Maven/internal/vision/vision.go

// Package vision is Maven's image-understanding seam (Vikunja #252,
// docs/plans/07-vision.md).
//
// One interface, Provider, with one method: describe an image, in words, in
// Russian, with an optional question about it. Text extraction is not a second
// method — "прочитай текст с картинки" is a prompt, and a vision-language model
// does not have a separate OCR mode to select.
//
// # What is deliberately NOT here
//
// The plan document called for a `RemoteProvider` calling "an OpenAI-compatible
// vision API endpoint". That step is refused: CLAUDE.md's surviving hard
// constraint after "never phones home" was deprecated is *no cloud model,
// inference stays on the box*, and a photo of his flat is the single worst thing
// to make an exception for. Endpoint is therefore checked at construction and
// must be a loopback or private address — a public host is a config error, not a
// deployment option. That check is the reason this package does not simply reuse
// internal/llm.Client.
//
// # State on this box, honestly
//
// The resident model is Qwen3-1.7B, which is text-only, and as of 2026-08-01
// there is no vision-capable gguf and no mmproj file anywhere under
// /mnt/hdd1/llms. So LocalProvider is written, tested against a fake server, and
// currently has nothing real to talk to: the describing half is BLOCKED on a
// model download (see docs/plans/07-vision.md for the candidates and the
// recipe). What works today without any download is the intake — an image
// arrives, is stored, is prepared — and the config seam that turns the rest on.
//
// Provider is nil-safe through Disabled, and vision is OFF unless configured,
// like the weather and telegram.
package vision

import (
	"bytes"
	"context"
	"encoding/json"
	"errors"
	"fmt"
	"net"
	"net/http"
	"net/url"
	"strings"
	"time"

	"github.com/kami/maven/internal/media"
	"github.com/kami/maven/internal/webfetch"
)

// DefaultTimeout — budget for one description. A small VLM doing prefill over
// an 896px image on a Vega iGPU is slow; 90s is generous because nobody is
// holding a conversation open on this path — the answer arrives as a reply or a
// note, and a too-tight timeout just means it never arrives at all.
const DefaultTimeout = 90 * time.Second

// DefaultMaxTokens — cap on the description. A paragraph is what a spoken
// answer can carry; a page is not.
const DefaultMaxTokens = 300

// DefaultPrompt — what she is asked when he did not ask anything specific,
// only sent a picture. Russian, because that is the channel language, and
// feminine self-reference is not needed here (the prompt is an instruction, the
// persona block is added by the caller that phrases the reply).
const DefaultPrompt = "Опиши, что на этом изображении. Коротко, 2-3 предложения. Если на нём есть текст, приведи его."

// Errors callers distinguish.
var (
	// ErrDisabled — vision is not configured. Returned by Disabled, which is
	// what the daemon wires when the config block is absent.
	ErrDisabled = errors.New("vision: not configured")
	// ErrNotPrivate — the configured endpoint is not on this box or its
	// network. Refused at construction; see the package comment.
	ErrNotPrivate = errors.New("vision: endpoint must be a local or private address")
	// ErrEmptyReply — the model returned nothing usable.
	ErrEmptyReply = errors.New("vision: empty description")
)

// Provider — the image-understanding contract. Describe takes an image already
// prepared by internal/media (decoded, downscaled, JPEG) and a prompt; an empty
// prompt means DefaultPrompt.
type Provider interface {
	Describe(ctx context.Context, im media.Image, prompt string) (string, error)
}

// Disabled — the floor Provider. Every call fails with ErrDisabled, which the
// caller turns into "я не умею смотреть картинки — зрение не настроено". It
// exists so that no call site needs a nil check and switching vision off cannot
// crash a turn.
type Disabled struct{}

// Describe always fails. The signature matches Provider.
func (Disabled) Describe(context.Context, media.Image, string) (string, error) {
	return "", ErrDisabled
}

// Config — how to reach the local vision server. Built from
// config.VisionConfig by the daemon; kept separate so this package does not
// import internal/config.
type Config struct {
	// Endpoint — base URL of a llama-server started with a vision model and its
	// mmproj (`llama-server -m model.gguf --mmproj mmproj.gguf`). Must be
	// loopback or private. The path is appended by the provider; give it
	// "http://127.0.0.1:8081".
	Endpoint string
	// Model — the model name to send. llama-server ignores it; it matters if the
	// endpoint is something else OpenAI-shaped on the same box.
	Model string
	// Timeout — per-description budget. 0 ⇒ DefaultTimeout.
	Timeout time.Duration
	// MaxTokens — cap on the reply. 0 ⇒ DefaultMaxTokens.
	MaxTokens int
	// Prompt — the default question. Empty ⇒ DefaultPrompt.
	Prompt string
}

// LocalProvider talks to a llama-server on this box over its
// /v1/chat/completions endpoint, sending the image as a data URI content part.
// It is the only real Provider, and it is a plain HTTP client: no subprocess
// spawning, because the daemon already owns llama-server lifecycle for the
// resident model and a second managed process is a bigger change than this task.
type LocalProvider struct {
	endpoint  string
	model     string
	prompt    string
	maxTokens int
	http      *http.Client
}

