|
| 1 | +import { describe, expect, it } from "vitest"; |
| 2 | +import { |
| 3 | + analyzeConversationState, |
| 4 | + closeToolLoopForThinking, |
| 5 | + hasPossibleCompactedThinking, |
| 6 | + looksLikeCompactedThinkingTurn, |
| 7 | + needsThinkingRecovery, |
| 8 | +} from "./thinking-recovery"; |
| 9 | + |
| 10 | +// ─── Fixtures ──────────────────────────────────────────────────────────────── |
| 11 | + |
| 12 | +function userMsg(text: string) { |
| 13 | + return { role: "user", parts: [{ text }] }; |
| 14 | +} |
| 15 | + |
| 16 | +function modelMsg(text: string) { |
| 17 | + return { role: "model", parts: [{ text }] }; |
| 18 | +} |
| 19 | + |
| 20 | +function modelWithThinking(text: string) { |
| 21 | + return { |
| 22 | + role: "model", |
| 23 | + parts: [{ thought: true, text: "thinking..." }, { text }], |
| 24 | + }; |
| 25 | +} |
| 26 | + |
| 27 | +function modelWithToolCall(name = "myTool") { |
| 28 | + return { |
| 29 | + role: "model", |
| 30 | + parts: [{ functionCall: { name, args: {} } }], |
| 31 | + }; |
| 32 | +} |
| 33 | + |
| 34 | +function modelWithThinkingAndToolCall(name = "myTool") { |
| 35 | + return { |
| 36 | + role: "model", |
| 37 | + parts: [ |
| 38 | + { thought: true, text: "reasoning..." }, |
| 39 | + { functionCall: { name, args: {} } }, |
| 40 | + ], |
| 41 | + }; |
| 42 | +} |
| 43 | + |
| 44 | +function toolResultMsg(name = "myTool") { |
| 45 | + return { |
| 46 | + role: "user", |
| 47 | + parts: [{ functionResponse: { name, response: { result: "ok" } } }], |
| 48 | + }; |
| 49 | +} |
| 50 | + |
| 51 | +// ─── analyzeConversationState ───────────────────────────────────────────────── |
| 52 | + |
| 53 | +describe("analyzeConversationState", () => { |
| 54 | + it("returns default state for empty contents", () => { |
| 55 | + const state = analyzeConversationState([]); |
| 56 | + expect(state.inToolLoop).toBe(false); |
| 57 | + expect(state.turnStartIdx).toBe(-1); |
| 58 | + expect(state.lastModelIdx).toBe(-1); |
| 59 | + }); |
| 60 | + |
| 61 | + it("returns default state for non-array input", () => { |
| 62 | + const state = analyzeConversationState(null as any); |
| 63 | + expect(state.inToolLoop).toBe(false); |
| 64 | + }); |
| 65 | + |
| 66 | + it("detects a simple user→model conversation (not in tool loop)", () => { |
| 67 | + const contents = [userMsg("hello"), modelMsg("hi there")]; |
| 68 | + const state = analyzeConversationState(contents); |
| 69 | + expect(state.inToolLoop).toBe(false); |
| 70 | + expect(state.lastModelIdx).toBe(1); |
| 71 | + expect(state.lastModelHasThinking).toBe(false); |
| 72 | + expect(state.lastModelHasToolCalls).toBe(false); |
| 73 | + }); |
| 74 | + |
| 75 | + it("detects thinking in last model message", () => { |
| 76 | + const contents = [userMsg("hello"), modelWithThinking("hi there")]; |
| 77 | + const state = analyzeConversationState(contents); |
| 78 | + expect(state.lastModelHasThinking).toBe(true); |
| 79 | + expect(state.turnHasThinking).toBe(true); |
| 80 | + }); |
| 81 | + |
| 82 | + it("detects tool loop: conversation ends with tool result", () => { |
| 83 | + const contents = [ |
| 84 | + userMsg("do something"), |
| 85 | + modelWithToolCall("search"), |
| 86 | + toolResultMsg("search"), |
| 87 | + ]; |
| 88 | + const state = analyzeConversationState(contents); |
| 89 | + expect(state.inToolLoop).toBe(true); |
