|
| 1 | +import { describe, test, expect } from "bun:test"; |
| 2 | +import { createPruningCompactor } from "./compactor.js"; |
| 3 | +import { condenseTurns } from "./summarizer.js"; |
| 4 | +import { |
| 5 | + extractContinuationFacts, |
| 6 | + repairSummary, |
| 7 | + verifyCompactionSummary, |
| 8 | + verifyOrRepair, |
| 9 | + VERIFY_REPAIR_HEADING, |
| 10 | +} from "./compaction-verify.js"; |
| 11 | +import type { |
| 12 | + ConversationTurn, |
| 13 | + ReactorState, |
| 14 | + StrategyContext, |
| 15 | +} from "@intx/types/runtime"; |
| 16 | + |
| 17 | +const mockStrategyCtx: StrategyContext = { |
| 18 | + state: {} as ReactorState, |
| 19 | + trigger: "test", |
| 20 | +}; |
| 21 | + |
| 22 | +function textTurn( |
| 23 | + role: ConversationTurn["role"], |
| 24 | + text: string, |
| 25 | + extra: Partial<ConversationTurn> = {}, |
| 26 | +): ConversationTurn { |
| 27 | + return { |
| 28 | + role, |
| 29 | + content: [{ type: "text", text }], |
| 30 | + timestamp: Date.now(), |
| 31 | + ...extra, |
| 32 | + }; |
| 33 | +} |
| 34 | + |
| 35 | +function allText(turns: ConversationTurn[]): string { |
| 36 | + return turns |
| 37 | + .flatMap((t) => |
| 38 | + t.content.filter((b) => b.type === "text").map((b) => b.text), |
| 39 | + ) |
| 40 | + .join("\n"); |
| 41 | +} |
| 42 | + |
| 43 | +// A dropped region with a standing goal, an exact path, a verification |
| 44 | +// command, and an unresolved failure. |
| 45 | +function droppedTurns(): ConversationTurn[] { |
| 46 | + return [ |
| 47 | + textTurn("user", "Migrate the auth module to opaque tokens in src/auth.ts"), |
| 48 | + textTurn("assistant", "Reading the handler first."), |
| 49 | + { |
| 50 | + role: "assistant", |
| 51 | + content: [ |
| 52 | + { |
| 53 | + type: "tool_call", |
| 54 | + id: "c1", |
| 55 | + name: "read_file", |
| 56 | + arguments: { path: "src/auth.ts" }, |
| 57 | + }, |
| 58 | + ], |
| 59 | + timestamp: 3, |
| 60 | + }, |
| 61 | + { |
| 62 | + role: "user", |
| 63 | + content: [ |
| 64 | + { |
| 65 | + type: "tool_result", |
| 66 | + callId: "c1", |
| 67 | + content: [{ type: "text", text: "handler source" }], |
| 68 | + }, |
| 69 | + ], |
| 70 | + timestamp: 4, |
| 71 | + }, |
| 72 | + { |
| 73 | + role: "assistant", |
| 74 | + content: [ |
| 75 | + { |
| 76 | + type: "tool_call", |
| 77 | + id: "c2", |
| 78 | + name: "run_shell", |
| 79 | + arguments: { command: "bun run test auth" }, |
| 80 | + }, |
| 81 | + ], |
| 82 | + timestamp: 5, |
| 83 | + }, |
| 84 | + { |
| 85 | + role: "user", |
| 86 | + content: [ |
| 87 | + { |
| 88 | + type: "tool_result", |
| 89 | + callId: "c2", |
| 90 | + isError: true, |
| 91 | + content: [{ type: "text", text: "token refresh assertion failed" }], |
| 92 | + }, |
| 93 | + ], |
| 94 | + timestamp: 6, |
| 95 | + }, |
| 96 | + textTurn("assistant", "Fix the token refresh assertion next."), |
| 97 | + ]; |
| 98 | +} |
| 99 | + |
| 100 | +describe("extractContinuationFacts", () => { |
| 101 | + test("lifts goal, exact names, verification, and blockers", () => { |
| 102 | + const facts = extractContinuationFacts(droppedTurns()); |
| 103 | + expect(facts.goal).toContain("opaque tokens"); |
| 104 | + expect(facts.exactNames).toContain("src/auth.ts"); |
| 105 | + expect(facts.verification).toContain("bun run test auth"); |
| 106 | + expect(facts.blockers.join("\n")).toContain("refresh assertion failed"); |
| 107 | + expect(facts.nextAction).toContain("refresh assertion next"); |
| 108 | + }); |
| 109 | + |
| 110 | + test("empty turns yield vacuous facts, never an abort", () => { |
| 111 | + const facts = extractContinuationFacts([]); |
| 112 | + const report = verifyCompactionSummary("anything", facts); |
| 113 | + expect(report.supported).toBe(true); |
| 114 | + }); |
| 115 | +}); |
| 116 | + |
| 117 | +describe("verifyCompactionSummary", () => { |
| 118 | + test("a faithful handoff is supported", () => { |
