feat(summarizer): add SummarizeForPush and use for Pushover; keep full WebUI/on-demand output; clamp only on push\ndocs: add AGENTS.md; revert CLAUDE.md release section

This commit is contained in:
Thomas Cravey 2025-09-05 06:58:38 -05:00
parent 2f9ab6a414
commit 8dc52976eb
7 changed files with 512 additions and 21 deletions

40
AGENTS.md Normal file
View file

@ -0,0 +1,40 @@
# Repository Guidelines
## Architecture Overview
- Role: Additional IRC client (not a bot) connected to a soju bouncer with your same nick. Uses a distinct clientId/username suffix for perclient history.
- Ingestion: Raw IRC via `internal/soju` with optional CHATHISTORY backfill, quiet hours, and rate limiting for alerts.
- Storage: SQLite (`internal/store`) for channel logs and metadata.
- Notifications: Pushover (`internal/notifier`) for mentions and scheduled digests.
- Summaries: OpenAI LLM (`internal/summarizer`, default gpt5) for link and channel summaries.
- Web UI: `internal/httpapi` serves SSE live view, link cards, inline article summaries, and an ondemand summary page.
## Project Structure
- `cmd/sojuboy/`: entrypoint and service wiring
- `internal/{soju,store,notifier,summarizer,httpapi,scheduler}`: core packages
- Root: `Dockerfile`, `docker-compose.yml`, `README.md`, `CHANGELOG.md`
## Build, Run, and Deploy
- Local: `go build -o sojuboy ./cmd/sojuboy` or `go run ./cmd/sojuboy`
- Docker (Synologyfriendly): `docker compose up --build` (exposes `8080`)
- Health: `./sojuboy --health`, `curl :8080/healthz`, `curl :8080/ready`
- Example .env: `SOJU_HOST=host` `SOJU_PORT=6697` `SOJU_TLS=true` `IRC_NICK=nick` `IRC_USERNAME=user/network@sojuboy` `IRC_PASSWORD=...` `CHANNELS=#chan1,#chan2` `NOTIFIER=pushover` `PUSHOVER_USER_KEY=...` `PUSHOVER_API_TOKEN=...` `LLM_PROVIDER=openai` `OPENAI_API_KEY=...` `DIGEST_CRON=0 0 9 * * *` `HTTP_TOKEN=longtoken`
## Coding Style & Naming
- Format: `gofmt -s -w .`; lint: `go vet ./...`; deps: `go mod tidy`
- Indentation: Go uses tabs; YAML/JS use 2 spaces
- Naming: packages lowercase; exported `CamelCase`; files `snake_case.go`
- Logging: use `slog` via `internal/logging`; include context
## Testing Guidelines
- Use standard `testing` with tabledriven tests; place beside code (`*_test.go`)
- For store tests, use inmemory DB: `store.Open(ctx, ":memory:")`
- Run: `go test ./...` and `go test -cover ./...`
## Commits & PRs
- Commits: short, imperative; conventional tags like `feat(ui):`, `fix(store):`, `docs:` (see history)
- PRs: description, linked issues, repro/verification steps; screenshots for Web UI changes
- Update `README.md`/`CHANGELOG.md` when flags, endpoints, or UX change
## Security & Config
- Keep secrets in `.env` (gitignored); never commit API keys
- Synology: map `/data` volume; protect UI with `HTTP_TOKEN` (cookie/Bearer/query); bind `127.0.0.1:8080` behind DSM Reverse Proxy if desired

