| 1 | package agent |
| 2 | |
| 3 | // Core agent-loop contract tests (agent-core simplification baseline). Each |
| 4 | // test pins one behavior the simplified loop promises; suites covering the |
| 5 | // remaining contract items are listed in docs/AGENT_CORE_SIMPLIFICATION.md. |
| 6 | |
| 7 | import ( |
| 8 | "context" |
| 9 | "reflect" |
| 10 | "strings" |
| 11 | "testing" |
| 12 | |
| 13 | "reasonix/internal/agent/testutil" |
| 14 | "reasonix/internal/event" |
| 15 | "reasonix/internal/provider" |
| 16 | "reasonix/internal/tool" |
| 17 | ) |
| 18 | |
| 19 | // TestContractCleanFinalMakesOneModelRequest pins the executor-only happy |
| 20 | // path: one user turn, one provider request, no host follow-ups. |
| 21 | func TestContractCleanFinalMakesOneModelRequest(t *testing.T) { |
| 22 | prov := &scriptedProvider{name: "p", turns: [][]provider.Chunk{ |
| 23 | {{Type: provider.ChunkText, Text: "the answer"}, {Type: provider.ChunkDone}}, |
| 24 | }} |
| 25 | a := New(prov, tool.NewRegistry(), NewSession(""), Options{}, event.Discard) |
| 26 | |
| 27 | if err := a.Run(context.Background(), "question"); err != nil { |
| 28 | t.Fatalf("Run: %v", err) |
| 29 | } |
| 30 | if len(prov.requests) != 1 { |
| 31 | t.Fatalf("provider requests = %d, want exactly one for a clean final", len(prov.requests)) |
| 32 | } |
| 33 | if got := lastAssistantContent(a.sess.conversation); got != "the answer" { |
| 34 | t.Fatalf("last assistant content = %q, want the final answer", got) |
| 35 | } |
| 36 | } |
| 37 | |
| 38 | // TestContractToolCallAdvancesToNextStep pins the tool-loop shape: a tool-call |
| 39 | // round executes and the loop continues to a second request for the final. |
| 40 | func TestContractToolCallAdvancesToNextStep(t *testing.T) { |
| 41 | mp := testutil.NewMock("m", |
| 42 | testutil.Turn{ToolCalls: []provider.ToolCall{{ID: "c1", Name: "echo", Arguments: `{"text":"hi"}`}}}, |
| 43 | testutil.Turn{Text: "all set"}, |
| 44 | ) |
| 45 | a := New(mp, echoRegistry(), NewSession(""), Options{}, event.Discard) |
| 46 | |
| 47 | if err := a.Run(withNoClosedLoop(context.Background()), "go"); err != nil { |
| 48 | t.Fatalf("Run: %v", err) |
| 49 | } |
| 50 | if got := mp.CallCount(); got != 2 { |
| 51 | t.Fatalf("provider calls = %d, want tool round + final", got) |
| 52 | } |
| 53 | if got := lastToolResult(a.Session(), "echo"); got == "" { |
| 54 | t.Fatal("tool result missing from session; tool round did not execute") |
| 55 | } |
| 56 | if got := lastAssistantContent(a.sess.conversation); got != "all set" { |
| 57 | t.Fatalf("last assistant content = %q, want the post-tool final", got) |
| 58 | } |
| 59 | } |
| 60 | |
| 61 | // TestContractThinkingSurvivesUnifiedRetry pins that the EMPTY_RESPONSE retry |
| 62 | // replays the same frozen request (thinking never disabled or reshaped) and |
| 63 | // that recovered reasoning stays on the committed assistant turn. |
| 64 | func TestContractThinkingSurvivesUnifiedRetry(t *testing.T) { |
| 65 | prov := &scriptedProvider{name: "p", turns: [][]provider.Chunk{ |
| 66 | {{Type: provider.ChunkDone}}, |
| 67 | { |
| 68 | {Type: provider.ChunkReasoning, Text: "need to think"}, |
| 69 | {Type: provider.ChunkText, Text: "the answer"}, |
| 70 | {Type: provider.ChunkDone}, |
| 71 | }, |
| 72 | }} |
| 73 | a := New(prov, tool.NewRegistry(), NewSession(""), Options{}, event.Discard) |
| 74 | |
| 75 | if err := a.Run(context.Background(), "question"); err != nil { |
| 76 | t.Fatalf("Run: %v", err) |
