| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "strings" |
| 6 | "testing" |
| 7 | |
| 8 | "reasonix/internal/event" |
| 9 | "reasonix/internal/provider" |
| 10 | "reasonix/internal/tool" |
| 11 | ) |
| 12 | |
| 13 | func reasoningTurn() [][]provider.Chunk { |
| 14 | return [][]provider.Chunk{{ |
| 15 | {Type: provider.ChunkReasoning, Text: "think A "}, |
| 16 | {Type: provider.ChunkReasoning, Text: "think B"}, |
| 17 | {Type: provider.ChunkText, Text: "the answer"}, |
| 18 | {Type: provider.ChunkDone}, |
| 19 | }} |
| 20 | } |
| 21 | |
| 22 | func recordReasoning(events *[]string) event.Sink { |
| 23 | return event.FuncSink(func(e event.Event) { |
| 24 | if e.Kind == event.Reasoning { |
| 25 | *events = append(*events, e.Text) |
| 26 | } |
| 27 | }) |
| 28 | } |
| 29 | |
| 30 | func assistantReasoning(msgs []provider.Message) string { |
| 31 | for _, m := range msgs { |
| 32 | if m.Role == provider.RoleAssistant { |
| 33 | return m.ReasoningContent |
| 34 | } |
| 35 | } |
| 36 | return "" |
| 37 | } |
| 38 | |
| 39 | type reasoningRoundTripScriptedProvider struct { |
| 40 | *scriptedProvider |
| 41 | } |
| 42 | |
| 43 | func (reasoningRoundTripScriptedProvider) RequiresReasoningRoundTrip() bool { return true } |
| 44 | |
| 45 | // TestPostLLMCallAbsentStreamsReasoningLive is the regression guard: with no |
| 46 | // PostLLMCall hook, reasoning must still stream chunk-by-chunk (one Reasoning |
| 47 | // event per delta) so the live "thinking…" display keeps working. |
| 48 | func TestPostLLMCallAbsentStreamsReasoningLive(t *testing.T) { |
| 49 | prov := &scriptedProvider{name: "p", turns: reasoningTurn()} |
| 50 | var reasoningEvents []string |
| 51 | a := New(prov, tool.NewRegistry(), NewSession(""), Options{}, recordReasoning(&reasoningEvents)) |
| 52 | |
| 53 | if err := a.Run(context.Background(), "go"); err != nil { |
| 54 | t.Fatalf("Run: %v", err) |
| 55 | } |
| 56 | |
| 57 | if len(reasoningEvents) != 2 { |
| 58 | t.Fatalf("want 2 live reasoning events (one per chunk), got %d: %v", len(reasoningEvents), reasoningEvents) |
| 59 | } |
| 60 | if joined := strings.Join(reasoningEvents, ""); joined != "think A think B" { |
| 61 | t.Fatalf("streamed reasoning = %q, want the full chain", joined) |
| 62 | } |
| 63 | if got := assistantReasoning(a.session.Messages); got != "think A think B" { |
| 64 | t.Fatalf("stored reasoning = %q, want the untransformed chain", got) |
| 65 | } |
| 66 | } |
| 67 | |
| 68 | // TestPostLLMCallTransformsReasoningOnce proves a configured hook suppresses the |
| 69 | // live stream, sees the full reasoning, and its output replaces both the single |
| 70 | // emitted Reasoning event and the stored reasoning_content. |
| 71 | func TestPostLLMCallTransformsReasoningOnce(t *testing.T) { |
| 72 | prov := &scriptedProvider{name: "p", turns: reasoningTurn()} |
| 73 | var reasoningEvents []string |
| 74 | h := &stubHooks{hasPostLLM: true, postLLMOut: "TRANSLATED"} |
| 75 | a := New(prov, tool.NewRegistry(), NewSession(""), Options{Hooks: h}, recordReasoning(&reasoningEvents)) |
| 76 | |
| 77 | if err := a.Run(context.Background(), "go"); err != nil { |
| 78 | t.Fatalf("Run: %v", err) |