| 119 | + const facts = extractContinuationFacts(droppedTurns()); |
| 120 | + const report = verifyCompactionSummary( |
| 121 | + "Migrating auth to opaque tokens. Read src/auth.ts, ran bun run test " + |
| 122 | + "auth; the token refresh assertion failed. Fix the token refresh " + |
| 123 | + "assertion next.", |
| 124 | + facts, |
| 125 | + ); |
| 126 | + expect(report.supported).toBe(true); |
| 127 | + expect(report.misses).toEqual([]); |
| 128 | + }); |
| 129 | + |
| 130 | + test("a lossy handoff misses the goal and the exact name", () => { |
| 131 | + const facts = extractContinuationFacts(droppedTurns()); |
| 132 | + const report = verifyCompactionSummary( |
| 133 | + "Work continues. Next: fix tests.", |
| 134 | + facts, |
| 135 | + ); |
| 136 | + expect(report.supported).toBe(false); |
| 137 | + const kinds = report.misses.map((m) => m.kind); |
| 138 | + expect(kinds).toContain("goal"); |
| 139 | + expect(kinds).toContain("exactName"); |
| 140 | + }); |
| 141 | + |
| 142 | + test("denying failure while errors were dropped is a contradiction", () => { |
| 143 | + const facts = extractContinuationFacts(droppedTurns()); |
| 144 | + const report = verifyCompactionSummary( |
| 145 | + "Auth migration done. No errors remain.", |
| 146 | + facts, |
| 147 | + ); |
| 148 | + expect(report.misses.some((m) => m.kind === "contradiction")).toBe(true); |
| 149 | + }); |
| 150 | +}); |
| 151 | + |
| 152 | +describe("verifyOrRepair", () => { |
| 153 | + test("repairs a lossy handoff instead of shipping it", () => { |
| 154 | + const facts = extractContinuationFacts(droppedTurns()); |
| 155 | + const outcome = verifyOrRepair( |
| 156 | + "Work continues. Next: fix tests.", |
| 157 | + facts, |
| 158 | + 4000, |
| 159 | + ); |
| 160 | + expect(outcome.aborted).toBe(false); |
| 161 | + expect(outcome.repaired).toBe(true); |
| 162 | + expect(outcome.summary).toContain(VERIFY_REPAIR_HEADING); |
| 163 | + const rereport = verifyCompactionSummary(outcome.summary, facts); |
| 164 | + expect(rereport.misses.some((m) => m.kind === "goal")).toBe(false); |
| 165 | + expect(rereport.misses.some((m) => m.kind === "exactName")).toBe(false); |
| 166 | + }); |
| 167 | + |
| 168 | + test("aborts a contradicting handoff instead of shipping a lie", () => { |
| 169 | + const facts = extractContinuationFacts(droppedTurns()); |
| 170 | + const outcome = verifyOrRepair( |
| 171 | + "Auth migration done. No errors remain.", |
| 172 | + facts, |
| 173 | + 4000, |
| 174 | + ); |
| 175 | + expect(outcome.aborted).toBe(true); |
| 176 | + expect(outcome.repaired).toBe(false); |
| 177 | + }); |
| 178 | + |
| 179 | + test("repairSummary names only what the handoff missed", () => { |
| 180 | + const facts = extractContinuationFacts(droppedTurns()); |
| 181 | + const repaired = repairSummary( |
| 182 | + "Migrating auth to opaque tokens in src/auth.ts.", |
| 183 | + facts, |
| 184 | + verifyCompactionSummary( |
| 185 | + "Migrating auth to opaque tokens in src/auth.ts.", |
| 186 | + facts, |
| 187 | + ).misses, |
| 188 | + ); |
| 189 | + expect(repaired).toContain(VERIFY_REPAIR_HEADING); |
| 190 | + }); |
| 191 | +}); |
| 192 | + |
| 193 | +describe("pruning compactor verify pass", () => { |
| 194 | + test("a lossy fold is repaired: the goal and exact path survive", async () => { |
| 195 | + const compactor = createPruningCompactor({ |
| 196 | + keepRecentTurns: 2, |
| 197 | + summaryMaxChars: 2000, |
| 198 | + summarize: async () => "Work continues. Next: fix tests.", |
| 199 | + }); |
| 200 | + const turns: ConversationTurn[] = [ |
| 201 | + ...droppedTurns(), |
| 202 | + textTurn("user", "recent ask"), |
| 203 | + textTurn("assistant", "recent reply"), |
| 204 | + ]; |
| 205 | + const result = await compactor.apply(turns, mockStrategyCtx); |
| 206 | + expect(allText(result.output)).toContain("opaque tokens"); |
