Files
Memoh/internal/agent/sential.go
T
Acbox Liu 1680316c7f refactor(agent): remove agent gateway instead of twilight sdk (#264)
* refactor(agent): replace TypeScript agent gateway with in-process Go agent using twilight-ai SDK

- Remove apps/agent (Bun/Elysia gateway), packages/agent (@memoh/agent),
  internal/bun runtime manager, and all embedded agent/bun assets
- Add internal/agent package powered by twilight-ai SDK for LLM calls,
  tool execution, streaming, sential logic, tag extraction, and prompts
- Integrate ToolGatewayService in-process for both built-in and user MCP
  tools, eliminating HTTP round-trips to the old gateway
- Update resolver to convert between sdk.Message and ModelMessage at the
  boundary (resolver_messages.go), keeping agent package free of
  persistence concerns
- Prepend user message before storeRound since SDK only returns output
  messages (assistant + tool)
- Clean up all Docker configs, TOML configs, nginx proxy, Dockerfile.agent,
  and Go config structs related to the removed agent gateway
- Update cmd/agent and cmd/memoh entry points with setter-based
  ToolGateway injection to avoid FX dependency cycles

* fix(web): move form declaration before computed properties that reference it

The `form` reactive object was declared after computed properties like
`selectedMemoryProvider` and `isSelectedMemoryProviderPersisted` that
reference it, causing a TDZ ReferenceError during setup.

* fix: prevent UTF-8 character corruption in streaming text output

StreamTagExtractor.Push() used byte-level string slicing to hold back
buffer tails for tag detection, which could split multi-byte UTF-8
characters. After json.Marshal replaced invalid bytes with U+FFFD,
the corruption became permanent — causing garbled CJK characters (�)
in agent responses.

Add safeUTF8SplitIndex() to back up split points to valid character
boundaries. Also fix byte-level truncation in command/formatter.go
and command/fs.go to use rune-aware slicing.

* fix: add agent error logging and fix Gemini tool schema validation

- Log agent stream errors in both SSE and WebSocket paths with bot/model context
- Fix send tool `attachments` parameter: empty `items` schema rejected by
  Google Gemini API (INVALID_ARGUMENT), now specifies `{"type": "string"}`
- Upgrade twilight-ai to d898f0b (includes raw body in API error messages)

* chore(ci): remove agent gateway from Docker build and release pipelines

Agent gateway has been replaced by in-process Go agent; remove the
obsolete Docker image matrix entry, Bun/UPX CI steps, and agent-binary
build logic from the release script.

* fix: preserve attachment filename, metadata, and container path through persistence

- Add `name` column to `bot_history_message_assets` (migration 0034) to
  persist original filenames across page refreshes.
- Add `metadata` JSONB column (migration 0035) to store source_path,
  source_url, and other context alongside each asset.
- Update SQL queries, sqlc-generated code, and all Go types (MessageAsset,
  AssetRef, OutboundAssetRef, FileAttachment) to carry name and metadata
  through the full lifecycle.
- Extract filenames from path/URL in AttachmentsResolver before clearing
  raw paths; enrich streaming event metadata with name, source_path, and
  source_url in both the WebSocket and channel inbound ingestion paths.
- Implement `LinkAssets` on message service and `LinkOutboundAssets` on
  flow resolver so WebSocket-streamed bot attachments are persisted to the
  correct assistant message after streaming completes.
- Frontend: update MessageAsset type with metadata field, pass metadata
  through to attachment items, and reorder attachment-block.vue template
  so container files (identified by metadata.source_path) open in the
  sidebar file manager instead of triggering a download.

* refactor(agent): decouple built-in tools from MCP, load via ToolProvider interface

Migrate all 13 built-in tool providers from internal/mcp/providers/ to
internal/agent/tools/ using the twilight-ai sdk.Tool structure. The agent
now loads tools through a ToolProvider interface instead of the MCP
ToolGatewayService, which is simplified to only manage external federation
sources. This enables selective tool loading and removes the coupling
between business tools and the MCP protocol layer.

* refactor(flow): split monolithic resolver.go into focused modules

Break the 1959-line resolver.go into 12 files organized by concern:
- resolver.go: core orchestration (Resolver struct, resolve, Chat, prepareRunConfig)
- resolver_stream.go: streaming (StreamChat, StreamChatWS, tryStoreStream)
- resolver_trigger.go: schedule/heartbeat triggers
- resolver_attachments.go: attachment routing, inlining, encoding
- resolver_history.go: message loading, deduplication, token trimming
- resolver_store.go: persistence (storeRound, storeMessages, asset linking)
- resolver_memory.go: memory provider integration
- resolver_model_selection.go: model selection and candidate matching
- resolver_identity.go: display name and channel identity resolution
- resolver_settings.go: bot settings, loop detection, inbox
- user_header.go: YAML front-matter formatting
- resolver_util.go: shared utilities (sanitize, normalize, dedup, UUID)

* fix(agent): enable Anthropic extended thinking by passing ReasoningConfig to provider

Anthropic's thinking requires WithThinking() at provider creation time,
unlike OpenAI which uses per-request ReasoningEffort. The config was
never wired through, so Claude models could not trigger thinking.

