| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "strings" |
| 6 | "testing" |
| 7 | |
| 8 | "reasonix/internal/provider" |
| 9 | ) |
| 10 | |
| 11 | type overflowSummaryProvider struct { |
| 12 | requests []provider.Request |
| 13 | } |
| 14 | |
| 15 | func (p *overflowSummaryProvider) Name() string { return "overflow-summary" } |
| 16 | |
| 17 | func (p *overflowSummaryProvider) ContextBudgetPolicy() provider.ContextBudgetPolicy { |
| 18 | return provider.ContextBudgetPolicy{ |
| 19 | WindowMode: provider.ContextWindowShared, |
| 20 | LimitMode: provider.OutputLimitAlways, |
| 21 | } |
| 22 | } |
| 23 | |
| 24 | func (p *overflowSummaryProvider) Stream(_ context.Context, req provider.Request) (<-chan provider.Chunk, error) { |
| 25 | p.requests = append(p.requests, req) |
| 26 | return chunks( |
| 27 | provider.Chunk{Type: provider.ChunkText, Text: "compact durable summary"}, |
| 28 | provider.Chunk{Type: provider.ChunkDone}, |
| 29 | ), nil |
| 30 | } |
| 31 | |
| 32 | func TestOverflowSummarizesLargestAdmissibleContiguousPrefix(t *testing.T) { |
| 33 | sess := foldableSessionOverForce(120) |
| 34 | prov := &overflowSummaryProvider{} |
| 35 | a := agentOverForceWindow(t, prov, sess, 60_000) |
| 36 | msgs := sess.Snapshot() |
| 37 | head, plannedEnd, ok := a.planFoldRegion(msgs, true, false) |
| 38 | if !ok { |
| 39 | t.Fatal("fixture has no foldable prefix") |
| 40 | } |
| 41 | safeEnd := a.maximumSafeSummaryPrefixEnd(msgs, head, plannedEnd, "") |
| 42 | if safeEnd <= head || safeEnd >= plannedEnd { |
| 43 | t.Fatalf("safe fold end = %d, want a non-empty prefix smaller than planned end %d", safeEnd, plannedEnd) |
| 44 | } |
| 45 | |
| 46 | if err := prepareContext(context.Background(), a, CompactionTriggerOverflow); err != nil { |
| 47 | t.Fatalf("overflow recovery: %v", err) |
| 48 | } |
| 49 | if len(prov.requests) != 1 { |
| 50 | t.Fatalf("summary requests = %d, want 1", len(prov.requests)) |
| 51 | } |
| 52 | req := prov.requests[0] |
| 53 | if got, max := a.estimatedRequestTokens(req), a.effectiveContextWindow()-a.summaryOutputBudget()-protocolReserveTokens; got > max { |
| 54 | t.Fatalf("summary request tokens = %d, exceeds admissible input %d", got, max) |
| 55 | } |
| 56 | if receipt := a.sess.compactionState.LastReceipt; receipt == nil || receipt.CoveredCount != safeEnd { |
| 57 | t.Fatalf("receipt = %+v, want covered prefix %d", receipt, safeEnd) |
| 58 | } |
| 59 | if current := a.contextManager().currentPrepared(); current.InputTokens >= a.hardInputCeiling() { |
| 60 | t.Fatalf("recovered projection tokens = %d, hard ceiling = %d", current.InputTokens, a.hardInputCeiling()) |
| 61 | } |
| 62 | } |
| 63 | |
| 64 | func TestMaximumSafeSummaryPrefixKeepsToolPairsTogether(t *testing.T) { |
| 65 | // A 10k window leaves 7244 prompt tokens after the scaled 2500-token |
| 66 | // summary budget and protocol reserve. The initial boundary splits the tool |
| 67 | // results, so it must retreat across the whole assistant/tool group. |
| 68 | msgs := []provider.Message{ |
| 69 | {Role: provider.RoleSystem, Content: "system"}, |
| 70 | {Role: provider.RoleUser, Content: "old task"}, |
| 71 | {Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{{ID: "call-1", Name: "read", Arguments: `{}`}}}, |