339
CLAUDE.md Normal file
View file

@ -0,0 +1,339 @@
# CLAUDE.md
This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
## Development Commands
### Building and Running
- `go build -o sojuboy ./cmd/sojuboy` - Build the binary
- `go run ./cmd/sojuboy` - Run directly with Go
- `docker-compose up -d --build` - Build and run with Docker Compose
- `docker build -t sojuboy .` - Build Docker image
### Health Checks
- `./sojuboy --health` - Local readiness check (exits 0/1)
- `curl http://localhost:8080/healthz` - HTTP health check
- `curl http://localhost:8080/ready` - HTTP readiness check (requires soju connection)
### Testing and Validation
- `go mod tidy` - Clean up module dependencies
- `go vet ./...` - Run Go static analysis
- `go fmt ./...` - Format Go code
## Architecture Overview
This is a Go service that acts as an IRC bouncer companion for soju, providing notifications, message storage, and AI summaries.
### Core Components
**Entry Point**: `cmd/sojuboy/main.go` - Main application entry point with service wiring and configuration
**Core Modules** (all in `internal/`):
- `config/` - Environment-based configuration loading
- `soju/` - IRC client for soju bouncer connection with IRCv3 capability negotiation
- `store/` - SQLite storage layer with WAL mode enabled
- `notifier/` - Pushover notification system (pluggable interface)
- `summarizer/` - OpenAI integration for AI digests and link summarization
- `scheduler/` - Cron-based digest scheduling and retention jobs
- `httpapi/` - HTTP API server with Web UI (templates + static assets)
- `logging/` - Structured logging setup
### Key Architectural Patterns
**IRC Integration**: Uses raw IRC protocol with sorcix/irc parser, implements irssi-style authentication for soju multi-client support (`username/network@client` format)
**Message Flow**: IRC messages → storage → mention detection → notifications + AI processing
**Web Architecture**: Go stdlib `net/http` with `html/template`, embedded static assets, SSE for real-time updates
**AI Integration**: OpenAI client with GPT-5 default (fallback to GPT-4o-mini), separate prompts for conversation digests vs link summarization
### Configuration
All configuration via environment variables (see README.md for complete reference). Key patterns:
- Soju connection: `SOJU_HOST`, `SOJU_PORT`, `SOJU_TLS`, `IRC_USERNAME` format critical for per-client history
- Required for functionality: `OPENAI_API_KEY`, `PUSHOVER_USER_KEY`, `PUSHOVER_API_TOKEN`
- Default data path: `/data/app.db` (SQLite with WAL mode)
### HTTP API Structure
- Authentication via `HTTP_TOKEN` (cookie or Bearer/query param)
- JSON APIs: `/tail`, `/history`, `/trigger`, `/linkcard`, `/linksummary`, `/metrics`
- Web UI with SSE tail, infinite scroll, link cards, on-demand summarization
- Channel-based operations (URL encode `#` as `%23` in requests)
### Storage Schema
SQLite with modernc.org/sqlite driver:
- Messages table with server-time timestamps
- Channel-based organization
- Retention management (daily cleanup at 03:00)
- WAL mode for concurrent access
### Development Notes
- Go 1.23+ required (see go.mod)
- Static binary compilation for distroless Docker image
- No external build tools (Make, npm, etc.) - pure Go toolchain
- Configuration via .env file for local development
- Docker health checks via `--health` flag
## Code Review Findings
### 🚀 Major Refactors for iOS Features
#### 1. Extract Message Processing Service
The `alert` function in `main.go:143-212` is doing too much:
- Message storage
- Notification logic
- Rate limiting
- DM vs channel handling
**Recommendation:** Create `internal/messaging/processor.go` with separate concerns:
```go
type MessageProcessor struct {
store Store
notifiers []Notifier // Multi-target for iOS + Pushover
rateLimiter RateLimiter
broadcaster Broadcaster
}
```
#### 2. Multi-Target Notification System
Current system hardcoded to single Pushover notifier. For iOS:
```go
type NotificationDispatcher struct {
targets []Notifier // Pushover, APNs, etc.
}
```
#### 3. Extract Rate Limiter
Move `rateLimiter` struct to `internal/ratelimit/` package for reusability and testing.
### ⚡ Critical Performance Issues
#### 1. SQLite Connection Bottleneck
`store.go:30` - `SetMaxOpenConns(1)` is extremely restrictive:
```go
// Current: Major bottleneck
db.SetMaxOpenConns(1)
// Better: Allow concurrent reads
db.SetMaxOpenConns(5)
db.SetMaxIdleConns(2)
```
#### 2. Inefficient Array Reversal
`store.go:147-150` - Manual reversal is inefficient:
```sql
-- Better: Use SQL ORDER BY instead
SELECT ... ORDER BY at ASC LIMIT ?
```