| 79 | } |
| 80 | |
| 81 | if len(reasoningEvents) != 1 || reasoningEvents[0] != "TRANSLATED" { |
| 82 | t.Fatalf("want one transformed reasoning event, got %v", reasoningEvents) |
| 83 | } |
| 84 | if len(h.postLLMSeen) != 1 || h.postLLMSeen[0] != "think A think B" { |
| 85 | t.Fatalf("hook saw %v, want the full original reasoning once", h.postLLMSeen) |
| 86 | } |
| 87 | if len(h.postLLMTurns) != 1 || h.postLLMTurns[0] != 1 { |
| 88 | t.Fatalf("hook turns = %v, want [1]", h.postLLMTurns) |
| 89 | } |
| 90 | if got := assistantReasoning(a.session.Messages); got != "TRANSLATED" { |
| 91 | t.Fatalf("stored reasoning = %q, want the hook's replacement", got) |
| 92 | } |
| 93 | } |
| 94 | |
| 95 | func TestPostLLMCallKeepsOriginalForReasoningRoundTripProvider(t *testing.T) { |
| 96 | prov := reasoningRoundTripScriptedProvider{&scriptedProvider{name: "p", turns: reasoningTurn()}} |
| 97 | h := &stubHooks{hasPostLLM: true, postLLMOut: "TRANSLATED"} |
| 98 | a := New(prov, tool.NewRegistry(), NewSession(""), Options{Hooks: h}, event.Discard) |
| 99 | |
| 100 | if err := a.Run(context.Background(), "go"); err != nil { |
| 101 | t.Fatalf("Run: %v", err) |
| 102 | } |
| 103 | if got := assistantReasoning(a.session.Messages); got != "think A think B" { |
| 104 | t.Fatalf("stored reasoning = %q, want raw provider reasoning for replay", got) |
| 105 | } |
| 106 | } |
| 107 | |
| 108 | // TestPostLLMCallKeepsOriginalForProviderReasoningMetadata proves that a |
| 109 | // provider-issued reasoning item ID/status pins the original reasoning text. |
| 110 | // Replaying hook-transformed text beside that metadata can make the provider |
| 111 | // reject the next request because the ID no longer identifies the same item. |
| 112 | func TestPostLLMCallKeepsOriginalForProviderReasoningMetadata(t *testing.T) { |
| 113 | prov := &scriptedProvider{name: "p", turns: [][]provider.Chunk{{ |
| 114 | {Type: provider.ChunkReasoning, Text: "think A "}, |
| 115 | {Type: provider.ChunkReasoning, Text: "think B"}, |
| 116 | {Type: provider.ChunkReasoning, ReasoningID: "rs_123", ReasoningStatus: "completed"}, |
| 117 | {Type: provider.ChunkText, Text: "answer"}, |
| 118 | {Type: provider.ChunkDone}, |
| 119 | }}} |
| 120 | var reasoningEvents []string |
| 121 | h := &stubHooks{hasPostLLM: true, postLLMOut: "TRANSLATED"} |
| 122 | a := New(prov, tool.NewRegistry(), NewSession(""), Options{Hooks: h}, recordReasoning(&reasoningEvents)) |
| 123 | |
| 124 | if err := a.Run(context.Background(), "go"); err != nil { |
| 125 | t.Fatalf("Run: %v", err) |
| 126 | } |
| 127 | if len(reasoningEvents) != 1 || reasoningEvents[0] != "TRANSLATED" { |
| 128 | t.Fatalf("want transformed reasoning shown live, got %v", reasoningEvents) |
| 129 | } |
| 130 | for _, m := range a.session.Messages { |
| 131 | if m.Role != provider.RoleAssistant { |
| 132 | continue |
| 133 | } |
| 134 | if m.ReasoningContent != "think A think B" { |
| 135 | t.Fatalf("stored reasoning = %q, want original provider text", m.ReasoningContent) |
| 136 | } |
| 137 | if m.ReasoningID != "rs_123" || m.ReasoningStatus != "completed" { |
