|
| 1 | +import { describe, expect, it } from "vitest"; |
| 2 | +import { |
| 3 | + NON_DELIVERABLE_TERMINAL_TURN_REASON, |
| 4 | + resolveAttemptTrajectoryTerminal, |
| 5 | + resolveTerminalAssistantTexts, |
| 6 | + type ResolveAttemptTrajectoryTerminalParams, |
| 7 | +} from "./attempt-trajectory-status.js"; |
| 8 | + |
| 9 | +function baseParams( |
| 10 | + overrides: Partial<ResolveAttemptTrajectoryTerminalParams> = {}, |
| 11 | +): ResolveAttemptTrajectoryTerminalParams { |
| 12 | + return { |
| 13 | + aborted: false, |
| 14 | + timedOut: false, |
| 15 | + assistantTexts: [], |
| 16 | + toolMetas: [], |
| 17 | + didSendViaMessagingTool: false, |
| 18 | + didSendDeterministicApprovalPrompt: false, |
| 19 | + messagingToolSentTexts: [], |
| 20 | + messagingToolSentMediaUrls: [], |
| 21 | + messagingToolSentTargets: [], |
| 22 | + successfulCronAdds: 0, |
| 23 | + synthesizedPayloadCount: 0, |
| 24 | + ...overrides, |
| 25 | + }; |
| 26 | +} |
| 27 | + |
| 28 | +describe("attempt trajectory status", () => { |
| 29 | + it("marks a terminal turn without visible text, tools, or delivery as an error", () => { |
| 30 | + expect(resolveAttemptTrajectoryTerminal(baseParams())).toEqual({ |
| 31 | + status: "error", |
| 32 | + terminalError: NON_DELIVERABLE_TERMINAL_TURN_REASON, |
| 33 | + }); |
| 34 | + }); |
| 35 | + |
| 36 | + it("keeps visible assistant text as success", () => { |
| 37 | + expect( |
| 38 | + resolveAttemptTrajectoryTerminal(baseParams({ assistantTexts: ["Visible answer."] })), |
| 39 | + ).toEqual({ status: "success" }); |
| 40 | + }); |
| 41 | + |
| 42 | + it("keeps committed messaging tool delivery as success even without assistant text", () => { |
| 43 | + expect( |
| 44 | + resolveAttemptTrajectoryTerminal( |
| 45 | + baseParams({ |
| 46 | + didSendViaMessagingTool: true, |
| 47 | + messagingToolSentTargets: [{ channel: "telegram" }], |
| 48 | + }), |
| 49 | + ), |
| 50 | + ).toEqual({ status: "success" }); |
| 51 | + }); |
| 52 | + |
| 53 | + it("does not treat an uncommitted messaging tool attempt as delivery", () => { |
| 54 | + expect( |
| 55 | + resolveAttemptTrajectoryTerminal( |
| 56 | + baseParams({ |
| 57 | + didSendViaMessagingTool: true, |
| 58 | + messagingToolSentTexts: [" "], |
| 59 | + messagingToolSentMediaUrls: [" "], |
| 60 | + }), |
| 61 | + ), |
| 62 | + ).toEqual({ |
| 63 | + status: "error", |
| 64 | + terminalError: NON_DELIVERABLE_TERMINAL_TURN_REASON, |
| 65 | + }); |
| 66 | + }); |
| 67 | + |
| 68 | + it("does not treat tool metadata alone as terminal progress", () => { |
| 69 | + expect( |
| 70 | + resolveAttemptTrajectoryTerminal( |
| 71 | + baseParams({ |
| 72 | + toolMetas: [{ toolName: "read" }], |
| 73 | + }), |
| 74 | + ), |
| 75 | + ).toEqual({ |
| 76 | + status: "error", |
| 77 | + terminalError: NON_DELIVERABLE_TERMINAL_TURN_REASON, |
| 78 | + }); |
| 79 | + }); |
| 80 | + |
| 81 | + it("keeps synthesized terminal payloads as success", () => { |
| 82 | + expect(resolveAttemptTrajectoryTerminal(baseParams({ synthesizedPayloadCount: 1 }))).toEqual({ |
| 83 | + status: "success", |
| 84 | + }); |
| 85 | + }); |
| 86 | + |
| 87 | + it("keeps heartbeat responses as success", () => { |