#### 3. OpenAI Client Recreation
`openai.go:60` - Client created per request. Should use singleton pattern with connection pooling.
### 🗑️ Dead Code Cleanup
1. **Entire unused package:** `internal/ircclient/base64.go` - appears completely unused
2. **Unused imports:** `io` package in `main.go` only for health check discard
3. **Dead helper:** `strconvI()` in `summarizer.go:475` - just use `strconv.Itoa` directly
4. **Rarely used export:** `config.GetEnvInt()` - used once, could be private
### 🔧 Easy Wins
#### 1. Fix Deprecated API
`templates.go:48` - `strings.Title()` is deprecated:
```go
// Replace with
data["Title"] = cases.Title(language.English).String(base)
```
#### 2. Hard-coded Values
- Health check port hardcoded to `8080` - use configured listen address
- Text limits hardcoded in summarizer - make configurable
- Temperature values hardcoded - make configurable
#### 3. Database Schema
Missing indexes for performance:
```sql
CREATE INDEX IF NOT EXISTS idx_messages_author ON messages(author);
CREATE INDEX IF NOT EXISTS idx_messages_time ON messages(at);
```
### 🛡️ Error Handling Improvements
#### 1. Silent Failures
`pushover.go:24` - Returns `nil` on error conditions without logging:
```go
// Better error handling
if p == nil || p.app == nil || p.userKey == "" {
return errors.New("pushover not configured")
}
```
#### 2. Message Truncation
No user feedback when messages truncated to 1024 chars.
### 📱 iOS-Specific Optimizations
#### 1. Message Broadcaster Enhancement
Current SSE broadcasting is basic - needs:
- Connection state management
- Message queuing for disconnected clients
- WebSocket support alongside SSE
#### 2. Auth Token Management
Simple token auth needs enhancement for mobile:
- JWT tokens with refresh capability
- Device-specific sessions
- Proper expiration handling
### 🎯 Implementation Priority
**Phase 1 (High Impact):**
1. Fix SQLite connection limit
2. Extract message processor service
3. Remove dead code
4. Fix deprecated APIs
**Phase 2 (iOS Foundation):**
1. Multi-target notification system
2. Enhanced message broadcasting
3. JWT authentication improvements
**Phase 3 (Performance):**
1. OpenAI client pooling
2. Database index optimization
3. Rate limiter improvements
## Application Behavior & Quirks
### Message Storage Logic
- **DM Handling**: Direct messages TO the bot are stored under the sender's nick, not the bot's nick (`main.go:146-151`)
- **Message Deduplication**: Uses `INSERT OR IGNORE` with optional `msgid` for soju message deduplication
- **Channel Normalization**: All channel lookups are case-insensitive via `lower()` SQL function
### Notification Logic
- **Rate Limiting**: Per-channel + per-keyword combinations, not global (`main.go:199`)
- **Quiet Hours**: Local timezone-based, supports midnight wrap-around (`config.go:267-271`)
- **Urgent Keywords**: Bypass quiet hours but still respect rate limits
- **Backfill Suppression**: Messages older than 5 minutes are not notified unless `NOTIFY_BACKFILL=true`
### IRC Protocol Specifics
- **Authentication**: Uses raw PASS/USER, not SASL (despite config hint)
- **Capability Negotiation**: Requests IRCv3 `draft/event-playback` for automatic message replay
- **Fallback History**: CHATHISTORY LATEST only used when event-playback unavailable
- **Connection Management**: Exponential backoff (1s → 30s max) with automatic reconnection
### AI Summarization Behavior
- **Message Grouping**: Same author within `SUMM_GROUP_WINDOW` (default 120s) are concatenated
- **Link Processing**: Extracts URLs via regex, fetches with readability for clean text
- **Image Handling**: Direct image URLs sent as vision inputs to GPT-5
- **Fallback Logic**: Local text-based summary if OpenAI fails
- **Model Selection**: Auto-detects reasoning models (gpt-5, o1) to skip temperature setting
### Web UI Implementation
- **SSE Streaming**: Real-time message delivery with 15s heartbeat
- **Link Cards**: Cached 24h, special handling for X/Twitter oEmbed and YouTube
- **Authentication**: Cookie-based with 7-day expiration, secure flag auto-detection
- **Channel Selector**: Populated from DB + config fallback for empty databases
### Storage Patterns
- **WAL Mode**: Enabled for concurrent read access despite single connection limit
- **Retention**: Daily cleanup at 03:00 local time, configurable days
- **Timestamps**: Always stored as UTC, server-time tags preferred over local time
- **Migration**: Best-effort schema updates with ignored errors for missing columns
### Error Handling Patterns
- **Silent Failures**: Many subsystems return `nil` errors to prevent cascading failures