* refactor(agent): extract prompts into embedded markdown templates

Move inline prompt strings from prompt.go into separate .md files under
internal/agent/prompts/, using {{key}} placeholders and a simple render
engine. Remove obsolete SystemPromptParams fields (Language,
MaxContextLoadTime, Channels, CurrentChannel) and their call-site usage.

* fix: lint
2026-03-19 13:31:54 +08:00

488 lines
12 KiB
Go

package agent
import (
"crypto/sha256"
"encoding/hex"
"encoding/json"
"fmt"
"sort"
"strings"
"unicode"
"unicode/utf8"
)
// Loop detection constants matching the TypeScript implementation.
const (
LoopDetectedAbortMessage = "loop detected, stream aborted"
LoopDetectedStreakThreshold = 3
LoopDetectedMinNewGramsPerChunk = 8
LoopDetectedProbeChars = 256
ToolLoopDetectedAbortMessage = "tool loop detected, stream aborted"
ToolLoopRepeatThreshold = 5
ToolLoopWarningsBeforeAbort = 1
ToolLoopWarningKey = "__memoh_tool_loop_warning" //nolint:gosec // internal warning key, not a credential
ToolLoopWarningText = "[MEMOH_TOOL_LOOP_WARNING] Repeated identical tool invocation (same tool + arguments) was detected more than 5 times. Stop looping this tool and either summarize current results or change strategy." //nolint:gosec // human-readable warning text, not a credential
defaultNgramSize = 10
defaultWindowSize = 1000
defaultOverlapThreshold = 0.75
defaultConsecutiveHits = 10
defaultMinNewGramsPerChunk = 1
)
// --- Sential: n-gram overlap detector ---
// SentialOptions configures the n-gram overlap detector.
type SentialOptions struct {
NgramSize int
WindowSize int
OverlapThreshold float64
}
// SentialResult is the output of an overlap inspection.
type SentialResult struct {
Hit bool
Overlap float64
MatchedGrams int
NewGrams int
}
// Sential detects text repetition via n-gram overlap.
type Sential struct {
ngramSize int
windowSize int
overlapThreshold float64
windowChars []rune
windowNgramQueue []string
historySet map[string]struct{}
historyCounts map[string]int
}
// NewSential creates a new n-gram overlap detector.
func NewSential(opts SentialOptions) *Sential {
ngramSize := opts.NgramSize
if ngramSize <= 0 {
ngramSize = defaultNgramSize
}
windowSize := opts.WindowSize
if windowSize <= 0 {
windowSize = defaultWindowSize
}
overlapThreshold := opts.OverlapThreshold
if overlapThreshold <= 0 {
overlapThreshold = defaultOverlapThreshold
}
return &Sential{
ngramSize: ngramSize,
windowSize: windowSize,
overlapThreshold: overlapThreshold,
historySet: make(map[string]struct{}),
historyCounts: make(map[string]int),
}
}
func (s *Sential) addHistoryGram(gram string) {
s.historyCounts[gram]++
if s.historyCounts[gram] == 1 {
s.historySet[gram] = struct{}{}
}
}
func (s *Sential) removeHistoryGram(gram string) {
count := s.historyCounts[gram]
if count <= 1 {
delete(s.historyCounts, gram)
delete(s.historySet, gram)
return
}
s.historyCounts[gram] = count - 1
}
func (s *Sential) pushWindowChar(ch rune) {
s.windowChars = append(s.windowChars, ch)
if len(s.windowChars) >= s.ngramSize {
start := len(s.windowChars) - s.ngramSize
gram := string(s.windowChars[start : start+s.ngramSize])
s.windowNgramQueue = append(s.windowNgramQueue, gram)
s.addHistoryGram(gram)
}
if len(s.windowChars) <= s.windowSize {
return
}
s.windowChars = s.windowChars[1:]
if len(s.windowNgramQueue) > 0 {
removed := s.windowNgramQueue[0]
s.windowNgramQueue = s.windowNgramQueue[1:]
s.removeHistoryGram(removed)
}
}
// Inspect checks a chunk of text for n-gram overlap with the sliding window.
func (s *Sential) Inspect(text string) SentialResult {
incoming := []rune(text)
if len(incoming) == 0 {
return SentialResult{}
}
contextSize := s.ngramSize - 1
if contextSize < 0 {
contextSize = 0
}
var contextChars []rune
if contextSize > 0 && len(s.windowChars) > 0 {
start := len(s.windowChars) - contextSize
if start < 0 {
start = 0
}
