| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "fmt" |
| 6 | "strings" |
| 7 | "testing" |
| 8 | |
| 9 | "reasonix/internal/provider" |
| 10 | ) |
| 11 | |
| 12 | // recordingSharedWindowProvider logs every request before delegating to the |
| 13 | // shared-window fake, so tests can assert each summary request stayed |
| 14 | // admissible. |
| 15 | type recordingSharedWindowProvider struct { |
| 16 | sharedWindowTestProvider |
| 17 | requests []provider.Request |
| 18 | } |
| 19 | |
| 20 | func (p *recordingSharedWindowProvider) Stream(ctx context.Context, req provider.Request) (<-chan provider.Chunk, error) { |
| 21 | p.requests = append(p.requests, req) |
| 22 | return p.sharedWindowTestProvider.Stream(ctx, req) |
| 23 | } |
| 24 | |
| 25 | // TestManualCompactEscapesOverCeilingDeadlock reproduces #9059: once the |
| 26 | // model-visible view sits at or above the hard input ceiling, a manual compact |
| 27 | // must still issue an admissible summary request and fold the view back under |
| 28 | // the ceiling instead of failing forever before the provider call. |
| 29 | func TestManualCompactEscapesOverCeilingDeadlock(t *testing.T) { |
| 30 | window := 50_000 |
| 31 | prov := &recordingSharedWindowProvider{sharedWindowTestProvider: sharedWindowTestProvider{budget: 128 * 1024, shared: true}} |
| 32 | sess := foldableSessionOverForce(6) |
| 33 | a := agentOverForceWindow(t, prov, sess, window) |
| 34 | |
| 35 | big := strings.Repeat("word ", 400) |
| 36 | for i := 0; i < 300 && a.estimatedVisibleRequestTokens(a.modelVisibleMessages()) < a.hardInputCeiling(); i++ { |
| 37 | sess.Add(provider.Message{Role: provider.RoleAssistant, Content: big}) |
| 38 | sess.Add(provider.Message{Role: provider.RoleUser, Content: "continue"}) |
| 39 | } |
| 40 | if est, hard := a.estimatedVisibleRequestTokens(a.modelVisibleMessages()), a.hardInputCeiling(); est < hard { |
| 41 | t.Fatalf("fixture did not reach hard ceiling: %d < %d", est, hard) |
| 42 | } |
| 43 | |
| 44 | if err := a.CompactNow(context.Background(), ""); err != nil { |
| 45 | t.Fatalf("manual compact over ceiling deadlocked: %v", err) |
| 46 | } |
| 47 | if est, hard := a.estimatedVisibleRequestTokens(a.modelVisibleMessages()), a.hardInputCeiling(); est >= hard { |
| 48 | t.Fatalf("manual compact left view at or above hard ceiling: %d >= %d", est, hard) |
| 49 | } |
| 50 | if len(prov.requests) == 0 { |
| 51 | t.Fatal("manual compact never reached the summarizer") |
| 52 | } |
| 53 | maxPrompt := window - a.summaryOutputBudget() - protocolReserveTokens |
| 54 | for i, req := range prov.requests { |
| 55 | if got := a.estimatedRequestTokens(req); got > maxPrompt { |
| 56 | t.Fatalf("summary request %d estimated %d tokens, exceeds admissible %d", i, got, maxPrompt) |
| 57 | } |
| 58 | } |
| 59 | summaries := 0 |
| 60 | for _, m := range a.modelVisibleMessages() { |
| 61 | if isCompactionSummary(m) { |
| 62 | summaries++ |
| 63 | } |
| 64 | } |
| 65 | if summaries != 1 { |
| 66 | t.Fatalf("projection summaries = %d, want exactly 1", summaries) |
| 67 | } |
| 68 | } |
| 69 | |
| 70 | // TestManualCompactOverCeilingPrunesBeforeSummary pins the rescue order: the |
| 71 | // free view-side prune runs before the first summarizer call, so oversized |