// NewLocal builds a LocalProvider, refusing a non-private endpoint. A bad URL
// or a public host is an error at construction so the daemon logs it once at
// startup instead of failing every turn.
func NewLocal(cfg Config) (*LocalProvider, error) {
	base := strings.TrimRight(strings.TrimSpace(cfg.Endpoint), "/")
	if base == "" {
		return nil, errors.New("vision: empty endpoint")
	}
	if err := checkPrivate(base); err != nil {
		return nil, err
	}
	timeout := cfg.Timeout
	if timeout <= 0 {
		timeout = DefaultTimeout
	}
	maxTokens := cfg.MaxTokens
	if maxTokens <= 0 {
		maxTokens = DefaultMaxTokens
	}
	prompt := strings.TrimSpace(cfg.Prompt)
	if prompt == "" {
		prompt = DefaultPrompt
	}
	return &LocalProvider{
		endpoint:  base,
		model:     cfg.Model,
		prompt:    prompt,
		maxTokens: maxTokens,
		http:      &http.Client{Timeout: timeout},
	}, nil
}

// Endpoint is the server this provider talks to. For logs and /dash.
func (p *LocalProvider) Endpoint() string { return p.endpoint }

// checkPrivate refuses any endpoint that is not on this box or its LAN. A
// hostname that is not an IP literal is refused too: "vision.example.com" could
// resolve anywhere, and resolving it here would be trusting DNS with his photos.
// localhost is the one name allowed, because it is the common case.
func checkPrivate(raw string) error {
	u, err := url.Parse(raw)
	if err != nil {
		return fmt.Errorf("vision: parse endpoint: %w", err)
	}
	if u.Scheme != "http" && u.Scheme != "https" {
		return fmt.Errorf("vision: endpoint scheme %q not supported", u.Scheme)
	}
	host := u.Hostname()
	if host == "" {
		return errors.New("vision: endpoint has no host")
	}
	if strings.EqualFold(host, "localhost") {
		return nil
	}
	ip := net.ParseIP(host)
	if ip == nil {
		return fmt.Errorf("%w: %q is a name, not an address", ErrNotPrivate, host)
	}
	if !webfetch.IsPrivateIP(ip) {
		return fmt.Errorf("%w: %s", ErrNotPrivate, host)
	}
	return nil
}

// chat request shapes. Content is the OpenAI multimodal array form: a text part
// and an image_url part whose url is a data URI.
type textPart struct {
	Type string `json:"type"`
	Text string `json:"text"`
}
type imageURL struct {
	URL string `json:"url"`
}
type imagePart struct {
	Type     string   `json:"type"`
	ImageURL imageURL `json:"image_url"`
}
type chatReq struct {
	Model     string  `json:"model,omitempty"`
	Messages  []any   `json:"messages"`
	MaxTokens int     `json:"max_tokens,omitempty"`
	Temp      float64 `json:"temperature"`
}
type userMsg struct {
	Role    string `json:"role"`
	Content []any  `json:"content"`
}
type chatResp struct {
	Choices []struct {
		Message struct {
			Content          string `json:"content"`
			ReasoningContent string `json:"reasoning_content"`
		} `json:"message"`
	} `json:"choices"`
}

// Describe sends the image and prompt and returns the model's answer. An empty
// prompt uses the configured default. Errors are wrapped, never fatal: the
// caller says she could not make out the picture and the turn continues.
func (p *LocalProvider) Describe(ctx context.Context, im media.Image, prompt string) (string, error) {
	if len(im.JPEG) == 0 {
		return "", media.ErrEmpty
	}
	q := strings.TrimSpace(prompt)
	if q == "" {
		q = p.prompt
	}
	body, err := json.Marshal(chatReq{
		Model:     p.model,
		MaxTokens: p.maxTokens,
		Messages: []any{userMsg{Role: "user", Content: []any{
			textPart{Type: "text", Text: q},
			imagePart{Type: "image_url", ImageURL: imageURL{URL: im.DataURI()}},
		}}},
	})
	if err != nil {
		return "", fmt.Errorf("vision: marshal: %w", err)
	}
	req, err := http.NewRequestWithContext(ctx, http.MethodPost,
		p.endpoint+"/v1/chat/completions", bytes.NewReader(body))
	if err != nil {
		return "", fmt.Errorf("vision: request: %w", err)
	}
	req.Header.Set("Content-Type", "application/json")

	resp, err := p.http.Do(req)
	if err != nil {
		return "", fmt.Errorf("vision: post: %w", err)
	}
	defer resp.Body.Close()
	if resp.StatusCode != http.StatusOK {
		return "", fmt.Errorf("vision: status %d", resp.StatusCode)
	}
	var out chatResp
	if err := json.NewDecoder(resp.Body).Decode(&out); err != nil {
		return "", fmt.Errorf("vision: decode: %w", err)
	}
	if len(out.Choices) == 0 {
		return "", ErrEmptyReply
	}
	text := strings.TrimSpace(out.Choices[0].Message.Content)
	if text == "" {
		// Same fallback as internal/llm: a Thinking model sometimes puts the
		// whole answer in reasoning_content and leaves content empty.
		text = strings.TrimSpace(out.Choices[0].Message.ReasoningContent)
	}
	if text == "" {
		return "", ErrEmptyReply
	}
	return text, nil
}