| 90 | + expect(state.lastModelIdx).toBe(1); |
| 91 | + expect(state.lastModelHasToolCalls).toBe(true); |
| 92 | + }); |
| 93 | + |
| 94 | + it("detects tool loop with multiple tool results", () => { |
| 95 | + const contents = [ |
| 96 | + userMsg("do two things"), |
| 97 | + { role: "model", parts: [ |
| 98 | + { functionCall: { name: "a", args: {} } }, |
| 99 | + { functionCall: { name: "b", args: {} } }, |
| 100 | + ]}, |
| 101 | + { role: "user", parts: [ |
| 102 | + { functionResponse: { name: "a", response: {} } }, |
| 103 | + { functionResponse: { name: "b", response: {} } }, |
| 104 | + ]}, |
| 105 | + ]; |
| 106 | + const state = analyzeConversationState(contents); |
| 107 | + expect(state.inToolLoop).toBe(true); |
| 108 | + }); |
| 109 | + |
| 110 | + it("is NOT in tool loop when last message is a real user message", () => { |
| 111 | + const contents = [ |
| 112 | + userMsg("task"), |
| 113 | + modelWithToolCall(), |
| 114 | + toolResultMsg(), |
| 115 | + modelMsg("done"), |
| 116 | + userMsg("thanks"), |
| 117 | + ]; |
| 118 | + const state = analyzeConversationState(contents); |
| 119 | + expect(state.inToolLoop).toBe(false); |
| 120 | + }); |
| 121 | + |
| 122 | + it("tracks turn start correctly across multi-step tool loop", () => { |
| 123 | + const contents = [ |
| 124 | + userMsg("first real user"), |
| 125 | + modelWithThinkingAndToolCall("step1"), |
| 126 | + toolResultMsg("step1"), |
| 127 | + modelWithToolCall("step2"), |
| 128 | + toolResultMsg("step2"), |
| 129 | + ]; |
| 130 | + const state = analyzeConversationState(contents); |
| 131 | + expect(state.turnStartIdx).toBe(1); // first model message in turn |
| 132 | + expect(state.turnHasThinking).toBe(true); |
| 133 | + expect(state.inToolLoop).toBe(true); |
| 134 | + }); |
| 135 | + |
| 136 | + it("turns NOT having thinking when first model msg has no thinking", () => { |
| 137 | + const contents = [ |
| 138 | + userMsg("go"), |
| 139 | + modelWithToolCall("t1"), |
| 140 | + toolResultMsg("t1"), |
| 141 | + ]; |
| 142 | + const state = analyzeConversationState(contents); |
| 143 | + expect(state.turnHasThinking).toBe(false); |
| 144 | + expect(state.inToolLoop).toBe(true); |
| 145 | + }); |
| 146 | +}); |
| 147 | + |
| 148 | +// ─── needsThinkingRecovery ──────────────────────────────────────────────────── |
| 149 | + |
| 150 | +describe("needsThinkingRecovery", () => { |
| 151 | + it("returns false when not in tool loop", () => { |
| 152 | + expect(needsThinkingRecovery({ inToolLoop: false, turnHasThinking: false, |
| 153 | + turnStartIdx: -1, lastModelIdx: -1, lastModelHasThinking: false, |
| 154 | + lastModelHasToolCalls: false })).toBe(false); |
| 155 | + }); |
| 156 | + |
| 157 | + it("returns false when in tool loop but turn had thinking", () => { |
| 158 | + expect(needsThinkingRecovery({ inToolLoop: true, turnHasThinking: true, |
| 159 | + turnStartIdx: 1, lastModelIdx: 2, lastModelHasThinking: false, |
| 160 | + lastModelHasToolCalls: true })).toBe(false); |
| 161 | + }); |
| 162 | + |
| 163 | + it("returns true when in tool loop without thinking", () => { |
| 164 | + expect(needsThinkingRecovery({ inToolLoop: true, turnHasThinking: false, |
| 165 | + turnStartIdx: 1, lastModelIdx: 2, lastModelHasThinking: false, |
| 166 | + lastModelHasToolCalls: true })).toBe(true); |
| 167 | + }); |
