1package core
2
3import (
4 "context"
5 "fmt"
6 "strconv"
7 "strings"
8 "time"
9
10 "github.com/patriceckhart/zot/packages/provider"
11)
12
13// Compact summarizes the agent's transcript via the LLM and replaces
14// it with a single synthetic user message carrying the summary. A
15// small tail of recent messages is optionally preserved for continuity.
16//
17// keepTail is the number of most-recent messages to keep verbatim after
18// the summary. 0 means summarize everything; a typical useful value is
19// 4-8 (last couple of exchanges).
20//
21// The method blocks until the summary request completes. Emitted
22// events via sink are limited to text deltas from the summary call so
23// the UI can show progress.
24func (a *Agent) Compact(ctx context.Context, keepTail int, sink func(delta string)) (summary string, err error) {
25 a.mu.Lock()
26 msgs := append([]provider.Message(nil), a.messages...)
27 a.mu.Unlock()
28
29 if len(msgs) == 0 {
30 return "", fmt.Errorf("nothing to compact")
31 }
32 if keepTail < 0 {
33 keepTail = 0
34 }
35 if keepTail > len(msgs) {
36 keepTail = len(msgs)
37 }
38 summarizable := msgs[:len(msgs)-keepTail]
39 if len(summarizable) == 0 {
40 return "", fmt.Errorf("nothing to compact: keep-tail covers the whole transcript")
41 }
42
43 // Serialize the summarizable transcript to text and wrap it in tags
44 // so the model treats it as material to summarize, not to continue.
45 transcript := serializeTranscript(summarizable)
46
47 prompt := "<conversation>\n" + transcript + "\n</conversation>\n\n" + compactionPrompt
48
49 req := provider.Request{
50 Model: a.Model,
51 System: summarizationSystem,
52 MaxTokens: 4096,
53 Temperature: a.Temperature,
54 Messages: []provider.Message{
55 {
56 Role: provider.RoleUser,
57 Content: []provider.Content{provider.TextBlock{Text: prompt}},
58 Time: time.Now(),
59 },
60 },
61 }
62
63 stream, err := a.Client.Stream(ctx, req)
64 if err != nil {
65 return "", err
66 }
67
68 var sb strings.Builder
69 for ev := range stream {
70 switch e := ev.(type) {
71 case provider.EventTextDelta:
72 sb.WriteString(e.Delta)
73 if sink != nil {
74 sink(e.Delta)
75 }
76 case provider.EventDone:
77 if e.Err != nil {
78 return "", e.Err
79 }
80 }
81 }
82 summary = strings.TrimSpace(sb.String())
83 if summary == "" {
84 return "", fmt.Errorf("empty summary from model")
85 }
86
87 // Estimate token count before compaction (rough: 1 token ~ 4 chars).
88 tokensBefore := len(transcript) / 4
89
90 // Replace transcript: one synthetic user message with the summary,
91 // followed by the preserved tail (if any).
92 synthetic := provider.Message{
93 Role: provider.RoleUser,
94 Content: []provider.Content{
95 provider.TextBlock{Text: "## Context Summary (compacted)\n\n" + summary},
96 },
97 Time: time.Now(),
98 Meta: map[string]string{
99 "compaction": "true",
100 "tokens_before": strconv.Itoa(tokensBefore),
101 },
102 }
103
104 tail := msgs[len(msgs)-keepTail:]
105 // Repair the tail: remove orphaned tool_result blocks whose
106 // matching tool_use was in the compacted (now-removed) portion.
107 // Anthropic rejects transcripts where a tool_result references
108 // a tool_use ID that doesn't exist.
109 tail = repairOrphanedToolResults(tail)
110
111 next := make([]provider.Message, 0, 1+len(tail))
112 next = append(next, synthetic)
113 next = append(next, tail...)
114
115 a.mu.Lock()
116 a.messages = next
117 a.rev++
118 onCompacted := a.OnTranscriptCompacted
119 persisted := append([]provider.Message(nil), next...)
120 a.mu.Unlock()
121
122 if onCompacted != nil {
123 onCompacted(persisted)
124 }
125
126 return summary, nil
127}
128
129// repairOrphanedToolResults removes tool_result content blocks (and
130// entire messages that become empty) when the matching tool_use ID
131// does not appear anywhere in the given messages. This happens after
132// compaction when the tail preserves a tool_result but the tool_use
133// that produced it was summarized away.
134func repairOrphanedToolResults(msgs []provider.Message) []provider.Message {
135 return provider.RepairOrphanedToolResults(msgs)
136}
137
138// serializeTranscript renders a list of provider.Message into a plain
139// text transcript the summarization model can read without trying to
140// continue the conversation.
141func serializeTranscript(msgs []provider.Message) string {
142 var sb strings.Builder
143 for _, m := range msgs {
144 switch m.Role {
145 case provider.RoleUser:
146 sb.WriteString("\n--- user ---\n")
147 case provider.RoleAssistant:
148 sb.WriteString("\n--- assistant ---\n")
149 case provider.RoleTool:
150 sb.WriteString("\n--- tool ---\n")
151 }
152 for _, c := range m.Content {
153 switch v := c.(type) {
154 case provider.TextBlock:
155 sb.WriteString(v.Text)
156 sb.WriteString("\n")
157 case provider.ImageBlock:
158 fmt.Fprintf(&sb, "[image: %s, %d bytes]\n", v.MimeType, len(v.Data))
159 case provider.ToolCallBlock:
160 fmt.Fprintf(&sb, "[tool_call %s %s]\n", v.Name, string(v.Arguments))
161 case provider.ToolResultBlock:
162 for _, inner := range v.Content {
163 if tb, ok := inner.(provider.TextBlock); ok {
164 sb.WriteString("[tool_result] ")
165 sb.WriteString(tb.Text)
166 sb.WriteString("\n")
167 }
168 }
169 }
170 }
171 }
172 return sb.String()
173}
174
175const summarizationSystem = `You are a context summarization assistant. Your task is to read a conversation between a user and an AI coding assistant, then produce a structured summary following the exact format specified.
176
177Do NOT continue the conversation. Do NOT respond to any questions in the conversation. ONLY output the structured summary.`
178
179const compactionPrompt = `The messages above are a conversation to summarize. Create a structured context checkpoint summary that another LLM will use to continue the work.
180
181Use this EXACT format:
182
183## Goal
184[What is the user trying to accomplish? Can be multiple items if the session covers different tasks.]
185
186## Constraints & Preferences
187- [Any constraints, preferences, or requirements mentioned by user]
188- [Or "(none)" if none were mentioned]
189
190## Progress
191### Done
192- [x] [Completed tasks/changes]
193
194### In Progress
195- [ ] [Current work]
196
197### Blocked
198- [Issues preventing progress, if any]
199
200## Key Decisions
201- **[Decision]**: [Brief rationale]
202
203## Next Steps
2041. [Ordered list of what should happen next]
205
206## Critical Context
207- [Any data, examples, or references needed to continue]
208- [Or "(none)" if not applicable]
209
210Keep each section concise. Preserve exact file paths, function names, and error messages.`