| 138 | t.Fatalf("stored reasoning metadata = (%q, %q), want (rs_123, completed)", m.ReasoningID, m.ReasoningStatus) |
| 139 | } |
| 140 | return |
| 141 | } |
| 142 | t.Fatal("assistant message was not stored") |
| 143 | } |
| 144 | |
| 145 | // TestPostLLMCallConfiguredButNoReasoning makes sure a hook with an empty |
| 146 | // reasoning chain neither calls the hook nor emits a stray Reasoning event. |
| 147 | func TestPostLLMCallConfiguredButNoReasoning(t *testing.T) { |
| 148 | prov := &scriptedProvider{name: "p", turns: [][]provider.Chunk{{ |
| 149 | {Type: provider.ChunkText, Text: "answer, no thinking"}, |
| 150 | {Type: provider.ChunkDone}, |
| 151 | }}} |
| 152 | var reasoningEvents []string |
| 153 | h := &stubHooks{hasPostLLM: true, postLLMOut: "should not be used"} |
| 154 | a := New(prov, tool.NewRegistry(), NewSession(""), Options{Hooks: h}, recordReasoning(&reasoningEvents)) |
| 155 | |
| 156 | if err := a.Run(context.Background(), "go"); err != nil { |
| 157 | t.Fatalf("Run: %v", err) |
| 158 | } |
| 159 | |
| 160 | if len(reasoningEvents) != 0 { |
| 161 | t.Fatalf("no reasoning should emit no Reasoning events, got %v", reasoningEvents) |
| 162 | } |
| 163 | if len(h.postLLMSeen) != 0 { |
| 164 | t.Fatalf("hook should not fire on empty reasoning, saw %v", h.postLLMSeen) |
| 165 | } |
| 166 | } |
| 167 | |
| 168 | // TestPostLLMCallKeepsSignedReasoningOriginal proves that when the reasoning is |
| 169 | // pinned by a provider signature (Anthropic extended thinking), a transform hook |
| 170 | // changes only the live display — the stored reasoning_content stays the original |
| 171 | // so the signed thinking block can be replayed verbatim on the next tool-call |
| 172 | // turn. Storing the transformed text under the original signature is a 400. |
| 173 | func TestPostLLMCallKeepsSignedReasoningOriginal(t *testing.T) { |
| 174 | prov := &scriptedProvider{name: "p", turns: [][]provider.Chunk{{ |
| 175 | {Type: provider.ChunkReasoning, Text: "think A "}, |
| 176 | {Type: provider.ChunkReasoning, Text: "think B", Signature: "sig-xyz"}, |
| 177 | {Type: provider.ChunkText, Text: "answer"}, |
| 178 | {Type: provider.ChunkDone}, |
| 179 | }}} |
| 180 | var reasoningEvents []string |
| 181 | h := &stubHooks{hasPostLLM: true, postLLMOut: "TRANSLATED"} |
| 182 | a := New(prov, tool.NewRegistry(), NewSession(""), Options{Hooks: h}, recordReasoning(&reasoningEvents)) |
| 183 | |
| 184 | if err := a.Run(context.Background(), "go"); err != nil { |
| 185 | t.Fatalf("Run: %v", err) |
| 186 | } |
| 187 | if len(reasoningEvents) != 1 || reasoningEvents[0] != "TRANSLATED" { |
| 188 | t.Fatalf("want the transformed reasoning shown live, got %v", reasoningEvents) |
| 189 | } |
| 190 | if got := assistantReasoning(a.session.Messages); got != "think A think B" { |
| 191 | t.Fatalf("stored reasoning = %q, want the original (signature pins it)", got) |
| 192 | } |
| 193 | for _, m := range a.session.Messages { |
| 194 | if m.Role == provider.RoleAssistant && m.ReasoningSignature != "sig-xyz" { |
| 195 | t.Fatalf("stored signature = %q, want sig-xyz alongside its original text", m.ReasoningSignature) |
| 196 | } |
| 197 | } |
| 198 | } |
| 199 |