| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "strings" |
| 6 | "testing" |
| 7 | |
| 8 | "reasonix/internal/agent/testutil" |
| 9 | "reasonix/internal/event" |
| 10 | "reasonix/internal/provider" |
| 11 | "reasonix/internal/tool" |
| 12 | ) |
| 13 | |
| 14 | func TestMaybeCompactIgnoresRetryAggregateBelowContextThreshold(t *testing.T) { |
| 15 | // compact_ratio is the sole trigger; LatestPromptTokens (context attempt) |
| 16 | // must drive the decision, not the billable retry aggregate. |
| 17 | const window = 10_000 |
| 18 | sess := &Session{Messages: []provider.Message{ |
| 19 | {Role: provider.RoleSystem, Content: "sys"}, |
| 20 | {Role: provider.RoleUser, Content: "task"}, |
| 21 | {Role: provider.RoleAssistant, Content: strings.Repeat("large earlier request ", 500)}, |
| 22 | {Role: provider.RoleUser, Content: "current request"}, |
| 23 | {Role: provider.RoleAssistant, Content: "current answer"}, |
| 24 | }} |
| 25 | a := New(&fakeProvider{reply: "summary"}, tool.NewRegistry(), sess, Options{ |
| 26 | ContextWindow: window, |
| 27 | CompactRatio: 0.85, |
| 28 | RecentKeep: 2, |
| 29 | }, event.Discard) |
| 30 | |
| 31 | // Aggregate 9000 would be above fold (8500); latest context attempt 4000 is not. |
| 32 | prepareForObservedUsage(a, context.Background(), &provider.Usage{ |
| 33 | PromptTokens: 9000, |
| 34 | ContextPromptTokens: 4000, |
| 35 | }) |
| 36 | |
| 37 | if hasCompactionSummary(visibleContext(a)) { |
| 38 | t.Fatal("retry-aggregate prompt tokens triggered compaction below the latest context threshold") |
| 39 | } |
| 40 | } |
| 41 | |
| 42 | func TestMaybeCompactStillTriggersAtLatestContextThreshold(t *testing.T) { |
| 43 | const window = 10_000 |
| 44 | sess := &Session{Messages: []provider.Message{ |
| 45 | {Role: provider.RoleSystem, Content: "sys"}, |
| 46 | {Role: provider.RoleUser, Content: "task"}, |
| 47 | {Role: provider.RoleAssistant, Content: strings.Repeat("large earlier request ", 500)}, |
| 48 | {Role: provider.RoleUser, Content: "current request"}, |
| 49 | {Role: provider.RoleAssistant, Content: "current answer"}, |
| 50 | }} |
| 51 | a := New(&fakeProvider{reply: "summary"}, tool.NewRegistry(), sess, Options{ |
| 52 | ContextWindow: window, |
| 53 | CompactRatio: 0.85, |
| 54 | RecentKeep: 2, |
| 55 | }, event.Discard) |
| 56 | |
| 57 | // Latest context attempt at/above fold trigger (8500) must install a summary. |
| 58 | prepareForObservedUsage(a, context.Background(), &provider.Usage{ |
| 59 | PromptTokens: 16000, |
| 60 | ContextPromptTokens: 8600, |
| 61 | }) |
| 62 | |
| 63 | if !hasCompactionSummary(visibleContext(a)) { |
| 64 | t.Fatal("latest prompt at the configured threshold did not trigger compaction") |
| 65 | } |
| 66 | } |
| 67 | |
| 68 | func TestMissingReasoningRetryAggregateDoesNotTriggerEarlyCompaction(t *testing.T) { |
| 69 | mp := testutil.NewMock("deepseek-proxy", |
| 70 | testutil.Turn{ |
| 71 | ToolCalls: []provider.ToolCall{{ID: "c1", Name: "echo", Arguments: `{"text":"hi"}`}}, |
| 72 | Usage: &provider.Usage{PromptTokens: 4000, CompletionTokens: 2, TotalTokens: 4002, CacheMissTokens: 4000, FinishReason: "tool_calls"}, |
| 73 | }, |
| 74 | testutil.Turn{ |
| 75 | Reasoning: "retry reasoning", |
| 76 | ToolCalls: []provider.ToolCall{{ID: "c1", Name: "echo", Arguments: `{"text":"hi"}`}}, |
| 77 | Usage: &provider.Usage{PromptTokens: 4000, CompletionTokens: 3, TotalTokens: 4003, CacheHitTokens: 4000, ReasoningTokens: 2, FinishReason: "tool_calls"}, |
| 78 | }, |
| 79 | testutil.Turn{ |
| 80 | Text: "done", |
| 81 | Usage: &provider.Usage{PromptTokens: 4100, CompletionTokens: 1, TotalTokens: 4101, CacheHitTokens: 4000, CacheMissTokens: 100}, |
| 82 | }, |
| 83 | ) |
| 84 | sess := &Session{Messages: []provider.Message{ |
| 85 | {Role: provider.RoleSystem, Content: "sys"}, |
| 86 | {Role: provider.RoleUser, Content: "original task"}, |
| 87 | {Role: provider.RoleAssistant, Content: strings.Repeat("old analysis ", 100)}, |
| 88 | {Role: provider.RoleUser, Content: "follow-up"}, |
| 89 | {Role: provider.RoleAssistant, Content: strings.Repeat("more old work ", 100)}, |
| 90 | }} |
| 91 | sink := &recordSink{} |
| 92 | a := New(strictToolCallReasoningProvider{mp}, echoRegistry(), sess, Options{ |
| 93 | ContextWindow: 10_000, |
| 94 | ArchiveDir: t.TempDir(), |
| 95 | }, sink) |
| 96 | |
| 97 | if err := a.Run(context.Background(), "go"); err != nil { |
| 98 | t.Fatalf("Run: %v", err) |
| 99 | } |
| 100 | if got := len(sink.kinds(event.CompactionStarted)); got != 0 { |
| 101 | t.Fatalf("compactions = %d, want none while latest prompt is below threshold", got) |
| 102 | } |
| 103 | usageEvents := sink.kinds(event.Usage) |
| 104 | if len(usageEvents) == 0 || usageEvents[0].Usage == nil { |
| 105 | t.Fatalf("missing recovery usage event: %+v", usageEvents) |
| 106 | } |
| 107 | if got := usageEvents[0].Usage; got.PromptTokens != 8000 || got.ContextPromptTokens != 4000 { |
| 108 | t.Fatalf("recovery usage = %+v, want aggregate prompt 8000 and latest context prompt 4000", got) |
| 109 | } |
| 110 | } |
| 111 |