| 168 | +}); |
| 169 | + |
| 170 | +// ─── closeToolLoopForThinking ───────────────────────────────────────────────── |
| 171 | + |
| 172 | +describe("closeToolLoopForThinking", () => { |
| 173 | + it("appends synthetic model + user messages", () => { |
| 174 | + const contents = [ |
| 175 | + userMsg("go"), |
| 176 | + modelWithToolCall("search"), |
| 177 | + toolResultMsg("search"), |
| 178 | + ]; |
| 179 | + const result = closeToolLoopForThinking(contents); |
| 180 | + expect(result.length).toBe(5); |
| 181 | + expect(result[3]?.role).toBe("model"); |
| 182 | + expect(result[4]?.role).toBe("user"); |
| 183 | + expect(result[4]?.parts[0]?.text).toBe("[Continue]"); |
| 184 | + }); |
| 185 | + |
| 186 | + it("strips thinking blocks from prior messages", () => { |
| 187 | + const contents = [ |
| 188 | + userMsg("hello"), |
| 189 | + modelWithThinking("response"), |
| 190 | + toolResultMsg(), |
| 191 | + ]; |
| 192 | + const result = closeToolLoopForThinking(contents); |
| 193 | + const modelMessages = result.filter((m) => m.role === "model"); |
| 194 | + for (const msg of modelMessages) { |
| 195 | + const parts: any[] = msg.parts ?? []; |
| 196 | + const hasThinking = parts.some((p: any) => p?.thought === true); |
| 197 | + expect(hasThinking).toBe(false); |
| 198 | + } |
| 199 | + }); |
| 200 | + |
| 201 | + it("uses singular message for single tool result", () => { |
| 202 | + const contents = [userMsg("go"), modelWithToolCall(), toolResultMsg()]; |
| 203 | + const result = closeToolLoopForThinking(contents); |
| 204 | + const syntheticModel = result[result.length - 2]; |
| 205 | + expect(syntheticModel?.parts[0]?.text).toBe("[Tool execution completed.]"); |
| 206 | + }); |
| 207 | + |
| 208 | + it("uses plural message for multiple tool results", () => { |
| 209 | + const contents = [ |
| 210 | + userMsg("go"), |
| 211 | + { role: "model", parts: [ |
| 212 | + { functionCall: { name: "a", args: {} } }, |
| 213 | + { functionCall: { name: "b", args: {} } }, |
| 214 | + ]}, |
| 215 | + { role: "user", parts: [ |
| 216 | + { functionResponse: { name: "a", response: {} } }, |
| 217 | + { functionResponse: { name: "b", response: {} } }, |
| 218 | + ]}, |
| 219 | + ]; |
| 220 | + const result = closeToolLoopForThinking(contents); |
| 221 | + const syntheticModel = result[result.length - 2]; |
| 222 | + expect(syntheticModel?.parts[0]?.text).toBe("[2 tool executions completed.]"); |
| 223 | + }); |
| 224 | + |
| 225 | + it("uses fallback message when no tool results present", () => { |
| 226 | + const contents = [userMsg("go"), modelMsg("working...")]; |
| 227 | + const result = closeToolLoopForThinking(contents); |
| 228 | + const syntheticModel = result[result.length - 2]; |
| 229 | + expect(syntheticModel?.parts[0]?.text).toBe("[Processing previous context.]"); |
| 230 | + }); |
| 231 | + |
| 232 | + it("does not mutate original contents array", () => { |
| 233 | + const contents = [userMsg("go"), modelWithToolCall(), toolResultMsg()]; |
| 234 | + const original = JSON.stringify(contents); |
| 235 | + closeToolLoopForThinking(contents); |
| 236 | + expect(JSON.stringify(contents)).toBe(original); |
| 237 | + }); |
| 238 | +}); |
| 239 | + |
| 240 | +// ─── looksLikeCompactedThinkingTurn ────────────────────────────────────────── |
| 241 | + |