- **Graceful Degradation**: Missing summarizer/notifier doesn't break core IRC functionality
- **Connection Resilience**: IRC client handles ERROR messages and reconnects automatically
### Security Considerations
- **Token Storage**: HTTP_TOKEN stored in plain text, suitable for internal deployments
- **CORS**: No CORS headers - designed for same-origin web UI access
- **TLS**: Configurable per connection, defaults to enabled
- **Input Sanitization**: Minimal - relies on template escaping and database parameterization
### Performance Characteristics
- **Memory Usage**: Link/summary caches unbounded - potential memory leak over time
- **Concurrent Requests**: Severely limited by single DB connection
- **OpenAI Rate Limits**: No built-in rate limiting or retry logic
- **SSE Scalability**: In-memory subscriber map, no persistence across restarts
### Development Workflow
- **Hot Reload**: Not supported - requires full restart for config/code changes
- **Logging**: Structured JSON to stdout, debug mode shows IRC protocol messages
- **Health Monitoring**: Separate metrics endpoint for Prometheus integration
- **Docker**: Multi-stage build with distroless final image for security
### Common Gotchas
- **Channel Names**: Must URL-encode `#` as `%23` in HTTP requests
- **Time Formats**: Mix of RFC3339 and RFC3339Nano depending on context
- **Case Sensitivity**: IRC nicks/channels case-insensitive, but stored as-received
- **CHATHISTORY**: Requires soju-specific timestamp format, may not work with other bouncers
- **Health Check**: Hard-coded to port 8080 regardless of configured listen address
## Planned iOS App Development
### Context
User wants to build a native iOS Swift app that consumes this Go server's HTTP APIs. The server should remain in Go (excellent fit for backend services), while the iOS app provides a native mobile experience.
### Required Server Enhancements for iOS
**1. iOS Push Notifications (Priority 1)**
- Add Apple Push Notification service (APNs) integration alongside Pushover
- New endpoints: `POST /api/devices` (device token registration)
- Enhanced notification system supporting multiple targets
- Environment variables: `APNS_KEY_ID`, `APNS_TEAM_ID`, `APNS_BUNDLE_ID`, `APNS_KEY_PATH`, `APNS_SANDBOX`
**2. Enhanced Real-time Communication**
- Add WebSocket support alongside existing SSE (better for iOS background handling)
- Message queuing for disconnected clients
- New endpoints: `GET /api/ws?channel=#chan`, `GET /api/missed?since=<timestamp>`
**3. Mobile-Optimized Authentication**
- JWT tokens with refresh capability (current simple token is too basic)
- Device-specific authentication for multiple iOS devices
- New endpoints: `POST /api/auth/login`, `POST /api/auth/refresh`, `POST /api/auth/logout`
**4. Offline Support & Sync**
- Cursor-based pagination (improve current `before` parameter approach)
- Message read/unread state tracking per device
- Enhanced endpoints: `GET /api/sync?cursor=<id>&limit=100`, `POST /api/read-receipts`, `GET /api/unread-count`
**5. iOS-Specific Optimizations**
- Batch operations: `POST /api/linkcards/batch`
- Image size variants: `GET /api/linkcard?url=...&size=thumb`
- Performance optimizations for mobile bandwidth
### iOS App Feature Parity Goals
The native app should match current web UI functionality:
- Live chat view with real-time updates
- Link cards and on-demand link summarization
- Native push notifications (replacing Pushover for mobile users)
- On-demand AI summarization
- Channel switching and history browsing
- Infinite scroll for chat history
### Database Schema Extensions Needed
- `devices` table - iOS device tokens and user associations
- `read_receipts` table - per-device read state tracking
- `message_queue` table - offline message queueing
### Implementation Priority for iOS Support
1. APNs integration (native notifications)
2. WebSocket + connection management
3. JWT authentication system
4. Offline sync capabilities
5. Mobile-specific API optimizations
### Architecture Decision: Keep Go Backend
**Question Asked:** Could this project be rewritten in Swift?
**Answer:** Technically feasible but not recommended.
**Pros of Swift rewrite:** Strong type safety, modern concurrency (async/await), cross-platform potential
**Cons:** Limited ecosystem for IRC libraries, SQLite libraries, cron scheduling, HTTP templating; larger Docker images; significant development effort to recreate working functionality
**Recommended Approach:** Keep Go backend (excellent fit for network services, mature ecosystem, single static binary deployment) + build native Swift iOS frontend. This leverages both platforms' strengths effectively.