contextChars = make([]rune, len(s.windowChars[start:]))
copy(contextChars, s.windowChars[start:])
}
candidate := append([]rune{}, contextChars...)
candidate = append(candidate, incoming...)
contextLength := len(contextChars)
matchedGrams := 0
newGrams := 0
if len(candidate) >= s.ngramSize {
for i := 0; i <= len(candidate)-s.ngramSize; i++ {
gramEndIndex := i + s.ngramSize - 1
if gramEndIndex < contextLength {
continue
}
gram := string(candidate[i : i+s.ngramSize])
newGrams++
if _, ok := s.historySet[gram]; ok {
matchedGrams++
}
}
}
overlap := 0.0
if newGrams > 0 {
overlap = float64(matchedGrams) / float64(newGrams)
}
hit := overlap > s.overlapThreshold
for _, ch := range incoming {
s.pushWindowChar(ch)
}
return SentialResult{
Hit: hit,
Overlap: overlap,
MatchedGrams: matchedGrams,
NewGrams: newGrams,
}
}
// Reset clears the detector state.
func (s *Sential) Reset() {
s.windowChars = nil
s.windowNgramQueue = nil
s.historySet = make(map[string]struct{})
s.historyCounts = make(map[string]int)
}
// --- Text Loop Guard ---
// TextLoopGuardResult extends SentialResult with streak and abort tracking.
type TextLoopGuardResult struct {
SentialResult
Streak int
Abort bool
}
// TextLoopGuard wraps Sential with consecutive-hit tracking.
type TextLoopGuard struct {
sential *Sential
consecutiveHitsToAbort int
minNewGramsPerChunk int
streak int
}
// NewTextLoopGuard creates a text loop guard.
func NewTextLoopGuard(consecutiveHits, minNewGrams int, opts SentialOptions) *TextLoopGuard {
if consecutiveHits <= 0 {
consecutiveHits = defaultConsecutiveHits
}
if minNewGrams <= 0 {
minNewGrams = defaultMinNewGramsPerChunk
}
return &TextLoopGuard{
sential: NewSential(opts),
consecutiveHitsToAbort: consecutiveHits,
minNewGramsPerChunk: minNewGrams,
}
}
// Inspect checks text and tracks consecutive overlap streaks.
func (g *TextLoopGuard) Inspect(text string) TextLoopGuardResult {
result := g.sential.Inspect(text)
if result.NewGrams >= g.minNewGramsPerChunk {
if result.Hit {
g.streak++
} else {
g.streak = 0
}
}
return TextLoopGuardResult{
SentialResult: result,
Streak: g.streak,
Abort: g.streak >= g.consecutiveHitsToAbort,
}
}
// Reset clears the guard state.
func (g *TextLoopGuard) Reset() {
g.sential.Reset()
g.streak = 0
}
// --- Text Loop Probe Buffer ---
// TextLoopProbeBuffer batches text into chunks before passing to an inspector.
type TextLoopProbeBuffer struct {
chunkSize int
inspect func(string)
chars []rune
offset int
}
// NewTextLoopProbeBuffer creates a probe buffer.
func NewTextLoopProbeBuffer(chunkSize int, inspect func(string)) *TextLoopProbeBuffer {
if chunkSize <= 0 {
chunkSize = LoopDetectedProbeChars
}
return &TextLoopProbeBuffer{
chunkSize: chunkSize,
inspect: inspect,
}
}
// Push adds text to the buffer, emitting full chunks to the inspector.
func (b *TextLoopProbeBuffer) Push(text string) {
if text == "" {
return
}
b.chars = append(b.chars, []rune(text)...)
for len(b.chars)-b.offset >= b.chunkSize {
chunk := string(b.chars[b.offset : b.offset+b.chunkSize])
b.offset += b.chunkSize
if len(chunk) > 0 {
b.inspect(chunk)
}
}
if b.offset >= b.chunkSize {
b.chars = b.chars[b.offset:]
b.offset = 0
}
}
// Flush emits any remaining content to the inspector.
func (b *TextLoopProbeBuffer) Flush() {
if len(b.chars)-b.offset > 0 {
remainder := string(b.chars[b.offset:])
if len(remainder) > 0 {
b.inspect(remainder)
}
}
b.chars = nil
b.offset = 0
}
// --- Tool Loop Guard ---
var defaultVolatileKeys = []string{
"toolcallid", "toolcallid", "requestid", "requestid",
"traceid", "traceid", "spanid", "spanid",
"sessionid", "sessionid", "timestamp",
"createdat", "createdat", "updatedat", "updatedat",
"expiresat", "expiresat", "nonce",
}
var volatileKeySuffixes = []string{
"requestid", "traceid", "sessionid", "toolcallid",
"timestamp", "createdat", "updatedat", "expiresat",
}
// ToolLoopInput represents a tool call for loop detection.