| 72 | // tool results in the never-folded tail shrink without a model round-trip. |
| 73 | func TestManualCompactOverCeilingPrunesBeforeSummary(t *testing.T) { |
| 74 | window := 50_000 |
| 75 | prov := &recordingSharedWindowProvider{sharedWindowTestProvider: sharedWindowTestProvider{budget: 128 * 1024, shared: true}} |
| 76 | msgs := []provider.Message{ |
| 77 | {Role: provider.RoleSystem, Content: "sys"}, |
| 78 | {Role: provider.RoleUser, Content: "dump everything"}, |
| 79 | } |
| 80 | for i := range 10 { |
| 81 | callID := "call-dump" |
| 82 | if i > 0 { |
| 83 | callID = fmt.Sprintf("call-dump-%d", i) |
| 84 | } |
| 85 | msgs = append(msgs, |
| 86 | provider.Message{Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{{ID: callID, Name: "dump", Arguments: `{}`}}}, |
| 87 | provider.Message{Role: provider.RoleTool, ToolCallID: callID, Name: "dump", Content: strings.Repeat("x", 24_000)}, |
| 88 | provider.Message{Role: provider.RoleUser, Content: "again"}, |
| 89 | ) |
| 90 | } |
| 91 | sess := &Session{Messages: msgs} |
| 92 | a := agentOverForceWindow(t, prov, sess, window) |
| 93 | if est, hard := a.estimatedVisibleRequestTokens(a.modelVisibleMessages()), a.hardInputCeiling(); est < hard { |
| 94 | t.Fatalf("fixture did not reach hard ceiling: %d < %d", est, hard) |
| 95 | } |
| 96 | |
| 97 | if err := a.CompactNow(context.Background(), ""); err != nil { |
| 98 | t.Fatalf("manual compact over tool-heavy ceiling: %v", err) |
| 99 | } |
| 100 | if len(prov.requests) == 0 { |
| 101 | t.Fatal("manual compact never reached the summarizer") |
| 102 | } |
| 103 | prunedSeen := false |
| 104 | for _, m := range prov.requests[0].Messages { |
| 105 | if m.Role != provider.RoleTool { |
| 106 | continue |
| 107 | } |
| 108 | if len(m.Content) < 24_000 && strings.Contains(m.Content, toolPruneMarker) { |
| 109 | prunedSeen = true |
| 110 | } |
| 111 | } |
| 112 | if !prunedSeen { |
| 113 | t.Fatal("first summary request carried unpruned tool results; prune did not run before the summarizer") |
| 114 | } |
| 115 | if est, hard := a.estimatedVisibleRequestTokens(a.modelVisibleMessages()), a.hardInputCeiling(); est >= hard { |
| 116 | t.Fatalf("manual compact left view at or above hard ceiling: %d >= %d", est, hard) |
| 117 | } |
| 118 | } |
| 119 | |
| 120 | // TestManualCompactMultiBatchWhenSingleFoldInsufficient drives a view several |
| 121 | // times the window: one admissible batch cannot free enough, so the rescue |
| 122 | // loop folds again with the prior digest merged into the next input. |
| 123 | func TestManualCompactMultiBatchWhenSingleFoldInsufficient(t *testing.T) { |
| 124 | window := 50_000 |
| 125 | prov := &recordingSharedWindowProvider{sharedWindowTestProvider: sharedWindowTestProvider{budget: 128 * 1024, shared: true}} |
| 126 | sess := foldableSessionOverForce(6) |
| 127 | a := agentOverForceWindow(t, prov, sess, window) |
| 128 | |
| 129 | big := strings.Repeat("word ", 400) |
| 130 | hard := a.hardInputCeiling() |
| 131 | for i := 0; i < 500 && a.estimatedVisibleRequestTokens(a.modelVisibleMessages()) < 2*hard+10_000; i++ { |
| 132 | sess.Add(provider.Message{Role: provider.RoleAssistant, Content: big}) |