1
TODO.md Normal file
View file

@ -0,0 +1 @@
[ ] Add Search fucntionality - to search chat logs

View file

@ -252,14 +252,17 @@ func main() {
logger.Error("digest fetch", "err", err)
continue
}
summary, err := sum.Summarize(ctx, ch, msgs, window)
summary, err := sum.SummarizeForPush(ctx, ch, msgs, window)
if err != nil {
logger.Error("digest summarize", "err", err)
continue
}
if nt != nil {
title := "IRC digest " + ch + " (" + window.String() + ")"
if err := nt.Notify(ctx, title, summary); err != nil {
// Safety clamp for provider limits
msg := strings.TrimSpace(summary)
if len(msg) > 1024 { msg = msg[:1024] }
if err := nt.Notify(ctx, title, msg); err != nil {
logger.Error("digest notify", "err", err)
} else {
atomic.AddInt64(&metrics.NotificationsSent, 1)

View file

@ -156,7 +156,14 @@ func (s *Server) handleTrigger(w http.ResponseWriter, r *http.Request) {
}
ctxSum, cancel := context.WithTimeout(ctx, tout)
defer cancel()
summary, err := s.Summarizer.Summarize(ctxSum, channel, msgs, window)
var summary string
var err error
// If we will push, use a push-optimized prompt; otherwise allow full-length
if strings.EqualFold(r.URL.Query().Get("push"), "1") || strings.EqualFold(r.URL.Query().Get("push"), "true") {
summary, err = s.Summarizer.SummarizeForPush(ctxSum, channel, msgs, window)
} else {
summary, err = s.Summarizer.Summarize(ctxSum, channel, msgs, window)
}
if err != nil {
if s.Logger != nil {
s.Logger.Error("http trigger summarizer", "err", err)
@ -167,13 +174,18 @@ func (s *Server) handleTrigger(w http.ResponseWriter, r *http.Request) {
}
// Only push if explicitly requested
pushFlag := strings.EqualFold(r.URL.Query().Get("push"), "1") || strings.EqualFold(r.URL.Query().Get("push"), "true")
if pushFlag && s.Notifier != nil {
title := fmt.Sprintf("IRC digest %s (%s)", channel, window)
_ = s.Notifier.Notify(ctx, title, summary)
if s.Metrics != nil {
atomic.AddInt64(&s.Metrics.NotificationsSent, 1)
}
}
if pushFlag && s.Notifier != nil {
title := fmt.Sprintf("IRC digest %s (%s)", channel, window)
// Trim only for push notifications (Pushover limit ~1024)
msg := strings.TrimSpace(summary)
if len(msg) > 1000 {
msg = msg[:1000] + "…"
}
_ = s.Notifier.Notify(ctx, title, msg)
if s.Metrics != nil {
atomic.AddInt64(&s.Metrics.NotificationsSent, 1)
}
}
w.Header().Set("Content-Type", "text/plain; charset=utf-8")
_, _ = w.Write([]byte(summary))
}
@ -720,13 +732,16 @@ func (s *Server) handleTriggerJSON(w http.ResponseWriter, r *http.Request) {
_, _ = w.Write([]byte("summarizer error"))
return
}
if push && s.Notifier != nil {
title := fmt.Sprintf("IRC digest %s (%s)", channel, dur)
_ = s.Notifier.Notify(r.Context(), title, sum)
if s.Metrics != nil {
atomic.AddInt64(&s.Metrics.NotificationsSent, 1)
}
}
if push && s.Notifier != nil {
title := fmt.Sprintf("IRC digest %s (%s)", channel, dur)
// Safety clamp for provider limits
msg := strings.TrimSpace(sum)
if len(msg) > 1024 { msg = msg[:1024] }
_ = s.Notifier.Notify(r.Context(), title, msg)
if s.Metrics != nil {
atomic.AddInt64(&s.Metrics.NotificationsSent, 1)
}
}
w.Header().Set("Content-Type", "application/json")
_ = json.NewEncoder(w).Encode(map[string]any{"summary": sum})
}