| 72 | {Role: provider.RoleTool, ToolCallID: "call-1", Name: "read", Content: strings.Repeat("first result payload. ", 150)}, |
| 73 | {Role: provider.RoleTool, ToolCallID: "call-1", Name: "read", Content: strings.Repeat("second result payload. ", 1500)}, |
| 74 | {Role: provider.RoleUser, Content: "recent task"}, |
| 75 | } |
| 76 | a := &Agent{ |
| 77 | agentConfig: agentConfig{contextWindow: 10_000}, |
| 78 | svc: agentServices{prov: &overflowSummaryProvider{}}, |
| 79 | sess: sessionRuntime{conversation: &Session{Messages: msgs}}, |
| 80 | } |
| 81 | end := a.maximumSafeSummaryPrefixEnd(msgs, 1, len(msgs)-1, "") |
| 82 | if end != 2 { |
| 83 | t.Fatalf("fold boundary = %d, want 2 so the assistant call and both results stay in the tail", end) |
| 84 | } |
| 85 | } |
| 86 | |
| 87 | // opaqueWindowProvider declares no ContextBudgetPolicy (Unknown window mode) |
| 88 | // and is never admitted, reproducing the #9572 shape: a freshly switched-to |
| 89 | // gateway whose summary request used to bypass the safe-prefix cap entirely. |
| 90 | type opaqueWindowProvider struct { |
| 91 | requests []provider.Request |
| 92 | } |
| 93 | |
| 94 | func (p *opaqueWindowProvider) Name() string { return "opaque-window" } |
| 95 | |
| 96 | func (p *opaqueWindowProvider) Stream(_ context.Context, req provider.Request) (<-chan provider.Chunk, error) { |
| 97 | p.requests = append(p.requests, req) |
| 98 | return chunks( |
| 99 | provider.Chunk{Type: provider.ChunkText, Text: "compact durable summary"}, |
| 100 | provider.Chunk{Type: provider.ChunkDone}, |
| 101 | ), nil |
| 102 | } |
| 103 | |
| 104 | func TestPressureSummaryCappedByConfiguredWindowWithoutAdmission(t *testing.T) { |
| 105 | sess := foldableSessionOverForce(120) |
| 106 | prov := &opaqueWindowProvider{} |
| 107 | a := agentOverForceWindow(t, prov, sess, 60_000) |
| 108 | if err := prepareContext(context.Background(), a, CompactionTriggerPressure); err != nil { |
| 109 | t.Fatalf("pressure maintenance: %v", err) |
| 110 | } |
| 111 | if len(prov.requests) != 1 { |
| 112 | t.Fatalf("summary requests = %d, want 1", len(prov.requests)) |
| 113 | } |
| 114 | req := prov.requests[0] |
| 115 | if got, max := a.estimatedRequestTokens(req), a.effectiveContextWindow()-a.summaryOutputBudget()-protocolReserveTokens; got > max { |
| 116 | t.Fatalf("summary request tokens = %d, exceeds configured-window cap %d", got, max) |
| 117 | } |
| 118 | } |
| 119 | |
| 120 | func TestPressureSummaryCappedByLearnedWindowAfterSessionReset(t *testing.T) { |
| 121 | sess := foldableSessionOverForce(120) |
| 122 | prov := &opaqueWindowProvider{} |
| 123 | a := agentOverForceWindow(t, prov, sess, 60_000) |
| 124 | a.contextWindow = 0 |
| 125 | a.sess.output.learned.Store(&learnedContextBudget{windowTokens: 60_000}) |
| 126 | a.storeAdmission(contextAdmission{ObservedWindow: 60_000}) |
| 127 | a.SetSession(sess) |
| 128 | if got := a.lastAdmission().ObservedWindow; got != 0 { |
| 129 | t.Fatalf("session reset retained observed window %d", got) |
| 130 | } |
| 131 | |
| 132 | msgs := sess.Snapshot() |
| 133 | head, plannedEnd, ok := a.planFoldRegion(msgs, true, false) |
| 134 | if !ok { |
| 135 | t.Fatal("fixture has no foldable prefix") |
| 136 | } |
