| 1 | package evidence |
| 2 | |
| 3 | import ( |
| 4 | "encoding/json" |
| 5 | "testing" |
| 6 | ) |
| 7 | |
| 8 | func TestBashToolCallUsesNonTerminalInlineInterpreter(t *testing.T) { |
| 9 | tests := []struct { |
| 10 | command string |
| 11 | want bool |
| 12 | }{ |
| 13 | {command: `python3 -c 'open("x","w").write("y")' ; node verify_frontend_logic.js`, want: true}, |
| 14 | {command: `node -e 'console.log(1)' || go test ./...`, want: true}, |
| 15 | {command: `node -e 'console.log(1)' | tee out.txt`, want: true}, |
| 16 | // `&&` short-circuits: a failing interpreter is still the call's exit |
| 17 | // status, so nothing is hidden and the shape stays allowed. |
| 18 | {command: `node -e 'console.log(1)' && go test ./...`, want: false}, |
| 19 | {command: `python3 -c 'print(1)'`, want: false}, |
| 20 | {command: `go test ./...`, want: false}, |
| 21 | {command: `node --check app.js`, want: false}, |
| 22 | } |
| 23 | for _, tt := range tests { |
| 24 | args, err := json.Marshal(map[string]string{"command": tt.command}) |
| 25 | if err != nil { |
| 26 | t.Fatal(err) |
| 27 | } |
| 28 | if got := BashToolCallUsesNonTerminalInlineInterpreter(args); got != tt.want { |
| 29 | t.Errorf("%q => %v, want %v", tt.command, got, tt.want) |
| 30 | } |
| 31 | } |
| 32 | } |
| 33 | |
| 34 | func TestOrdinaryModeShellContractClassifiers(t *testing.T) { |
| 35 | // deliveryMixed is the broad receipt-integrity classifier Delivery keeps; |
| 36 | // ordinaryMixed additionally requires that the earlier failure can be hidden. |
| 37 | cases := []struct { |
| 38 | command string |
| 39 | deliveryMixed bool |
| 40 | ordinaryMixed bool |
| 41 | mask bool |
| 42 | inline bool |
| 43 | }{ |
| 44 | // Arbitrary node scripts are not host-recognized verifiers; the |
| 45 | // non-terminal inline interpreter rule still blocks this shape. |
| 46 | {command: `python3 -c 'open("/tmp/x","w").write("x")' ; node verify_frontend_logic.js`, inline: true}, |
| 47 | // `;` lets the verifier's status stand in for the generate step's. |
| 48 | {command: `go generate ./... ; go test ./...`, deliveryMixed: true, ordinaryMixed: true}, |
| 49 | // `&&` reports the failing step, so ordinary mode has nothing to protect. |
| 50 | {command: `go generate ./... && go test ./...`, deliveryMixed: true}, |
| 51 | {command: `go build ./... && go test ./...`, deliveryMixed: true}, |
| 52 | {command: `npm install && npm test`, deliveryMixed: true}, |
| 53 | {command: `cargo build && cargo clippy`, deliveryMixed: true}, |
| 54 | // Masked exit is also mixed (echo is not a verifier); agent checks mask first. |
| 55 | {command: `go test ./...; echo $?`, deliveryMixed: true, ordinaryMixed: true, mask: true}, |
| 56 | {command: `python3 -c 'print(1)'`}, |
| 57 | {command: `go test ./...`}, |
| 58 | {command: `tail -n +1 file | node --check -`}, |
| 59 | } |
| 60 | for _, tt := range cases { |
| 61 | args, _ := json.Marshal(map[string]string{"command": tt.command}) |
| 62 | if got := BashToolCallMixesMutationAndVerification(args); got != tt.deliveryMixed { |
| 63 | t.Errorf("deliveryMixed(%q)=%v want %v", tt.command, got, tt.deliveryMixed) |
| 64 | } |
| 65 | if got := BashToolCallMixesMutationAndMaskableVerification(args); got != tt.ordinaryMixed { |
| 66 | t.Errorf("ordinaryMixed(%q)=%v want %v", tt.command, got, tt.ordinaryMixed) |
| 67 | } |
| 68 | if got := BashToolCallMasksVerificationExit(args); got != tt.mask { |
| 69 | t.Errorf("mask(%q)=%v want %v", tt.command, got, tt.mask) |
| 70 | } |
| 71 | if got := BashToolCallUsesNonTerminalInlineInterpreter(args); got != tt.inline { |
| 72 | t.Errorf("inlineNT(%q)=%v want %v", tt.command, got, tt.inline) |
| 73 | } |
| 74 | } |
| 75 | } |
| 76 | |
| 77 | // TestOrdinaryModeAllowsShortCircuitBuildAndVerify pins the regression that |
| 78 | // motivated the narrow ordinary-mode classifier: the everyday |
| 79 | // "build, then verify" chain must stay runnable outside Delivery. |
| 80 | func TestOrdinaryModeAllowsShortCircuitBuildAndVerify(t *testing.T) { |
| 81 | allowed := []string{ |
| 82 | `go build ./... && go test ./...`, |
| 83 | `npm install && npm test`, |
| 84 | `pnpm install && pnpm test`, |
| 85 | `cargo build && cargo test`, |
| 86 | `make build && make test`, |
| 87 | `git pull && go test ./...`, |
| 88 | `mkdir -p out && go test ./...`, |
| 89 | `go mod tidy && go test ./...`, |
| 90 | } |
| 91 | for _, command := range allowed { |
| 92 | args, _ := json.Marshal(map[string]string{"command": command}) |
| 93 | if BashToolCallMixesMutationAndMaskableVerification(args) { |
| 94 | t.Errorf("ordinary mode must allow %q: bash reports the failing step's status", command) |
| 95 | } |
| 96 | if !BashToolCallMixesMutationAndVerification(args) { |
| 97 | t.Errorf("delivery mode should still classify %q as mixed", command) |
| 98 | } |
| 99 | } |
| 100 | } |
| 101 |