View file

@ -111,7 +111,8 @@ func (o *OpenAI) Summarize(ctx context.Context, channel string, msgs []store.Mes
b.WriteString("- Focus on what happened and why it matters.\n")
b.WriteString("- Integrate linked content and pasted multi-line posts naturally.\n")
b.WriteString("- Avoid rigid sections; use short paragraphs or light bullets if helpful.\n")
b.WriteString("- Keep it compact but dont omit important context.\n")
b.WriteString("- Keep it compact but dont omit important context.\n")
prompt := b.String()
sys := "You summarize IRC transcripts. Be concise, natural, and informative."
@ -162,6 +163,95 @@ func (o *OpenAI) Summarize(ctx context.Context, channel string, msgs []store.Mes
return out, nil
}
// SummarizeForPush produces a digest tailored for push notifications (e.g., Pushover ~1024 chars).
// It uses a slightly more constrained prompt to encourage succinct output.
func (o *OpenAI) SummarizeForPush(ctx context.Context, channel string, msgs []store.Message, window time.Duration) (string, error) {
if o == nil || o.apiKey == "" {
return "", nil
}
cfg := openai.DefaultConfig(o.apiKey)
if strings.TrimSpace(o.baseURL) != "" {
cfg.BaseURL = o.baseURL
}
client := openai.NewClientWithConfig(cfg)
grouped := groupMessages(msgs, o.groupWindow)
if o.maxGroups > 0 && len(grouped) > o.maxGroups {
grouped = grouped[len(grouped)-o.maxGroups:]
}
links := extractLinks(grouped)
var imageURLs []string
var nonImageLinks []linkSnippet
for _, l := range links {
if isImageURL(l.url) { imageURLs = append(imageURLs, l.url) } else { nonImageLinks = append(nonImageLinks, l) }
}
if o.followLinks && len(nonImageLinks) > 0 {
nonImageLinks = fetchLinkSnippets(ctx, nonImageLinks, o.linkTimeout, o.linkMaxBytes, o.maxLinks)
}
var b strings.Builder
b.WriteString("Channel: ")
b.WriteString(channel)
b.WriteString("\nTime window: ")
b.WriteString(window.String())
b.WriteString("\n\nTranscript (grouped by author):\n")
for _, g := range grouped {
b.WriteString(g.time.Format(time.RFC3339))
b.WriteString(" ")
b.WriteString(g.author)
b.WriteString(": ")
b.WriteString(g.text)
b.WriteString("\n")
}
if len(nonImageLinks) > 0 {
b.WriteString("\nReferenced content (snippets):\n")
for _, ln := range nonImageLinks {
b.WriteString("- ")
b.WriteString(ln.url)
b.WriteString(" → ")
b.WriteString(ln.snippet)
b.WriteString("\n")
}
}
b.WriteString("\nWrite a concise, readable summary of the conversation above.\n")
b.WriteString("- Focus on what happened and why it matters.\n")
b.WriteString("- Integrate linked content and pasted multi-line posts naturally.\n")
b.WriteString("- Avoid rigid sections; use short paragraphs or light bullets if helpful.\n")
b.WriteString("- Keep it compact but dont omit important context.\n")
b.WriteString("- Keep the final output under ~900 characters suitable for a single push notification.\n")
prompt := b.String()