| 137 | safeEnd := a.maximumSafeSummaryPrefixEnd(msgs, head, plannedEnd, "") |
| 138 | if safeEnd <= head || safeEnd >= plannedEnd { |
| 139 | t.Fatalf("safe fold end = %d, want a non-empty prefix smaller than planned end %d", safeEnd, plannedEnd) |
| 140 | } |
| 141 | request := a.summaryRequest(msgs[head:safeEnd], "") |
| 142 | if got, max := a.estimatedRequestTokens(request), a.effectiveContextWindow()-a.summaryOutputBudget()-protocolReserveTokens; got > max { |
| 143 | t.Fatalf("summary request tokens = %d, exceeds learned-window cap %d", got, max) |
| 144 | } |
| 145 | } |
| 146 | |
| 147 | func smallWindowPressureSession() *Session { |
| 148 | return &Session{Messages: []provider.Message{ |
| 149 | {Role: provider.RoleSystem, Content: "sys"}, |
| 150 | {Role: provider.RoleUser, Content: strings.Repeat("x", 20_000)}, |
| 151 | {Role: provider.RoleAssistant, Content: strings.Repeat("y", 12_000)}, |
| 152 | {Role: provider.RoleUser, Content: "recent"}, |
| 153 | }} |
| 154 | } |
| 155 | |
| 156 | func TestPressureSummaryScalesOutputBudgetForSmallWindow(t *testing.T) { |
| 157 | const window = 10_000 |
| 158 | prov := &opaqueWindowProvider{} |
| 159 | sess := smallWindowPressureSession() |
| 160 | a := agentOverForceWindow(t, prov, sess, window) |
| 161 | est := a.estimatedVisibleRequestTokens(a.modelVisibleMessages()) |
| 162 | if est < a.compactTrigger() || est >= a.hardInputCeiling() { |
| 163 | t.Fatalf("fixture tokens = %d, want pressure range [%d, %d)", est, a.compactTrigger(), a.hardInputCeiling()) |
| 164 | } |
| 165 | if err := prepareContext(context.Background(), a, CompactionTriggerPressure); err != nil { |
| 166 | t.Fatalf("small-window pressure maintenance: %v", err) |
| 167 | } |
| 168 | if len(prov.requests) != 1 { |
| 169 | t.Fatalf("summary requests = %d, want 1", len(prov.requests)) |
| 170 | } |
| 171 | req := prov.requests[0] |
| 172 | if req.MaxTokens != window/4 { |
| 173 | t.Fatalf("summary max tokens = %d, want scaled budget %d", req.MaxTokens, window/4) |
| 174 | } |
| 175 | if got := a.estimatedRequestTokens(req) + req.MaxTokens + protocolReserveTokens; got > window { |
| 176 | t.Fatalf("summary request total = %d, exceeds small window %d", got, window) |
| 177 | } |
| 178 | } |
| 179 | |
| 180 | func TestPressureNoSafePrefixRecordsBudgetReason(t *testing.T) { |
| 181 | const window = 10_000 |
| 182 | prov := &opaqueWindowProvider{} |
| 183 | a := agentOverForceWindow(t, prov, smallWindowPressureSession(), window) |
| 184 | |
| 185 | _, err := a.contextManager().Prepare(context.Background(), ContextPreparePolicy{ |
| 186 | Trigger: CompactionTriggerPressure, Instructions: strings.Repeat("preserve this focus ", window), |
| 187 | }) |
| 188 | if err != nil { |
| 189 | t.Fatalf("pressure maintenance should soft-skip: %v", err) |
| 190 | } |
| 191 | if len(prov.requests) != 0 { |
| 192 | t.Fatalf("summary requests = %d, want none when no prefix fits", len(prov.requests)) |
| 193 | } |
| 194 | receipt := a.sess.compactionState.LastReceipt |
| 195 | if receipt == nil || receipt.Status != "blocked" || !strings.Contains(receipt.Reason, "no balanced prefix leaves enough room") { |
| 196 | t.Fatalf("receipt = %+v, want precise summary-budget reason", receipt) |
| 197 | } |
| 198 | } |
| 199 |