| 1 | package boot |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "strings" |
| 6 | "testing" |
| 7 | ) |
| 8 | |
| 9 | // TestBuildComposesByteStableSystemPrompt is the boot-level byte-stability |
| 10 | // guard: two Builds over the same workspace and config must compose the exact |
| 11 | // same system prompt. The system prompt is the provider-cached prefix of every |
| 12 | // request in every session — any byte of nondeterminism here (probe flaps, |
| 13 | // unsorted iteration, time-dependent content) cold-starts the provider cache |
| 14 | // for the whole machine, which is precisely the "desktop costs more" class |
| 15 | // (#2945). Environment probes are covered cross-process by the persisted |
| 16 | // snapshot tests in internal/environment; this test pins the rest of the |
| 17 | // composition (memory, skills index, output style, workspace line, policies). |
| 18 | func TestBuildComposesByteStableSystemPrompt(t *testing.T) { |
| 19 | isolateConfigHome(t) |
| 20 | dir := robustTempDir(t) |
| 21 | t.Chdir(dir) |
| 22 | |
| 23 | writeFile(t, dir, "reasonix.toml", ` |
| 24 | default_model = "test-model" |
| 25 | |
| 26 | [agent] |
| 27 | system_prompt = "BASE SYSTEM PROMPT" |
| 28 | |
| 29 | [[providers]] |
| 30 | name = "test-model" |
| 31 | kind = "openai" |
| 32 | base_url = "https://example.invalid" |
| 33 | model = "x" |
| 34 | api_key_env = "REASONIX_TEST_KEY_UNSET" |
| 35 | `) |
| 36 | writeFile(t, dir, "REASONIX.md", "Project rule: keep the prompt prefix stable.") |
| 37 | |
| 38 | first, err := Build(context.Background(), Options{}) |
| 39 | if err != nil { |
| 40 | t.Fatalf("first Build: %v", err) |
| 41 | } |
| 42 | firstPrompt := systemMessage(first.History()) |
| 43 | first.Close() |
| 44 | if strings.TrimSpace(firstPrompt) == "" { |
| 45 | t.Fatal("first Build composed an empty system prompt") |
| 46 | } |
| 47 | |
| 48 | second, err := Build(context.Background(), Options{}) |
| 49 | if err != nil { |
| 50 | t.Fatalf("second Build: %v", err) |
| 51 | } |
| 52 | secondPrompt := systemMessage(second.History()) |
| 53 | second.Close() |
| 54 | |
| 55 | if firstPrompt != secondPrompt { |
| 56 | t.Fatalf("system prompt is not byte-stable across identical Builds:\nfirst (%d bytes)\nsecond (%d bytes)\nfirst diff site: %q", |
| 57 | len(firstPrompt), len(secondPrompt), firstDivergence(firstPrompt, secondPrompt)) |
| 58 | } |
| 59 | } |
| 60 | |
| 61 | // firstDivergence returns a small window around the first differing byte so a |
| 62 | // failure names the drifting prompt section instead of dumping both prompts. |
| 63 | func firstDivergence(a, b string) string { |
| 64 | limit := len(a) |
| 65 | if len(b) < limit { |
| 66 | limit = len(b) |
| 67 | } |
| 68 | i := 0 |
| 69 | for i < limit && a[i] == b[i] { |
| 70 | i++ |
| 71 | } |
| 72 | start := i - 40 |
| 73 | if start < 0 { |
| 74 | start = 0 |
| 75 | } |
| 76 | endA := i + 40 |
| 77 | if endA > len(a) { |
| 78 | endA = len(a) |
| 79 | } |
| 80 | endB := i + 40 |
| 81 | if endB > len(b) { |
| 82 | endB = len(b) |
| 83 | } |
| 84 | return "..." + a[start:endA] + "... vs ..." + b[start:endB] + "..." |
| 85 | } |
| 86 |