| 242 | +describe("looksLikeCompactedThinkingTurn", () => { |
| 243 | + it("returns false for null / undefined", () => { |
| 244 | + expect(looksLikeCompactedThinkingTurn(null)).toBe(false); |
| 245 | + expect(looksLikeCompactedThinkingTurn(undefined)).toBe(false); |
| 246 | + }); |
| 247 | + |
| 248 | + it("returns false for message with no parts", () => { |
| 249 | + expect(looksLikeCompactedThinkingTurn({ role: "model", parts: [] })).toBe(false); |
| 250 | + }); |
| 251 | + |
| 252 | + it("returns false for message without function calls", () => { |
| 253 | + expect(looksLikeCompactedThinkingTurn(modelMsg("just text"))).toBe(false); |
| 254 | + }); |
| 255 | + |
| 256 | + it("returns false when message has thinking blocks alongside function call", () => { |
| 257 | + const msg = { |
| 258 | + role: "model", |
| 259 | + parts: [ |
| 260 | + { thought: true, text: "thinking" }, |
| 261 | + { functionCall: { name: "t", args: {} } }, |
| 262 | + ], |
| 263 | + }; |
| 264 | + expect(looksLikeCompactedThinkingTurn(msg)).toBe(false); |
| 265 | + }); |
| 266 | + |
| 267 | + it("returns false when text appears before function call (non-compacted)", () => { |
| 268 | + const msg = { |
| 269 | + role: "model", |
| 270 | + parts: [ |
| 271 | + { text: "I will now call the tool." }, |
| 272 | + { functionCall: { name: "t", args: {} } }, |
| 273 | + ], |
| 274 | + }; |
| 275 | + expect(looksLikeCompactedThinkingTurn(msg)).toBe(false); |
| 276 | + }); |
| 277 | + |
| 278 | + it("returns true for bare function call with no preceding text (looks compacted)", () => { |
| 279 | + const msg = modelWithToolCall("search"); |
| 280 | + expect(looksLikeCompactedThinkingTurn(msg)).toBe(true); |
| 281 | + }); |
| 282 | +}); |
| 283 | + |
| 284 | +// ─── hasPossibleCompactedThinking ──────────────────────────────────────────── |
| 285 | + |
| 286 | +describe("hasPossibleCompactedThinking", () => { |
| 287 | + it("returns false for empty contents", () => { |
| 288 | + expect(hasPossibleCompactedThinking([], 0)).toBe(false); |
| 289 | + }); |
| 290 | + |
| 291 | + it("returns false for invalid turnStartIdx", () => { |
| 292 | + expect(hasPossibleCompactedThinking([modelMsg("hi")], -1)).toBe(false); |
| 293 | + }); |
| 294 | + |
| 295 | + it("returns false when no model messages look compacted", () => { |
| 296 | + const contents = [userMsg("go"), modelWithThinkingAndToolCall(), toolResultMsg()]; |
| 297 | + expect(hasPossibleCompactedThinking(contents, 1)).toBe(false); |
| 298 | + }); |
| 299 | + |
| 300 | + it("returns true when a model message in turn looks compacted", () => { |
| 301 | + const contents = [ |
| 302 | + userMsg("go"), |
| 303 | + modelWithToolCall("search"), |
| 304 | + toolResultMsg("search"), |
| 305 | + ]; |
| 306 | + expect(hasPossibleCompactedThinking(contents, 1)).toBe(true); |
| 307 | + }); |
| 308 | + |
| 309 | + it("ignores model messages before turnStartIdx", () => { |
| 310 | + const contents = [ |
| 311 | + userMsg("first turn"), |
| 312 | + modelWithToolCall("old"), |
| 313 | + toolResultMsg("old"), |
| 314 | + userMsg("second turn"), |
| 315 | + modelWithThinkingAndToolCall("new"), |
| 316 | + toolResultMsg("new"), |
| 317 | + ]; |
| 318 | + // turnStart is 4 (second model message) |
| 319 | + expect(hasPossibleCompactedThinking(contents, 4)).toBe(false); |
| 320 | + }); |
| 321 | +}); |
0 commit comments