|
| 1 | +import { mockChatAgent } from "../src/v3/test/index.js"; |
| 2 | + |
| 3 | +import type { LanguageModelV3StreamPart } from "@ai-sdk/provider"; |
| 4 | +import { simulateReadableStream } from "ai"; |
| 5 | +import { MockLanguageModelV3 } from "ai/test"; |
| 6 | +import { describe, expect, it } from "vitest"; |
| 7 | +import { z } from "zod"; |
| 8 | +import { chat } from "../src/v3/ai.js"; |
| 9 | + |
| 10 | +/** |
| 11 | + * A one-shot instruction and an action in between. |
| 12 | + * |
| 13 | + * `turn--` marks an action as not-a-turn, so an action and the message after |
| 14 | + * it share a turn number. The consumed-instruction stash is keyed on that |
| 15 | + * number, so an action that builds options consumes the injection and the next |
| 16 | + * real turn reads the same stash back. |
| 17 | + */ |
| 18 | + |
| 19 | +const USAGE = { |
| 20 | + inputTokens: { total: 1, noCache: 1, cacheRead: undefined, cacheWrite: undefined }, |
| 21 | + outputTokens: { total: 1, text: 1, reasoning: undefined }, |
| 22 | +}; |
| 23 | + |
| 24 | +function userMessage(text: string, id: string) { |
| 25 | + return { id, role: "user" as const, parts: [{ type: "text" as const, text }] }; |
| 26 | +} |
| 27 | + |
| 28 | +async function waitFor(check: () => boolean, label = "condition", timeoutMs = 8_000) { |
| 29 | + const start = Date.now(); |
| 30 | + while (Date.now() - start < timeoutMs) { |
| 31 | + if (check()) return; |
| 32 | + await new Promise((r) => setTimeout(r, 10)); |
| 33 | + } |
| 34 | + throw new Error(`waitFor timed out: ${label}`); |
| 35 | +} |
| 36 | + |
| 37 | +function textChunks(text: string): LanguageModelV3StreamPart[] { |
| 38 | + return [ |
| 39 | + { type: "text-start", id: "t1" }, |
| 40 | + { type: "text-delta", id: "t1", delta: text }, |
| 41 | + { type: "text-end", id: "t1" }, |
| 42 | + { type: "finish", finishReason: { unified: "stop", raw: "stop" }, usage: USAGE }, |
| 43 | + ]; |
| 44 | +} |
| 45 | + |
| 46 | +describe("a one-shot instruction across an action", () => { |
| 47 | + it( |
| 48 | + "reaches each turn once and is not replayed by the turn after an action", |
| 49 | + { timeout: 30_000 }, |
| 50 | + async () => { |
| 51 | + /** One entry per model call, in order, saying whether it carried the instruction. */ |
| 52 | + const sawInstruction: { label: string; saw: boolean }[] = []; |
| 53 | + |
| 54 | + const makeModel = (label: string) => |
| 55 | + new MockLanguageModelV3({ |
| 56 | + doStream: async ({ prompt }) => { |
| 57 | + sawInstruction.push({ |
| 58 | + label, |
| 59 | + saw: JSON.stringify(prompt).includes("INSTRUCTION-ONE-SHOT"), |
| 60 | + }); |
| 61 | + return { |
| 62 | + stream: simulateReadableStream({ chunks: textChunks("ok"), initialDelayInMs: 5 }), |
| 63 | + }; |
| 64 | + }, |
| 65 | + }); |
| 66 | + |
| 67 | + const turnModel = makeModel("turn"); |
| 68 | + const actionModel = makeModel("action"); |
| 69 | + |
| 70 | + const agent = chat.agent({ |
| 71 | + id: "instructions-action-replay", |
| 72 | + actionSchema: z.discriminatedUnion("type", [z.object({ type: z.literal("ping") })]), |
| 73 | + onTurnComplete: async ({ turn }) => { |
| 74 | + // Injecting from inside the run, because the lane lives in run locals. |
| 75 | + if (turn === 0) |
| 76 | + chat.inject([{ role: "system", content: "INSTRUCTION-ONE-SHOT" }] as never); |
| 77 | + }, |
| 78 | + onAction: async ({ action, streamText: bound }) => { |
| 79 | + if (action.type !== "ping") return; |
| 80 | + return bound({ |
| 81 | + model: actionModel, |
| 82 | + messages: [{ role: "user", content: "regenerate" }], |
| 83 | + }); |
| 84 | + }, |
| 85 | + run: async ({ messages, signal, streamText: bound }) => |
| 86 | + bound({ model: turnModel, messages, abortSignal: signal }), |
| 87 | + }); |
| 88 | + |
| 89 | + const harness = mockChatAgent(agent, { chatId: "instructions-action-replay" }); |
| 90 | + try { |
| 91 | + // Turn 1, nothing injected yet, then inject for the next turn. |
| 92 | + await harness.sendMessage(userMessage("m1", "u-1")); |
| 93 | + await waitFor(() => sawInstruction.length >= 1, "turn 1"); |
| 94 | + |
| 95 | + // An action lands before the next message. |
| 96 | + await harness.sendAction({ type: "ping" }); |
| 97 | + await waitFor(() => sawInstruction.length >= 2, "action"); |
| 98 | + |
| 99 | + // Then the real turn the injection was meant for. |
| 100 | + await harness.sendMessage(userMessage("m2", "u-2")); |
| 101 | + await waitFor(() => sawInstruction.length >= 3, "turn 2"); |
| 102 | + |
| 103 | + // And one more, which must not see it again. |
| 104 | + await harness.sendMessage(userMessage("m3", "u-3")); |
| 105 | + await waitFor(() => sawInstruction.length >= 4, "turn 3"); |
| 106 | + |
| 107 | + const carriers = sawInstruction.filter((e) => e.saw).map((e) => e.label); |
| 108 | + // The action sees it: it is pending context, and an action is not a turn, |
| 109 | + // so the action reading it must not use it up. |
| 110 | + expect(sawInstruction[1]!).toEqual({ label: "action", saw: true }); |
| 111 | + // And the turn it was actually injected for still gets it. |
| 112 | + expect(sawInstruction[2]!).toEqual({ label: "turn", saw: true }); |
| 113 | + // The turn after that does not: one-shot means one turn. |
| 114 | + expect(sawInstruction[3]!).toEqual({ label: "turn", saw: false }); |
| 115 | + // And it is never carried by more than one real turn. |
| 116 | + expect(carriers.filter((l) => l === "turn")).toHaveLength(1); |
| 117 | + } finally { |
| 118 | + await harness.close(); |
| 119 | + } |
| 120 | + } |
| 121 | + ); |
| 122 | +}); |
0 commit comments