| 133 | sess.Add(provider.Message{Role: provider.RoleUser, Content: "continue"}) |
| 134 | } |
| 135 | if est := a.estimatedVisibleRequestTokens(a.modelVisibleMessages()); est < 2*hard { |
| 136 | t.Fatalf("fixture did not reach double ceiling: %d < %d", est, 2*hard) |
| 137 | } |
| 138 | |
| 139 | if err := a.CompactNow(context.Background(), ""); err != nil { |
| 140 | t.Fatalf("multi-batch manual compact: %v", err) |
| 141 | } |
| 142 | if len(prov.requests) < 2 { |
| 143 | t.Fatalf("summary calls = %d, want at least 2 rescue batches", len(prov.requests)) |
| 144 | } |
| 145 | maxPrompt := window - a.summaryOutputBudget() - protocolReserveTokens |
| 146 | for i, req := range prov.requests { |
| 147 | if got := a.estimatedRequestTokens(req); got > maxPrompt { |
| 148 | t.Fatalf("summary request %d estimated %d tokens, exceeds admissible %d", i, got, maxPrompt) |
| 149 | } |
| 150 | } |
| 151 | var second strings.Builder |
| 152 | for _, m := range prov.requests[1].Messages { |
| 153 | second.WriteString(m.Content) |
| 154 | } |
| 155 | if !strings.Contains(second.String(), summaryTagOpen) { |
| 156 | t.Fatal("second rescue batch did not merge the prior digest into its input") |
| 157 | } |
| 158 | if est := a.estimatedVisibleRequestTokens(a.modelVisibleMessages()); est >= hard { |
| 159 | t.Fatalf("multi-batch manual compact left view over ceiling: %d >= %d", est, hard) |
| 160 | } |
| 161 | summaries := 0 |
| 162 | for _, m := range a.modelVisibleMessages() { |
| 163 | if isCompactionSummary(m) { |
| 164 | summaries++ |
| 165 | } |
| 166 | } |
| 167 | if summaries != 1 { |
| 168 | t.Fatalf("projection summaries = %d, want exactly 1", summaries) |
| 169 | } |
| 170 | } |
| 171 | |
| 172 | // TestManualCompactBelowCeilingKeepsSingleFold pins the ordinary manual |
| 173 | // contract: below the hard ceiling a manual compact still prunes nothing, |
| 174 | // makes exactly one summarizer call, and installs exactly one projection. |
| 175 | func TestManualCompactBelowCeilingKeepsSingleFold(t *testing.T) { |
| 176 | prov := &recordingSharedWindowProvider{sharedWindowTestProvider: sharedWindowTestProvider{budget: 128 * 1024, shared: true}} |
| 177 | sess := foldableSessionOverForce(60) |
| 178 | a := agentOverForceWindow(t, prov, sess, 50_000) |
| 179 | before := a.estimatedVisibleRequestTokens(a.modelVisibleMessages()) |
| 180 | if before >= a.hardInputCeiling() || before < a.compactTrigger() { |
| 181 | t.Fatalf("fixture est = %d, want between trigger %d and hard %d", before, a.compactTrigger(), a.hardInputCeiling()) |
| 182 | } |
| 183 | |
| 184 | if err := a.CompactNow(context.Background(), ""); err != nil { |
| 185 | t.Fatalf("manual compact below ceiling: %v", err) |
| 186 | } |
| 187 | if prov.calls != 1 { |
| 188 | t.Fatalf("summary calls = %d, want exactly 1", prov.calls) |
| 189 | } |
| 190 | if got := a.currentProjectionVersion(); got != 1 { |
| 191 | t.Fatalf("projection version = %d, want exactly 1 (single fold, no prune)", got) |
| 192 | } |
| 193 | visible := a.modelVisibleMessages() |
| 194 | if last := visible[len(visible)-1]; last.Role != provider.RoleUser || last.Content != "continue" { |
| 195 | t.Fatalf("verbatim tail not retained: %+v", last) |
| 196 | } |
| 197 | } |
| 198 |