpoints]
ollama   = "http://localhost:11434/v1"
llamacpp = "http://localhost:8080/v1"

[permission]
mode = "auto"      # default | accept_edits | bypass | deny | plan | auto

[session]
max_keep = 20      # sessions retained per project
```

### Profiles

Drop multiple configs under `~/.config/gnoma/profiles/` and switch with
`--profile <name>` or `/profile <name>`. Each profile keeps its own router
quality data and session history. Full details: [docs/profiles.md](docs/profiles.md).

---

## Routing defaults

Discovered arms ship with opinionated defaults — `Strengths` (per-task
preference) and `MaxComplexity` (ceiling above which the arm won't be
picked) — so a freshly-pulled fleet routes sensibly without any
`[[arms]]` config. Defaults match against the model ID with
longest-prefix-wins; size-keyed families (Qwen 3, Ministral 3, tiny3.5,
etc.) scale `MaxComplexity` down for smaller variants automatically.

Non-chat models (`embeddinggemma`, `whisper-base`, `kokoros`,
`vibevoice`, `*-asr`, `*-tts`, `*-audio`, `*-reranker`,
`*-embedding`) are skipped during discovery so they never register
as broken chat arms.

| Local family | Strengths | MaxComplexity |
|---|---|---|
| `qwen3-coder` / `devstral` | Generation, Refactor, Debug | 0.85 |
| `qwen2.5-coder` | Generation, Refactor, UnitTest | 0.70 |
| `phi-4` | Planning, Debug, Review | 0.65 |
| `gemma4` (base ~9B) | Explain, Review, Generation | 0.70 |
| `gemma4-e` / `gemma-4-e` (edge 2B–4B) | Explain, Boilerplate | 0.45 |
| `mistral-small-3` | Orchestration, Review | 0.65 |
| `qwen3` | Generation, Refactor, Debug | 0.50–0.75 (size-keyed) |
| `qwen3.5` | Boilerplate, Explain, Orchestration | 0.40–0.65 |
| `ministral-3` | Orchestration, Planning | 0.35–0.70 |
| `tiny3.5` | Boilerplate, Explain | 0.20–0.30 |
| `phi-4-mini` / `llama3.2` / `granite` | Boilerplate, Explain | 0.30–0.35 |
| `functiongemma` | (Disabled — reserved for tool-router role) | 0.40 |

| Cloud model | Strengths | CostWeight |
|---|---|---|
| `claude-opus-4-7` | Planning, SecurityReview, Debug, Refactor | 0.3 |
| `claude-sonnet-4-6` | Generation, Refactor, Review | 0.7 |
| `gpt-5.5` | Planning, SecurityReview, Generation | 0.3 |
| `gpt-5.3-codex` | Generation, Refactor, Debug, UnitTest | 0.6 |
| `gpt-5.2` | Orchestration, Review | 0.8 |
| `gemini-3.1-pro` | Planning, Review, Orchestration | 0.5 |
| `gemini-3.5-flash` | Boilerplate, Explain, Orchestration | 1.2 |

`CostWeight` scales how much $/Mtok matters in scoring: values below
1.0 keep expensive frontier arms competitive on high-stakes tasks
(Planning, SecurityReview); values above 1.0 penalize cost more so
cheap fast arms only win when cost is genuinely decisive.

### Overriding the defaults

Drop an `[[arms]]` block in `config.toml` to override per-arm
`Strengths` or `CostWeight`. User values win — defaults only fill
zero fields:

```toml
[[arms]]
id          = "anthropic/claude-opus-4-7"
strengths   = ["security_review", "planning", "debug"]
cost_weight = 0.2  # weight cost even less than the default 0.3

[[arms]]
id        = "ollama/qwen3-coder:30b"
strengths = ["generation", "refactor"]
```

Full rationale and benchmark sources behind these defaults:
[`docs/superpowers/plans/2026-05-23-routing-defaults-refresh.md`](docs/superpowers/plans/2026-05-23-routing-defaults-refresh.md).

### Preferring local vs cloud

`[router].prefer` biases routing toward one camp without hard-filtering
the other:

```toml
[router]
prefer = "auto"   # auto (default) | local | cloud
```

| Value | Effect |
|---|---|
| `"auto"` | No bias. Tier order (SLM → CLI-agent → local → cloud) decides, with Strengths and quality scores breaking ties. Default. |
| `"local"` | Cloud arms are demoted by 2 tiers. Local + CLI-agent arms always win unless no local option is feasible. |
| `"cloud"` | Local arms are demoted by 2 tiers. Cloud arms win, **except** for tier-0 SLMs — a small specialist arm whose `MaxComplexity` ceiling fits the task still wins, by design (the SLM is for small stuff). |

Three things still take priority over `prefer`:

- `--provider X` pins the forced arm.
- Incognito (`Ctrl+X` or `--incognito`) hard-filters cloud arms — `prefer = "cloud"` under incognito still picks a local arm.
- A `Strengths`-tagged arm always wins its tagged task type, regardless of `prefer`. Tag Opus with `[security_review]` under `prefer = "local"` and Opus still wins SecurityReview tasks.

CLI-agent subprocess arms (`claude`, `gemini`, `vibe`) count as **local** for this knob — they proxy to cloud but run as local processes. Use `--provider <name>` if you need to pin a specific subprocess.

---

## SLM (small-language-model) routing

gnoma can run a tiny local model alongside the main provider to:

- **Classify** each prompt (task type + complexity + tool requirement) so the
  router picks the right arm.
- **Execute** trivial tasks itself (knowledge questions, single file reads,
  anything with complexity ≤ 0.3), keeping the heavy provider for real work.

```toml
[slm]
enabled         = true
backend         = "auto"      # ollama | llamacpp | llamafile | openaicompat | auto | disabled
model           = "qwen3:0.6b"
register_as_arm = true        # default; set to false to make the SLM classifier-only
                              # (e.g. for FunctionGemma, code-completion-tuned models)
classify_timeout = "15s"      # default; bump higher for slow cold-loads
```

Setup, presets, and verification: [docs/slm-backends.md](docs/slm-backends.md).
The `auto` backend probes Ollama → llama.cpp → llamafile on startup and picks
the first reachable option. Inspect with `gnoma slm status` and
`gnoma router stats`.

---

## Session persistence

Sessions are auto-saved per project under `.gnoma/sessions/<id>/` after each
completed turn. On a crash you lose at most the current in-flight turn.

```sh
gnoma --resume              # interactive picker
gnoma --resume <id>         # restore by ID
gnoma -r                    # shorthand
gnoma --incognito           # no save, no router learning
```

Inside the TUI: `/resume`, `/resume <id>`, `Ctrl+X` (incognito toggle).

Router-quality data (EMA scores) is stored at
`~/.config/gnoma/quality.json` (or `quality-<profile>.json` in profile mode).

---

## Extensibility

### MCP servers

Connect any [MCP](https://modelcontextprotocol.io)-compatible server:

```toml
[[mcp_servers]]
name    = "git"
command = "mcp-server-git"
args    = ["--repo", "."]
timeout = "30s"

# Optionally replace a built-in tool with an MCP one
[mcp_servers.replace_default]
exec = "bash"
```

MCP tools appear as `mcp__{server}__{tool}` unless mapped via `replace_default`.

### Skills

Drop markdown files into `.gnoma/skills/` or `~/.config/gnoma/skills/`. Invoke
with `/<skill-name>`. List with `/skills`.

### Hooks

Shell commands run on tool events (`pre_tool_use`, `post_tool_use`, etc.):

```toml
[[hooks]]
name         = "block-rm-rf"
event        = "pre_tool_use"
type         = "command"
exec         = "bash-safety-check.sh"
tool_pattern = "bash*"
```

Ordering rules: [ADR-004](docs/essentials/decisions/004-posttooluse-hook-ordering.md).

### Plugins

Plugins bundle skills, hooks, and MCP server configs. Drop a plugin directory
into `~/.config/gnoma/plugins/` (global) or `<project>/.gnoma/plugins/`
(project-local); gnoma auto-discovers them on startup.

Each plugin's `plugin.json` is pinned by SHA-256 on first load
(Trust-On-First-Use). A manifest that changes between runs is refused with a
clear error and a re-enrolment hint. Full model:
[docs/plugins-trust.md](docs/plugins-trust.md) and
[ADR-003](docs/essentials/decisions/003-plugin-trust.md).

### Elfs (sub-agents)

The `spawn_elfs` tool decomposes work into parallel sub-tasks. See
[`internal/skill/skills/batch.md`](internal/skill/skills/batch.md) for the
built-in batching skill.

---

## Subcommands

| Command | What it does |
|---|---|
| `gnoma providers` | List every discovered provider, model, and CLI agent |
| `gnoma profile list` / `show <name>` | Profile diagnostics |
| `gnoma router stats` | Quality EMA + classifier source breakdown |
| `gnoma slm setup` / `slm status` | Manage the llamafile-backed SLM |

`gnoma --help` for the full flag set.

---

## Security

gnoma runs tools and shell commands on your behalf. The
[`internal/security`](internal/security) package canonicalises every path
(TOCTOU-safe), scans every outgoing LLM message and incoming tool result
for secrets (regex + Shannon entropy) before it reaches the model, and
sanitizes Unicode (homoglyphs, BiDi tricks). The `SafeProvider` boundary
keeps incognito-mode data out of long-lived stores.

> **Scope note.** The current "firewall" is a content boundary — it
> redacts/blocks secrets in inputs and outputs. It is **not** a
> network-egress firewall: outgoing HTTP from tools and providers goes
> through stock `http.Client`, with no per-host allowlist or
> dial-layer enforcement. Per-host egress rules and a per-session
> audit log of blocked/redacted events are tracked in
> [TODO.md](TODO.md).
>
> **Data flow.** gnoma itself emits no telemetry to external services
> — no analytics, no metrics endpoint, no remote logging. When you
> route to a cloud provider (Anthropic, OpenAI, Google, Mistral),
> prompts and tool data are sent to that provider as required to
> fulfill the request — by design. For fully on-device operation,
> use Ollama or llama.cpp and `--incognito`.

### Entropy false-positive reduction

The secret scanner also computes Shannon entropy on long unstructured
tokens to catch unknown-format secrets. Under a lowered threshold or
`redact_high_entropy = true`, this can fire on shapes that are never
secrets (UUIDs, SHA digests, ISO-8601 timestamps, URLs). Opt into the
format-aware safelist to skip them:

```toml
[security]
entropy_threshold    = 3.5
redact_high_entropy  = true
entropy_safelist     = ["uuid", "sha_hex", "iso8601", "url"]
```

Default is an empty list — pre-safelist behaviour. Skips are logged
(`Debug`-level, per pattern, token length only — never the bytes) so the
real false-positive rate is measurable on real workloads.

### Startup safety check

gnoma classifies the current working directory before launch and
refuses, warns, or allows based on tier:

| Tier | What | Behavior |
|---|---|---|
| **Refuse** | `/`, `/etc`, `/sys`, `/proc`, `/usr`, `/var`, `/bin`, `/sbin`, `/boot`, `/root`, `/dev` (and macOS equivalents `/System`, `/Library`, `/private`, `/Applications`) | Refuses to start. Exit code 2. |
| **Warn** | `$HOME`, `~/Desktop`, `~/Downloads`, `~/Documents`, `~/.config`, `~/.local`, `~/.cache`, `/tmp` | Prints a warning banner and waits for `y` keypress to continue. Anything else (including piped EOF) aborts with exit 1. |
| **OK** | Anywhere with a project marker (`.gnoma/`, `go.mod`, `package.json`, `pyproject.toml`, `Cargo.toml`, `Makefile`, `Dockerfile`, `build.gradle`, `pom.xml`) or inside a git repo | No prompt. |

A project marker anywhere — including inside `$HOME` — promotes the
directory to OK. The banner is shown for every tier and summarizes
cwd, git branch, project type, provider, model, modes, and a
top-level sensitive-file inventory (`.env`, SSH keys, `*.pem`,
`.ssh/`, `.aws/`, etc.).

```toml
[safety]
refuse_in_system_dirs  = true   # default
warn_in_home           = true   # default
require_project_marker = false  # default — being inside a git repo is enough
```

Bypass all safety checks with `--dangerously-allow-anywhere`. Required
for non-interactive invocations (piped stdin, CI) in warn-tier dirs,
since there's no human present to consent.

Containers (`/.dockerenv` or `/run/.containerenv` present) automatically
downgrade refuse-tier paths to warn-tier — devcontainers commonly run
from `/` or `/workspace`.

Full design:
[`docs/superpowers/plans/2026-05-23-startup-safety-banner.md`](docs/superpowers/plans/2026-05-23-startup-safety-banner.md).

Architecture references:

- [docs/essentials/INDEX.md](docs/essentials/INDEX.md) — full architecture map
- [docs/essentials/decisions/](docs/essentials/decisions/) — ADRs 001–004

---

## Development

```sh
make build          # ./bin/gnoma
make test           # unit tests
make test-integration  # //go:build integration — requires real API keys
make cover          # coverage.html
make lint           # golangci-lint
make check          # fmt + vet + lint + test
```

Architecture, conventions, and TDD workflow: [CONTRIBUTING.md](CONTRIBUTING.md).

---

## About

Named after the northern pygmy-owl (*Glaucidium gnoma*); agents are called
**elfs** (elf owl).

- **Upstream:** <https://somegit.dev/Owlibou/gnoma>
- **GitHub mirror:** <https://github.com/VikingOwl91/gnoma> (read-only;
  PRs go to upstream Gitea)

## License

Apache License 2.0. See [LICENSE](LICENSE) and [NOTICE](NOTICE).