type ToolLoopInput struct {
ToolName string
Input any
}
// ToolLoopResult is the output of a tool loop inspection.
type ToolLoopResult struct {
Hash string
RepeatCount int
BreachCount int
Warn bool
Abort bool
}
// ToolLoopGuard detects repeated identical tool calls.
type ToolLoopGuard struct {
repeatThreshold int
warningsBeforeAbort int
volatileKeySet map[string]struct{}
lastHash string
repeatCount int
breachCount int
breachHash string
}
// NewToolLoopGuard creates a tool loop guard.
func NewToolLoopGuard(repeatThreshold, warningsBeforeAbort int) *ToolLoopGuard {
if repeatThreshold <= 0 {
repeatThreshold = ToolLoopRepeatThreshold
}
if warningsBeforeAbort <= 0 {
warningsBeforeAbort = ToolLoopWarningsBeforeAbort
}
volatileSet := make(map[string]struct{})
for _, k := range defaultVolatileKeys {
volatileSet[normalizeKeyName(k)] = struct{}{}
}
return &ToolLoopGuard{
repeatThreshold: repeatThreshold,
warningsBeforeAbort: warningsBeforeAbort,
volatileKeySet: volatileSet,
}
}
// Inspect checks a tool call for repetition.
func (g *ToolLoopGuard) Inspect(input ToolLoopInput) ToolLoopResult {
hash := computeToolLoopHash(input, g.volatileKeySet)
if hash == g.lastHash {
g.repeatCount++
} else {
g.lastHash = hash
g.repeatCount = 1
}
if g.breachHash != hash {
g.breachHash = hash
g.breachCount = 0
}
warn := false
abort := false
if g.repeatCount > g.repeatThreshold {
if g.breachCount < g.warningsBeforeAbort {
g.breachCount++
warn = true
g.lastHash = ""
g.repeatCount = 0
} else {
g.breachCount++
abort = true
}
}
return ToolLoopResult{
Hash: hash,
RepeatCount: g.repeatCount,
BreachCount: g.breachCount,
Warn: warn,
Abort: abort,
}
}
// Reset clears the guard state.
func (g *ToolLoopGuard) Reset() {
g.lastHash = ""
g.repeatCount = 0
g.breachCount = 0
g.breachHash = ""
}
func normalizeKeyName(key string) string {
var b strings.Builder
for _, r := range strings.TrimSpace(strings.ToLower(key)) {
if unicode.IsLetter(r) || unicode.IsDigit(r) {
b.WriteRune(r)
}
}
return b.String()
}
func isVolatileKey(key string, volatileSet map[string]struct{}) bool {
normalized := normalizeKeyName(key)
if normalized == "" {
return false
}
if _, ok := volatileSet[normalized]; ok {
return true
}
for _, suffix := range volatileKeySuffixes {
if strings.HasSuffix(normalized, suffix) {
return true
}
}
return false
}
func normalizeToolLoopValue(value any, volatileSet map[string]struct{}) any {
if value == nil {
return nil
}
switch v := value.(type) {
case string:
return v
case bool:
return v
case float64:
return v
case json.Number:
return v.String()
case []any:
result := make([]any, len(v))
for i, item := range v {
result[i] = normalizeToolLoopValue(item, volatileSet)
}
return result
case map[string]any:
keys := make([]string, 0, len(v))
for k := range v {
keys = append(keys, k)
}
sort.Strings(keys)
result := make(map[string]any, len(v))
for _, k := range keys {
if isVolatileKey(k, volatileSet) {
continue
}
normalized := normalizeToolLoopValue(v[k], volatileSet)
if normalized != nil {
result[k] = normalized
}
}
return result
default:
return fmt.Sprintf("%v", v)
}
}
func computeToolLoopHash(input ToolLoopInput, volatileSet map[string]struct{}) string {
payload := map[string]any{
"toolName": strings.TrimSpace(input.ToolName),
"input": normalizeToolLoopValue(input.Input, volatileSet),
}
serialized, _ := json.Marshal(payload)
h := sha256.Sum256(serialized)
return hex.EncodeToString(h[:])
}
// --- Helper to check for repetitions in text ---
func isNonEmptyString(s string) bool {
return utf8.RuneCountInString(strings.TrimSpace(s)) > 0
}
// Guard wraps tools with tool loop detection. Returns a wrapper execute function.
func (g *ToolLoopGuard) Guard(toolName string, input any) (warn bool, abort bool) {
result := g.Inspect(ToolLoopInput{ToolName: toolName, Input: input})
return result.Warn, result.Abort
}