| 207 | + expect(allText(result.output)).toContain("auth.ts"); |
| 208 | + expect(result.record.decisions).toMatchObject({ verifyRepaired: 1 }); |
| 209 | + }); |
| 210 | + |
| 211 | + test("a contradicting fold aborts: prior context is kept", async () => { |
| 212 | + const compactor = createPruningCompactor({ |
| 213 | + keepRecentTurns: 2, |
| 214 | + summaryMaxChars: 2000, |
| 215 | + summarize: async () => "Auth migration done. No errors remain.", |
| 216 | + }); |
| 217 | + const turns: ConversationTurn[] = [ |
| 218 | + ...droppedTurns(), |
| 219 | + textTurn("user", "recent ask"), |
| 220 | + textTurn("assistant", "recent reply"), |
| 221 | + ]; |
| 222 | + const result = await compactor.apply(turns, mockStrategyCtx); |
| 223 | + expect(result.output).toBe(turns); |
| 224 | + expect(result.record.reason).toBe("verify failed — keeping prior context"); |
| 225 | + expect(result.record.decisions).toMatchObject({ verifyAborted: 1 }); |
| 226 | + }); |
| 227 | + |
| 228 | + test("a faithful fold ships without repair markers", async () => { |
| 229 | + const compactor = createPruningCompactor({ |
| 230 | + keepRecentTurns: 2, |
| 231 | + summaryMaxChars: 2000, |
| 232 | + summarize: async () => |
| 233 | + "Migrating auth to opaque tokens. Read src/auth.ts, ran bun run " + |
| 234 | + "test auth; the token refresh assertion failed. Fix the token " + |
| 235 | + "refresh assertion next.", |
| 236 | + }); |
| 237 | + const turns: ConversationTurn[] = [ |
| 238 | + ...droppedTurns(), |
| 239 | + textTurn("user", "recent ask"), |
| 240 | + textTurn("assistant", "recent reply"), |
| 241 | + ]; |
| 242 | + const result = await compactor.apply(turns, mockStrategyCtx); |
| 243 | + expect(allText(result.output)).not.toContain(VERIFY_REPAIR_HEADING); |
| 244 | + expect(result.record.decisions).not.toMatchObject({ verifyRepaired: 1 }); |
| 245 | + }); |
| 246 | +}); |
| 247 | + |
| 248 | +describe("continuation facts survive many folds", () => { |
| 249 | + test("goal, exact path, and next action hold after five lossy folds", async () => { |
| 250 | + const compactor = createPruningCompactor({ |
| 251 | + keepRecentTurns: 2, |
| 252 | + summaryMaxChars: 4000, |
| 253 | + summarize: async () => "Work continues. Next: fix tests.", |
| 254 | + }); |
| 255 | + let turns: ConversationTurn[] = [ |
| 256 | + ...droppedTurns(), |
| 257 | + textTurn("user", "recent ask"), |
| 258 | + textTurn("assistant", "recent reply"), |
| 259 | + ]; |
| 260 | + for (let fold = 0; fold < 5; fold++) { |
| 261 | + const result = await compactor.apply(turns, mockStrategyCtx); |
| 262 | + // No fold may abort the eval: the lossy stub is repaired, not denied. |
| 263 | + expect(result.record.reason).not.toBe( |
| 264 | + "verify failed — keeping prior context", |
| 265 | + ); |
| 266 | + turns = [ |
| 267 | + ...result.output, |
| 268 | + textTurn("user", `follow-up ${fold}`), |
| 269 | + textTurn("assistant", `progress note ${fold}`), |
| 270 | + ]; |
| 271 | + } |
| 272 | + const text = allText(turns); |
| 273 | + expect(text).toContain("opaque tokens"); |
| 274 | + expect(text).toContain("auth.ts"); |
| 275 | + expect(text).toContain("refresh assertion"); |
| 276 | + }); |
| 277 | +}); |
| 278 | + |
| 279 | +describe("condenseTurns keep-set", () => { |
| 280 | + test("pins the standing goal ahead of the recency window", () => { |
| 281 | + const turns: ConversationTurn[] = [ |
| 282 | + textTurn("user", "Standing goal: migrate auth to opaque tokens"), |
| 283 | + ...Array.from({ length: 8 }, (_, i) => |
| 284 | + textTurn("user", `follow-up dump number ${i}`), |
| 285 | + ), |
| 286 | + ]; |
| 287 | + const condensed = condenseTurns(turns); |
| 288 | + expect(condensed).toContain("opaque tokens"); |
| 289 | + expect(condensed).toContain("Goal (first user message)"); |
| 290 | + }); |
| 291 | +}); |
0 commit comments