| 88 | + expect( |
| 89 | + resolveAttemptTrajectoryTerminal( |
| 90 | + baseParams({ |
| 91 | + heartbeatToolResponse: { notify: false, summary: "ok" }, |
| 92 | + }), |
| 93 | + ), |
| 94 | + ).toEqual({ |
| 95 | + status: "success", |
| 96 | + }); |
| 97 | + }); |
| 98 | + |
| 99 | + it("does not treat expected silent turns as non-deliverable failures", () => { |
| 100 | + expect(resolveAttemptTrajectoryTerminal(baseParams({ silentExpected: true }))).toEqual({ |
| 101 | + status: "success", |
| 102 | + }); |
| 103 | + }); |
| 104 | + |
| 105 | + it("does not treat eligible empty silent replies as non-deliverable failures", () => { |
| 106 | + expect( |
| 107 | + resolveAttemptTrajectoryTerminal(baseParams({ emptyAssistantReplyIsSilent: true })), |
| 108 | + ).toEqual({ |
| 109 | + status: "success", |
| 110 | + }); |
| 111 | + }); |
| 112 | + |
| 113 | + it("does not let the raw silent policy hide ineligible empty failures", () => { |
| 114 | + expect( |
| 115 | + resolveAttemptTrajectoryTerminal(baseParams({ emptyAssistantReplyIsSilent: false })), |
| 116 | + ).toEqual({ |
| 117 | + status: "error", |
| 118 | + terminalError: NON_DELIVERABLE_TERMINAL_TURN_REASON, |
| 119 | + }); |
| 120 | + }); |
| 121 | + |
| 122 | + it("uses safe last-assistant fallback text for terminal delivery status", () => { |
| 123 | + expect( |
| 124 | + resolveTerminalAssistantTexts({ |
| 125 | + assistantTexts: [], |
| 126 | + lastAssistantStopReason: "stop", |
| 127 | + lastAssistantVisibleText: "Fallback answer.", |
| 128 | + }), |
| 129 | + ).toEqual(["Fallback answer."]); |
| 130 | + expect( |
| 131 | + resolveTerminalAssistantTexts({ |
| 132 | + assistantTexts: [], |
| 133 | + lastAssistantStopReason: "error", |
| 134 | + lastAssistantVisibleText: "Raw provider error", |
| 135 | + }), |
| 136 | + ).toEqual([]); |
| 137 | + }); |
| 138 | + |
| 139 | + it("marks terminal tool-use attempts as non-deliverable without explicit delivery", () => { |
| 140 | + expect( |
| 141 | + resolveAttemptTrajectoryTerminal( |
| 142 | + baseParams({ |
| 143 | + assistantTexts: ["I will update that file."], |
| 144 | + toolMetas: [{ toolName: "write" }], |
| 145 | + lastAssistantStopReason: "toolUse", |
| 146 | + }), |
| 147 | + ), |
| 148 | + ).toEqual({ |
| 149 | + status: "error", |
| 150 | + terminalError: NON_DELIVERABLE_TERMINAL_TURN_REASON, |
| 151 | + }); |
| 152 | + expect( |
| 153 | + resolveAttemptTrajectoryTerminal( |
| 154 | + baseParams({ |
| 155 | + assistantTexts: ["I sent the reply."], |
| 156 | + didSendViaMessagingTool: true, |
| 157 | + messagingToolSentTexts: ["sent"], |
| 158 | + lastAssistantStopReason: "toolUse", |
| 159 | + }), |
| 160 | + ), |
| 161 | + ).toEqual({ status: "success" }); |
| 162 | + }); |
| 163 | + |
| 164 | + it("preserves prompt errors and interrupts", () => { |
| 165 | + expect( |
| 166 | + resolveAttemptTrajectoryTerminal(baseParams({ promptError: new Error("boom") })), |
| 167 | + ).toEqual({ status: "error" }); |
| 168 | + expect(resolveAttemptTrajectoryTerminal(baseParams({ timedOut: true }))).toEqual({ |
| 169 | + status: "interrupted", |
| 170 | + }); |
| 171 | + }); |
| 172 | +}); |
0 commit comments