sys := "You summarize IRC transcripts for a push notification. Be concise, natural, and informative."
model := o.model
if strings.TrimSpace(model) == "" { model = "gpt-4o-mini" }
reasoningLike := strings.HasPrefix(model, "gpt-5") || strings.HasPrefix(model, "o1") || strings.Contains(model, "reasoning")
var userParts []openai.ChatMessagePart
userParts = append(userParts, openai.ChatMessagePart{Type: openai.ChatMessagePartTypeText, Text: prompt})
for _, u := range imageURLs {
userParts = append(userParts, openai.ChatMessagePart{Type: openai.ChatMessagePartTypeImageURL, ImageURL: &openai.ChatMessageImageURL{URL: u}})
}
req := openai.ChatCompletionRequest{
Model: model,
Messages: []openai.ChatCompletionMessage{
{Role: openai.ChatMessageRoleSystem, Content: sys},
{Role: openai.ChatMessageRoleUser, MultiContent: userParts},
},
MaxCompletionTokens: o.maxTokens,
}
if !reasoningLike { req.Temperature = 0.3 }
resp, err := client.CreateChatCompletion(ctx, req)
if err != nil { return "", err }
if len(resp.Choices) == 0 { return localFallbackSummary(grouped, append(nonImageLinks, linksFromImages(imageURLs)...)), nil }
out := strings.TrimSpace(resp.Choices[0].Message.Content)
if out == "" { return localFallbackSummary(grouped, append(nonImageLinks, linksFromImages(imageURLs)...)), nil }
return out, nil
}
func (o *OpenAI) SummarizeLink(ctx context.Context, rawURL string) (string, error) {
if o == nil || o.apiKey == "" {
return "", nil
@ -274,7 +364,7 @@ func (o *OpenAI) SummarizeLink(ctx context.Context, rawURL string) (string, erro
b.WriteString(content)
b.WriteString("\n\n")
}
b.WriteString("Write a short, skimmable summary of the page/video/image above. If relevant, include key takeaways and any notable cautions. Keep it under a few short paragraphs.")
b.WriteString("Write a short, skimmable summary of the page/video/image above. If relevant, include key takeaways and any notable cautions. Keep it under a few short paragraphs.")
userParts = append(userParts, openai.ChatMessagePart{Type: openai.ChatMessagePartTypeText, Text: b.String()})
if img != "" {
userParts = append(userParts, openai.ChatMessagePart{Type: openai.ChatMessagePartTypeImageURL, ImageURL: &openai.ChatMessageImageURL{URL: img}})

View file

@ -8,6 +8,9 @@ import (
)
type Summarizer interface {
Summarize(ctx context.Context, channel string, msgs []store.Message, window time.Duration) (string, error)
SummarizeLink(ctx context.Context, rawURL string) (string, error)
Summarize(ctx context.Context, channel string, msgs []store.Message, window time.Duration) (string, error)
// SummarizeForPush creates a digest tuned for push notifications (e.g., Pushover limits).
// Implementations should keep the output succinct and within ~1k characters.
SummarizeForPush(ctx context.Context, channel string, msgs []store.Message, window time.Duration) (string, error)
SummarizeLink(ctx context.Context, rawURL string) (string, error)
}