| 77 | } |
| 78 | if len(prov.requests) != 2 || !reflect.DeepEqual(prov.requests[0], prov.requests[1]) { |
| 79 | t.Fatalf("retry must replay the frozen request verbatim:\nfirst=%#v\nsecond=%#v", prov.requests[0], prov.requests[1]) |
| 80 | } |
| 81 | msgs := a.sess.conversation.Snapshot() |
| 82 | last := msgs[len(msgs)-1] |
| 83 | if last.Role != provider.RoleAssistant || last.ReasoningContent != "need to think" { |
| 84 | t.Fatalf("committed assistant turn lost reasoning: %+v", last) |
| 85 | } |
| 86 | } |
| 87 | |
| 88 | // TestContractNoLongLivedFallbackStateAfterRetryExhaustion pins that exhausted |
| 89 | // protocol retries fail the turn without installing any persistent degraded |
| 90 | // mode: the next turn runs a normal single-request loop. |
| 91 | func TestContractNoLongLivedFallbackStateAfterRetryExhaustion(t *testing.T) { |
| 92 | turns := [][]provider.Chunk{} |
| 93 | for range maxSamplingAttempts { |
| 94 | turns = append(turns, []provider.Chunk{{Type: provider.ChunkDone}}) |
| 95 | } |
| 96 | turns = append(turns, []provider.Chunk{{Type: provider.ChunkText, Text: "recovered"}, {Type: provider.ChunkDone}}) |
| 97 | prov := &scriptedProvider{name: "p", turns: turns} |
| 98 | a := New(prov, tool.NewRegistry(), NewSession(""), Options{}, event.Discard) |
| 99 | |
| 100 | if err := a.Run(context.Background(), "question"); err == nil { |
| 101 | t.Fatal("exhausted empty-response retries must fail the turn") |
| 102 | } |
| 103 | for _, m := range a.sess.conversation.Snapshot() { |
| 104 | if m.Role == provider.RoleAssistant && strings.TrimSpace(m.Content) != "" { |
| 105 | t.Fatalf("failed turn committed assistant content %q", m.Content) |
| 106 | } |
| 107 | } |
| 108 | if err := a.Run(context.Background(), "try again"); err != nil { |
| 109 | t.Fatalf("second Run after exhausted retries: %v", err) |
| 110 | } |
| 111 | if got := len(prov.requests) - maxSamplingAttempts; got != 1 { |
| 112 | t.Fatalf("second turn made %d requests, want a normal single request (no fallback mode)", got) |
| 113 | } |
| 114 | if got := lastAssistantContent(a.sess.conversation); got != "recovered" { |
| 115 | t.Fatalf("last assistant content = %q, want the clean recovery answer", got) |
| 116 | } |
| 117 | } |
| 118 | |
| 119 | // TestContractCleanFinalAddsNoSyntheticContinuation pins that a direct Run |
| 120 | // ends at the first clean final: no host-generated user message may appear in |
| 121 | // the session, whatever the model text promises. |
| 122 | func TestContractCleanFinalAddsNoSyntheticContinuation(t *testing.T) { |
| 123 | prov := &scriptedProvider{name: "p", turns: [][]provider.Chunk{ |
| 124 | {{Type: provider.ChunkText, Text: "I will handle it next round."}, {Type: provider.ChunkDone}}, |
| 125 | }} |
| 126 | a := New(prov, tool.NewRegistry(), NewSession(""), Options{}, event.Discard) |
| 127 | |
| 128 | if err := a.Run(context.Background(), "question"); err != nil { |
| 129 | t.Fatalf("Run: %v", err) |
| 130 | } |
| 131 | if len(prov.requests) != 1 { |
| 132 | t.Fatalf("provider requests = %d, want one; a promised-later answer must not buy a continuation", len(prov.requests)) |
| 133 | } |
| 134 | for _, prefix := range []string{StandardTodoContinuationPrefix, "visible answer", executorHandoffMarker} { |
| 135 | if sessionHasUserMessageContaining(a.sess.conversation, prefix) { |
| 136 | t.Fatalf("session contains synthetic continuation prompt %q", prefix) |
| 137 | } |
| 138 | } |
| 139 | } |
| 140 |